M3: Copilot device-flow, token exchange, provider

Adds the Copilot provider: device-flow login modal support, copilot_internal/v2/token exchange with a direct-Bearer fallback, and routing across the three codecs, per the dual-path mitigation noted for Copilot token variance.

Cargo.toml

@@ -57,6 +57,9 @@ tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 tracing-appender = "0.2"
 dirs = "5"
 shell-words = "1"
+libc = "0.2"
+tempfile = "3"
+bytes = "1"
 
 [workspace.lints.rust]
 unsafe_code = "deny"

crates/harness-app/Cargo.toml

@@ -10,6 +10,14 @@ harness-providers = { workspace = true }
 harness-tools = { workspace = true }
 harness-mcp = { workspace = true }
 harness-lsp = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+dirs = { workspace = true }
+thiserror = { workspace = true }
+tracing = { workspace = true }
+
+[dev-dependencies]
+tempfile = { workspace = true }
 
 [lints]
 workspace = true

crates/harness-app/src/lib.rs

@@ -1 +1,562 @@
-// Composition root: builds the Engine and registers providers/tools/mcp/lsp. Lands in M1+.
+//! Composition root: builds the pieces `harness-tui` (and later `harness-server`) need —
+//! config, store, event bus, permission service, tool registry, provider registry — and
+//! exposes a small headless API (`run_prompt`) used by the `harness run -p` debug command.
+
+use std::collections::hash_map::DefaultHasher;
+use std::collections::HashMap;
+use std::hash::Hasher;
+use std::path::Path;
+use std::path::PathBuf;
+use std::sync::{Arc, Mutex};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use harness_core::config::{self, Config, ConfigError};
+use harness_core::engine::{run_session, RunConfig, StepContext};
+use harness_core::event::{AppEvent, EventBus, RunOutcome};
+use harness_core::permission::{PermissionReply, PermissionService};
+use harness_core::store::{Store, StoreError};
+use harness_core::tool::ToolRegistry;
+use harness_core::types::{
+    Message, MessageId, ModelRef, Part, PartBody, PartId, Session, SessionId,
+};
+use harness_providers::{AnthropicProvider, ModelCatalog, OpenAiProvider, ProviderRegistry};
+use tokio::task::JoinHandle;
+use tokio_util::sync::CancellationToken;
+
+/// Placeholder until M4's markdown agent registry lands (`assets/agents/orchestrator.md`).
+pub const DEFAULT_AGENT_PROMPT: &str =
+    "You are a helpful coding assistant with access to tools for reading, editing, running \
+     commands, and searching code. Use them as needed to satisfy the user's request, then \
+     reply with a concise final answer.";
+
+const DEFAULT_MAX_STEPS: u32 = 50;
+
+#[derive(Debug, thiserror::Error)]
+pub enum AppError {
+    #[error(transparent)]
+    Config(#[from] ConfigError),
+    #[error(transparent)]
+    Store(#[from] StoreError),
+    #[error("model ref must be \"provider/model\", got {0:?}")]
+    InvalidModelRef(String),
+    #[error("no provider registered for {0:?} (missing API key?)")]
+    UnknownProvider(String),
+    #[error("session {0} already has an active run")]
+    SessionRunning(SessionId),
+}
+
+fn now_ms() -> i64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .unwrap_or_default()
+        .as_millis() as i64
+}
+
+fn db_path(cwd: &Path) -> PathBuf {
+    let base = dirs::data_dir()
+        .unwrap_or_else(std::env::temp_dir)
+        .join("ai-harness")
+        .join("db");
+    let cwd_str = cwd.to_string_lossy();
+    let sanitized: String = cwd_str
+        .chars()
+        .map(|c| if c.is_alphanumeric() { c } else { '_' })
+        .collect();
+    let mut hasher = DefaultHasher::new();
+    hasher.write(cwd_str.as_bytes());
+    let hash = hasher.finish();
+    let hash_hex = format!("{:016x}", hash);
+    let slug = format!("{}_{}", sanitized, &hash_hex[..8]);
+    base.join(format!("{}.sqlite", slug))
+}
+
+/// Non-blocking, multi-turn engine API used by the TUI.
+#[derive(Clone)]
+pub struct EngineHandle {
+    inner: Arc<EngineInner>,
+}
+
+struct EngineInner {
+    config: Config,
+    store: Store,
+    bus: EventBus,
+    permissions: Arc<PermissionService>,
+    tools: ToolRegistry,
+    providers: ProviderRegistry,
+    catalog: ModelCatalog,
+    cwd: PathBuf,
+    data_dir: PathBuf,
+    runs: Mutex<HashMap<SessionId, RunHandle>>,
+}
+
+struct RunHandle {
+    cancel: CancellationToken,
+}
+
+impl EngineHandle {
+    /// Persistent SQLite-backed store. Used by the TUI and the default headless `App`.
+    pub fn init(cwd: PathBuf) -> Result<Self, AppError> {
+        let path = db_path(&cwd);
+        let store = Store::open(&path)?;
+        let handle = Self::new(cwd, store)?;
+        handle.spawn_catalog_refresh();
+        Ok(handle)
+    }
+
+    /// In-memory store — useful for tests and ephemeral sessions.
+    pub fn init_in_memory(cwd: PathBuf) -> Result<Self, AppError> {
+        let store = Store::open_in_memory()?;
+        Self::new(cwd, store)
+    }
+
+    fn new(cwd: PathBuf, store: Store) -> Result<Self, AppError> {
+        let config = config::load(&cwd)?;
+        let bus = EventBus::new();
+        let permissions = Arc::new(PermissionService::new(bus.clone()));
+
+        let mut tools = ToolRegistry::new();
+        harness_tools::register_builtins(&mut tools);
+
+        let mut providers = ProviderRegistry::new();
+        if let Some(key) = config
+            .providers
+            .get("anthropic")
+            .and_then(|p| p.api_key.clone())
+        {
+            providers.register(Arc::new(AnthropicProvider::new(key)));
+        }
+        if let Some(key) = config
+            .providers
+            .get("openai")
+            .and_then(|p| p.api_key.clone())
+        {
+            let provider = match config
+                .providers
+                .get("openai")
+                .and_then(|p| p.base_url.clone())
+            {
+                Some(base_url) => OpenAiProvider::with_base_url(key, base_url),
+                None => OpenAiProvider::new(key),
+            };
+            providers.register(Arc::new(provider));
+        }
+
+        let data_dir = dirs::data_dir()
+            .unwrap_or_else(std::env::temp_dir)
+            .join("ai-harness")
+            .join("tool-output");
+
+        // No network at construction: use the cached copy if fresh, else the baked snapshot.
+        // `init` warms the cache in the background for the next launch.
+        let catalog = ModelCatalog::load_cached_or_baked(&ModelCatalog::default_cache_path());
+
+        Ok(Self {
+            inner: Arc::new(EngineInner {
+                config,
+                store,
+                bus,
+                permissions,
+                tools,
+                providers,
+                catalog,
+                cwd,
+                data_dir,
+                runs: Mutex::new(HashMap::new()),
+            }),
+        })
+    }
+
+    /// Fire-and-forget refresh of the models.dev cache so the next launch has current pricing.
+    fn spawn_catalog_refresh(&self) {
+        tokio::spawn(async {
+            if let Err(e) = ModelCatalog::refresh_default_cache().await {
+                tracing::debug!(error = %e, "models.dev refresh failed; using cached/baked metadata");
+            }
+        });
+    }
+
+    pub fn bus(&self) -> EventBus {
+        self.inner.bus.clone()
+    }
+
+    pub fn permissions(&self) -> Arc<PermissionService> {
+        self.inner.permissions.clone()
+    }
+
+    pub fn config(&self) -> Config {
+        self.inner.config.clone()
+    }
+
+    pub async fn new_session(&self, agent: &str, model_ref: &str) -> Result<SessionId, AppError> {
+        let (provider_id, model_id) = model_ref
+            .split_once('/')
+            .ok_or_else(|| AppError::InvalidModelRef(model_ref.to_string()))?;
+        let model = ModelRef::new(provider_id, model_id);
+        let now = now_ms();
+        let session = Session::new_root(agent, model, now);
+        let session_id = session.id.clone();
+        self.inner.store.upsert_session(session.clone()).await?;
+        self.inner.bus.publish(AppEvent::SessionCreated { session });
+        Ok(session_id)
+    }
+
+    pub async fn prompt(
+        &self,
+        session_id: SessionId,
+        text: String,
+        model_ref: &str,
+    ) -> Result<(), AppError> {
+        let (provider_id, model_id) = model_ref
+            .split_once('/')
+            .ok_or_else(|| AppError::InvalidModelRef(model_ref.to_string()))?;
+        let provider = self
+            .inner
+            .providers
+            .get(provider_id)
+            .ok_or_else(|| AppError::UnknownProvider(provider_id.to_string()))?;
+        let model = ModelRef::new(provider_id, model_id);
+
+        {
+            let runs = self.inner.runs.lock().unwrap();
+            if runs.contains_key(&session_id) {
+                return Err(AppError::SessionRunning(session_id));
+            }
+        }
+
+        let now = now_ms();
+        let user_message = Message::new_user(session_id.clone(), now);
+        self.inner
+            .store
+            .upsert_message(user_message.clone())
+            .await?;
+        // The store actor does not emit events, so publish the user message and its part
+        // ourselves — otherwise the TUI (which builds its live transcript purely from bus
+        // events) never shows the prompt the user just typed until the session is reloaded.
+        self.inner.bus.publish(AppEvent::MessageCreated {
+            message: user_message.clone(),
+        });
+
+        let user_part = Part {
+            id: PartId::new(),
+            message_id: user_message.id.clone(),
+            session_id: session_id.clone(),
+            idx: 0,
+            body: PartBody::Text {
+                text,
+                synthetic: false,
+            },
+        };
+        self.inner.store.upsert_part(user_part.clone()).await?;
+        self.inner
+            .bus
+            .publish(AppEvent::PartUpdated { part: user_part });
+
+        let ctx = StepContext {
+            store: self.inner.store.clone(),
+            bus: self.inner.bus.clone(),
+            tools: self.inner.tools.clone(),
+            permissions: self.inner.permissions.clone(),
+            static_rules: self.inner.config.permissions.clone(),
+            extra_rules: Arc::new(Mutex::new(Vec::new())),
+            session_id: session_id.clone(),
+            cwd: self.inner.cwd.clone(),
+            data_dir: self.inner.data_dir.join(session_id.to_string()),
+            cancel: CancellationToken::new(),
+            now,
+        };
+        let run_config = RunConfig {
+            agent_name: "orchestrator".to_string(),
+            agent_prompt: DEFAULT_AGENT_PROMPT.to_string(),
+            model,
+            temperature: None,
+            max_steps: DEFAULT_MAX_STEPS,
+            instructions: self.inner.config.instructions.clone(),
+            cost: Some(self.inner.catalog.cost(provider_id, model_id)),
+        };
+
+        // Reserve the run slot *before* spawning. If we inserted after spawning, a run that
+        // finished quickly could remove its (not-yet-inserted) entry first, and our later
+        // insert would then strand the session as permanently "running".
+        let cancel = ctx.cancel.clone();
+        {
+            let mut runs = self.inner.runs.lock().unwrap();
+            if runs.contains_key(&session_id) {
+                return Err(AppError::SessionRunning(session_id));
+            }
+            runs.insert(
+                session_id.clone(),
+                RunHandle {
+                    cancel: cancel.clone(),
+                },
+            );
+        }
+
+        self.inner.bus.publish(AppEvent::RunStarted {
+            session_id: session_id.clone(),
+        });
+
+        let inner = self.inner.clone();
+        let spawn_session_id = session_id.clone();
+        tokio::spawn(async move {
+            let outcome = run_session(provider, ctx, &run_config, now_ms).await;
+            inner.bus.publish(AppEvent::RunFinished {
+                session_id: spawn_session_id.clone(),
+                outcome: outcome.clone(),
+            });
+            let mut runs = inner.runs.lock().unwrap();
+            runs.remove(&spawn_session_id);
+        });
+
+        Ok(())
+    }
+
+    pub fn abort(&self, session_id: &SessionId) {
+        let runs = self.inner.runs.lock().unwrap();
+        if let Some(run) = runs.get(session_id) {
+            run.cancel.cancel();
+        }
+    }
+
+    pub fn permission_reply(&self, id: &str, reply: PermissionReply) -> bool {
+        self.inner.permissions.reply(id, reply)
+    }
+
+    pub async fn list_sessions(&self) -> Result<Vec<Session>, AppError> {
+        Ok(self.inner.store.sessions().await?)
+    }
+
+    pub async fn session_messages(&self, session_id: SessionId) -> Result<Vec<Message>, AppError> {
+        Ok(self.inner.store.messages(session_id).await?)
+    }
+
+    pub async fn message_parts(&self, message_id: MessageId) -> Result<Vec<Part>, AppError> {
+        Ok(self.inner.store.parts(message_id).await?)
+    }
+
+    pub fn is_running(&self, session_id: &SessionId) -> bool {
+        let runs = self.inner.runs.lock().unwrap();
+        runs.contains_key(session_id)
+    }
+
+    pub async fn final_text(&self, session_id: &SessionId) -> Result<String, AppError> {
+        let messages = self.inner.store.messages(session_id.clone()).await?;
+        let Some(last) = messages.last() else {
+            return Ok(String::new());
+        };
+        let parts = self.inner.store.parts(last.id.clone()).await?;
+        let text = parts
+            .iter()
+            .filter_map(|p| match &p.body {
+                PartBody::Text { text, .. } => Some(text.as_str()),
+                _ => None,
+            })
+            .collect::<Vec<_>>()
+            .join("");
+        Ok(text)
+    }
+}
+
+pub struct App {
+    engine: EngineHandle,
+    // Keeps the auto-approve task alive for the lifetime of the App.
+    _auto_approve_handle: Option<JoinHandle<()>>,
+}
+
+impl App {
+    /// Persistent SQLite-backed store with an auto-approve permission frontend — the default
+    /// headless configuration used by `harness run -p`.
+    pub fn init(cwd: PathBuf) -> Result<Self, AppError> {
+        let engine = EngineHandle::init(cwd)?;
+        let _auto_approve_handle = Some(spawn_auto_approve_task(&engine));
+        Ok(Self {
+            engine,
+            _auto_approve_handle,
+        })
+    }
+
+    /// In-memory store with an auto-approve permission frontend — useful for tests.
+    pub fn init_in_memory(cwd: PathBuf) -> Result<Self, AppError> {
+        let engine = EngineHandle::init_in_memory(cwd)?;
+        let _auto_approve_handle = Some(spawn_auto_approve_task(&engine));
+        Ok(Self {
+            engine,
+            _auto_approve_handle,
+        })
+    }
+
+    pub fn engine(&self) -> &EngineHandle {
+        &self.engine
+    }
+
+    pub fn config(&self) -> Config {
+        self.engine.config()
+    }
+
+    /// Runs a single headless turn: creates a root session, appends `prompt` as the user
+    /// message, and drives the engine loop to completion. Returns the outcome plus the
+    /// session id so the caller can fetch the transcript via `final_text`.
+    pub async fn run_prompt(
+        &self,
+        prompt: String,
+        model_ref: &str,
+    ) -> Result<(RunOutcome, SessionId), AppError> {
+        let session_id = self.engine.new_session("orchestrator", model_ref).await?;
+        let mut rx = self.engine.bus().subscribe();
+        // Subscribe before starting the run so we cannot miss the RunFinished event
+        // even for an instant (mock) provider.
+        self.engine
+            .prompt(session_id.clone(), prompt, model_ref)
+            .await?;
+
+        while let Ok(event) = rx.recv().await {
+            if let AppEvent::RunFinished {
+                session_id: sid,
+                outcome,
+            } = event
+            {
+                if sid == session_id {
+                    return Ok((outcome, session_id));
+                }
+            }
+        }
+
+        // Bus closed unexpectedly; report a stopped run so the caller can continue.
+        Ok((RunOutcome::Stopped, session_id))
+    }
+
+    /// Concatenates the `Text` parts of the last message in the session — the final
+    /// assistant reply for a `harness run` invocation to print.
+    pub async fn final_text(&self, session_id: &SessionId) -> Result<String, AppError> {
+        self.engine.final_text(session_id).await
+    }
+}
+
+fn spawn_auto_approve_task(engine: &EngineHandle) -> JoinHandle<()> {
+    let bus = engine.bus();
+    let permissions = engine.permissions();
+    tokio::spawn(async move {
+        let mut rx = bus.subscribe();
+        while let Ok(event) = rx.recv().await {
+            if let AppEvent::PermissionAsked { request } = event {
+                permissions.reply(&request.id, PermissionReply::Once);
+            }
+        }
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn init_loads_defaults_with_no_providers_configured() {
+        let dir = tempfile::tempdir().unwrap();
+        std::env::remove_var("ANTHROPIC_API_KEY");
+        let app = App::init_in_memory(dir.path().to_path_buf()).unwrap();
+        assert!(app.engine.inner.providers.get("anthropic").is_none());
+        assert_eq!(app.engine.inner.tools.all().len(), 6);
+    }
+
+    #[tokio::test]
+    async fn run_prompt_errors_on_unregistered_provider() {
+        let dir = tempfile::tempdir().unwrap();
+        std::env::remove_var("ANTHROPIC_API_KEY");
+        let app = App::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let err = app
+            .run_prompt("hi".into(), "anthropic/claude-sonnet-4-5")
+            .await
+            .unwrap_err();
+        assert!(matches!(err, AppError::UnknownProvider(_)));
+    }
+
+    #[tokio::test]
+    async fn run_prompt_errors_on_malformed_model_ref() {
+        let dir = tempfile::tempdir().unwrap();
+        let app = App::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let err = app.run_prompt("hi".into(), "no-slash").await.unwrap_err();
+        assert!(matches!(err, AppError::InvalidModelRef(_)));
+    }
+
+    #[tokio::test]
+    async fn final_text_is_empty_for_unknown_session() {
+        let dir = tempfile::tempdir().unwrap();
+        let app = App::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let text = app.final_text(&SessionId::new()).await.unwrap();
+        assert_eq!(text, "");
+    }
+
+    #[tokio::test]
+    async fn engine_init_in_memory_creates_handle() {
+        let dir = tempfile::tempdir().unwrap();
+        let engine = EngineHandle::init_in_memory(dir.path().to_path_buf()).unwrap();
+        assert!(!engine.config().providers.contains_key("anthropic"));
+    }
+
+    #[tokio::test]
+    async fn engine_new_session_persists_session() {
+        let dir = tempfile::tempdir().unwrap();
+        let engine = EngineHandle::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let id = engine
+            .new_session("orchestrator", "anthropic/claude")
+            .await
+            .unwrap();
+        let sessions = engine.list_sessions().await.unwrap();
+        assert_eq!(sessions.len(), 1);
+        assert_eq!(sessions[0].id, id);
+        assert_eq!(sessions[0].agent, "orchestrator");
+    }
+
+    #[tokio::test]
+    async fn engine_list_sessions_returns_persisted_sessions() {
+        let dir = tempfile::tempdir().unwrap();
+        let engine = EngineHandle::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let id1 = engine.new_session("a", "anthropic/claude").await.unwrap();
+        let id2 = engine.new_session("b", "anthropic/claude").await.unwrap();
+        let sessions = engine.list_sessions().await.unwrap();
+        assert_eq!(sessions.len(), 2);
+        assert!(sessions.iter().any(|s| s.id == id1));
+        assert!(sessions.iter().any(|s| s.id == id2));
+    }
+
+    #[tokio::test]
+    async fn engine_prompt_errors_on_unregistered_provider() {
+        let dir = tempfile::tempdir().unwrap();
+        std::env::remove_var("ANTHROPIC_API_KEY");
+        let engine = EngineHandle::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let id = engine
+            .new_session("orchestrator", "anthropic/claude")
+            .await
+            .unwrap();
+        let err = engine
+            .prompt(id, "hi".into(), "anthropic/claude")
+            .await
+            .unwrap_err();
+        assert!(matches!(err, AppError::UnknownProvider(_)));
+    }
+
+    #[tokio::test]
+    async fn engine_prompt_errors_on_malformed_model_ref() {
+        let dir = tempfile::tempdir().unwrap();
+        let engine = EngineHandle::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let id = engine
+            .new_session("orchestrator", "anthropic/claude")
+            .await
+            .unwrap();
+        let err = engine
+            .prompt(id, "hi".into(), "no-slash")
+            .await
+            .unwrap_err();
+        assert!(matches!(err, AppError::InvalidModelRef(_)));
+    }
+
+    #[tokio::test]
+    async fn engine_is_running_false_for_inactive_session() {
+        let dir = tempfile::tempdir().unwrap();
+        let engine = EngineHandle::init_in_memory(dir.path().to_path_buf()).unwrap();
+        let id = engine
+            .new_session("orchestrator", "anthropic/claude")
+            .await
+            .unwrap();
+        assert!(!engine.is_running(&id));
+    }
+}

crates/harness-core/Cargo.toml

@@ -14,12 +14,15 @@ serde_json = { workspace = true }
 schemars = { workspace = true }
 rusqlite = { workspace = true }
 globset = { workspace = true }
+jsonc-parser = { workspace = true }
+dirs = { workspace = true }
 ulid = { workspace = true }
 thiserror = { workspace = true }
 tracing = { workspace = true }
 
 [dev-dependencies]
 tokio = { workspace = true, features = ["test-util", "macros"] }
+tempfile = { workspace = true }
 
 [lints]
 workspace = true

crates/harness-core/src/config/load.rs

@@ -0,0 +1,269 @@
+use std::path::{Path, PathBuf};
+
+use serde_json::Value;
+
+use super::schema::Config;
+
+#[derive(Debug, thiserror::Error)]
+pub enum ConfigError {
+    #[error("reading {path}: {source}")]
+    Io {
+        path: PathBuf,
+        #[source]
+        source: std::io::Error,
+    },
+    #[error("parsing {path}: {source}")]
+    Jsonc {
+        path: PathBuf,
+        #[source]
+        source: jsonc_parser::errors::ParseError,
+    },
+    #[error("deserializing merged config: {0}")]
+    Serde(#[from] serde_json::Error),
+}
+
+/// Env vars consulted directly (not via `{env:...}` interpolation) — highest precedence.
+const ENV_MODEL: &str = "AI_HARNESS_MODEL";
+const ENV_CONFIG_PATH: &str = "AI_HARNESS_CONFIG";
+/// `(provider id, env var)` pairs used to fill in `providers.<id>.api_key` when unset.
+const PROVIDER_ENV_KEYS: &[(&str, &str)] = &[
+    ("anthropic", "ANTHROPIC_API_KEY"),
+    ("openai", "OPENAI_API_KEY"),
+];
+
+fn read_jsonc(path: &Path) -> Result<Value, ConfigError> {
+    let text = std::fs::read_to_string(path).map_err(|source| ConfigError::Io {
+        path: path.to_path_buf(),
+        source,
+    })?;
+    let value =
+        jsonc_parser::parse_to_serde_value(&text, &Default::default()).map_err(|source| {
+            ConfigError::Jsonc {
+                path: path.to_path_buf(),
+                source,
+            }
+        })?;
+    Ok(value.unwrap_or_else(|| Value::Object(Default::default())))
+}
+
+/// Objects deep-merge (overlay wins per key); anything else (including arrays) is replaced.
+fn merge(base: &mut Value, overlay: Value) {
+    match (base, overlay) {
+        (Value::Object(base_map), Value::Object(overlay_map)) => {
+            for (key, value) in overlay_map {
+                match base_map.get_mut(&key) {
+                    Some(existing) => merge(existing, value),
+                    None => {
+                        base_map.insert(key, value);
+                    }
+                }
+            }
+        }
+        (base_slot, overlay_value) => {
+            *base_slot = overlay_value;
+        }
+    }
+}
+
+/// Exact-string placeholders only (opencode-style): `"{env:VAR}"` and `"{file:path}"`.
+/// `path` in `{file:...}` resolves relative to `base_dir`.
+fn interpolate(value: &mut Value, base_dir: &Path) {
+    match value {
+        Value::String(s) => {
+            if let Some(var) = s.strip_prefix("{env:").and_then(|s| s.strip_suffix('}')) {
+                *s = std::env::var(var).unwrap_or_default();
+            } else if let Some(rel) = s.strip_prefix("{file:").and_then(|s| s.strip_suffix('}')) {
+                *s = std::fs::read_to_string(base_dir.join(rel)).unwrap_or_default();
+            }
+        }
+        Value::Array(items) => {
+            for item in items {
+                interpolate(item, base_dir);
+            }
+        }
+        Value::Object(map) => {
+            for (_, v) in map.iter_mut() {
+                interpolate(v, base_dir);
+            }
+        }
+        _ => {}
+    }
+}
+
+fn global_config_path() -> Option<PathBuf> {
+    dirs::config_dir().map(|d| d.join("ai-harness").join("config.json"))
+}
+
+/// Ancestors of `cwd` that have a `.harness/config.json`, root-most first (lowest
+/// precedence among the project chain; `cwd`'s own file, if any, applies last).
+fn project_chain(cwd: &Path) -> Vec<PathBuf> {
+    let mut found: Vec<PathBuf> = cwd
+        .ancestors()
+        .map(|dir| dir.join(".harness").join("config.json"))
+        .filter(|p| p.is_file())
+        .collect();
+    found.reverse();
+    found
+}
+
+pub fn load(cwd: &Path) -> Result<Config, ConfigError> {
+    let mut merged = serde_json::to_value(Config::default())?;
+
+    if let Some(path) = global_config_path() {
+        if path.is_file() {
+            merge(&mut merged, read_jsonc(&path)?);
+        }
+    }
+
+    for path in project_chain(cwd) {
+        merge(&mut merged, read_jsonc(&path)?);
+    }
+
+    if let Ok(path) = std::env::var(ENV_CONFIG_PATH) {
+        let path = PathBuf::from(path);
+        if path.is_file() {
+            merge(&mut merged, read_jsonc(&path)?);
+        }
+    }
+
+    interpolate(&mut merged, cwd);
+
+    let mut config: Config = serde_json::from_value(merged)?;
+
+    if let Ok(model) = std::env::var(ENV_MODEL) {
+        config.model = Some(model);
+    }
+    for (provider_id, env_var) in PROVIDER_ENV_KEYS {
+        if let Ok(key) = std::env::var(env_var) {
+            let entry = config
+                .providers
+                .entry((*provider_id).to_string())
+                .or_default();
+            if entry.api_key.is_none() {
+                entry.api_key = Some(key);
+            }
+        }
+    }
+
+    Ok(config)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::sync::Mutex;
+
+    // Env var mutation races across tests in the same process; serialize them.
+    static ENV_LOCK: Mutex<()> = Mutex::new(());
+
+    #[test]
+    fn merge_deep_merges_objects_and_replaces_arrays() {
+        let mut base = serde_json::json!({
+            "a": {"x": 1, "y": 2},
+            "list": [1, 2, 3],
+        });
+        let overlay = serde_json::json!({
+            "a": {"y": 20, "z": 3},
+            "list": [9],
+        });
+        merge(&mut base, overlay);
+        assert_eq!(
+            base,
+            serde_json::json!({"a": {"x": 1, "y": 20, "z": 3}, "list": [9]})
+        );
+    }
+
+    #[test]
+    fn interpolate_replaces_env_placeholder() {
+        let _lock = ENV_LOCK.lock().unwrap();
+        std::env::set_var("HARNESS_TEST_VAR", "secret");
+        let mut value = serde_json::json!({"api_key": "{env:HARNESS_TEST_VAR}"});
+        interpolate(&mut value, Path::new("."));
+        assert_eq!(value, serde_json::json!({"api_key": "secret"}));
+        std::env::remove_var("HARNESS_TEST_VAR");
+    }
+
+    #[test]
+    fn interpolate_replaces_file_placeholder() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("prompt.md"), "You are a reviewer.").unwrap();
+        let mut value = serde_json::json!({"prompt": "{file:prompt.md}"});
+        interpolate(&mut value, dir.path());
+        assert_eq!(value, serde_json::json!({"prompt": "You are a reviewer."}));
+    }
+
+    #[test]
+    fn load_applies_bundled_defaults_when_nothing_else_present() {
+        let _lock = ENV_LOCK.lock().unwrap();
+        std::env::remove_var(ENV_MODEL);
+        std::env::remove_var(ENV_CONFIG_PATH);
+        let dir = tempfile::tempdir().unwrap();
+        let config = load(dir.path()).unwrap();
+        assert_eq!(config.orchestration.depth_limit, 3);
+        assert!(config.model.is_none());
+    }
+
+    #[test]
+    fn load_merges_project_chain_root_most_first() {
+        let _lock = ENV_LOCK.lock().unwrap();
+        std::env::remove_var(ENV_MODEL);
+        std::env::remove_var(ENV_CONFIG_PATH);
+
+        let root = tempfile::tempdir().unwrap();
+        let child = root.path().join("child");
+        std::fs::create_dir_all(child.join(".harness")).unwrap();
+        std::fs::create_dir_all(root.path().join(".harness")).unwrap();
+
+        std::fs::write(
+            root.path().join(".harness/config.json"),
+            r#"{ "model": "root-model", "orchestration": { "depth_limit": 5 } }"#,
+        )
+        .unwrap();
+        std::fs::write(
+            child.join(".harness/config.json"),
+            r#"{ "model": "child-model" }"#,
+        )
+        .unwrap();
+
+        let config = load(&child).unwrap();
+        // child's config.json overrides root's `model` (applied later in the chain)...
+        assert_eq!(config.model.as_deref(), Some("child-model"));
+        // ...but root's untouched `orchestration.depth_limit` survives the merge.
+        assert_eq!(config.orchestration.depth_limit, 5);
+    }
+
+    #[test]
+    fn env_model_override_wins_over_everything() {
+        let _lock = ENV_LOCK.lock().unwrap();
+        std::env::remove_var(ENV_CONFIG_PATH);
+        std::env::set_var(ENV_MODEL, "env-model");
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::create_dir_all(dir.path().join(".harness")).unwrap();
+        std::fs::write(
+            dir.path().join(".harness/config.json"),
+            r#"{ "model": "project-model" }"#,
+        )
+        .unwrap();
+        let config = load(dir.path()).unwrap();
+        assert_eq!(config.model.as_deref(), Some("env-model"));
+        std::env::remove_var(ENV_MODEL);
+    }
+
+    #[test]
+    fn provider_env_key_fills_in_missing_api_key_only() {
+        let _lock = ENV_LOCK.lock().unwrap();
+        std::env::remove_var(ENV_MODEL);
+        std::env::remove_var(ENV_CONFIG_PATH);
+        std::env::set_var("ANTHROPIC_API_KEY", "from-env");
+        let dir = tempfile::tempdir().unwrap();
+        let config = load(dir.path()).unwrap();
+        assert_eq!(
+            config
+                .providers
+                .get("anthropic")
+                .and_then(|p| p.api_key.clone()),
+            Some("from-env".to_string())
+        );
+        std::env::remove_var("ANTHROPIC_API_KEY");
+    }
+}

crates/harness-core/src/config/mod.rs

@@ -0,0 +1,8 @@
+pub mod load;
+pub mod schema;
+
+pub use load::{load, ConfigError};
+pub use schema::{
+    AgentPatch, Config, LspServerConfig, McpServerConfig, OrchestrationConfig, ProviderConfig,
+    TuiConfig,
+};

crates/harness-core/src/config/schema.rs

@@ -0,0 +1,86 @@
+use std::collections::HashMap;
+
+use serde::{Deserialize, Serialize};
+
+use crate::permission::Rule;
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(default)]
+pub struct ProviderConfig {
+    pub enabled: Option<bool>,
+    pub api_key: Option<String>,
+    pub base_url: Option<String>,
+}
+
+/// Patch semantics over a loaded `AgentDef` (M4): only the fields set here override the
+/// bundled/markdown definition. An unknown agent name with at least `model` or `prompt`
+/// set creates a custom agent.
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(default)]
+pub struct AgentPatch {
+    pub disable: Option<bool>,
+    pub mode: Option<String>,
+    pub model: Option<String>,
+    pub temperature: Option<f32>,
+    pub prompt: Option<String>,
+    pub tools: Option<HashMap<String, bool>>,
+    pub permission: Option<Vec<Rule>>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(default)]
+pub struct OrchestrationConfig {
+    pub depth_limit: u32,
+    pub job_board: bool,
+    pub max_reusable_per_agent: u32,
+    pub reminders: Option<HashMap<String, String>>,
+}
+
+impl Default for OrchestrationConfig {
+    fn default() -> Self {
+        Self {
+            depth_limit: 3,
+            job_board: true,
+            max_reusable_per_agent: 2,
+            reminders: None,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(default)]
+pub struct McpServerConfig {
+    pub command: String,
+    pub args: Vec<String>,
+    pub env: HashMap<String, String>,
+    pub enabled: Option<bool>,
+}
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(default)]
+pub struct LspServerConfig {
+    pub command: String,
+    pub args: Vec<String>,
+    pub extensions: Vec<String>,
+}
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(default)]
+pub struct TuiConfig {
+    pub theme: Option<String>,
+}
+
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+#[serde(default)]
+pub struct Config {
+    pub model: Option<String>,
+    pub small_model: Option<String>,
+    pub providers: HashMap<String, ProviderConfig>,
+    pub agents: HashMap<String, AgentPatch>,
+    pub permissions: Vec<Rule>,
+    pub orchestration: OrchestrationConfig,
+    pub mcp: HashMap<String, McpServerConfig>,
+    pub lsp: HashMap<String, LspServerConfig>,
+    pub tui: TuiConfig,
+    pub instructions: Vec<String>,
+}

crates/harness-core/src/engine/doomloop.rs

@@ -0,0 +1,75 @@
+use std::collections::VecDeque;
+use std::hash::{Hash, Hasher};
+
+/// opencode's `DOOM_LOOP_THRESHOLD`: force a permission ask once the model repeats the
+/// exact same tool call this many times in a row.
+const DOOM_LOOP_THRESHOLD: usize = 3;
+
+fn fingerprint(tool_name: &str, input: &serde_json::Value) -> (String, u64) {
+    let mut hasher = std::collections::hash_map::DefaultHasher::new();
+    input.to_string().hash(&mut hasher);
+    (tool_name.to_string(), hasher.finish())
+}
+
+/// Tracks the last `DOOM_LOOP_THRESHOLD` `(tool_name, hash(input))` pairs for one run.
+pub struct DoomLoopGuard {
+    history: VecDeque<(String, u64)>,
+}
+
+impl DoomLoopGuard {
+    pub fn new() -> Self {
+        Self {
+            history: VecDeque::with_capacity(DOOM_LOOP_THRESHOLD),
+        }
+    }
+
+    /// Records this call and returns `true` if it's the `DOOM_LOOP_THRESHOLD`-th identical
+    /// call in a row (the caller should force a permission ask regardless of the ruleset).
+    pub fn record_and_check(&mut self, tool_name: &str, input: &serde_json::Value) -> bool {
+        let fp = fingerprint(tool_name, input);
+        self.history.push_back(fp.clone());
+        if self.history.len() > DOOM_LOOP_THRESHOLD {
+            self.history.pop_front();
+        }
+        self.history.len() == DOOM_LOOP_THRESHOLD && self.history.iter().all(|h| *h == fp)
+    }
+}
+
+impl Default for DoomLoopGuard {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn triggers_after_three_identical_calls() {
+        let mut guard = DoomLoopGuard::new();
+        let input = serde_json::json!({"command": "ls"});
+        assert!(!guard.record_and_check("bash", &input));
+        assert!(!guard.record_and_check("bash", &input));
+        assert!(guard.record_and_check("bash", &input));
+    }
+
+    #[test]
+    fn different_input_resets_the_streak() {
+        let mut guard = DoomLoopGuard::new();
+        assert!(!guard.record_and_check("bash", &serde_json::json!({"command": "ls"})));
+        assert!(!guard.record_and_check("bash", &serde_json::json!({"command": "pwd"})));
+        assert!(!guard.record_and_check("bash", &serde_json::json!({"command": "ls"})));
+    }
+
+    #[test]
+    fn keeps_triggering_while_the_streak_continues() {
+        let mut guard = DoomLoopGuard::new();
+        let input = serde_json::json!({"command": "ls"});
+        for _ in 0..2 {
+            guard.record_and_check("bash", &input);
+        }
+        assert!(guard.record_and_check("bash", &input));
+        assert!(guard.record_and_check("bash", &input));
+    }
+}

crates/harness-core/src/engine/loop.rs

@@ -0,0 +1,581 @@
+use std::sync::Arc;
+
+use crate::event::RunOutcome;
+use crate::llm::{
+    FinishReason, Initiator, LlmRequest, Provider, ProviderError, Role as WireRole, ToolSchema,
+    WireContent, WireMessage,
+};
+use crate::store::Store;
+use crate::types::{Message, Part, PartBody, Role, SessionId, ToolState};
+
+use super::doomloop::DoomLoopGuard;
+use super::processor::{process_step, StepContext, StepResult};
+use super::retry;
+use super::system;
+
+pub struct RunConfig {
+    pub agent_name: String,
+    pub agent_prompt: String,
+    pub model: crate::types::ModelRef,
+    pub temperature: Option<f32>,
+    pub max_steps: u32,
+    pub instructions: Vec<String>,
+    /// Pricing for `model`, from models.dev metadata. `None` leaves cost at 0.
+    pub cost: Option<crate::types::ModelCost>,
+}
+
+/// Adds a step's usage/cost onto the persisted session and republishes it. Cost accounting is
+/// best-effort: a store error here is logged, not surfaced as a run failure.
+async fn accumulate_session_usage(
+    ctx: &StepContext,
+    usage: &crate::types::TokenUsage,
+    cost: f64,
+    now: i64,
+) {
+    match ctx.store.session(ctx.session_id.clone()).await {
+        Ok(Some(mut session)) => {
+            session.usage.add(usage);
+            session.cost += cost;
+            session.updated_at = now;
+            if let Err(e) = ctx.store.upsert_session(session.clone()).await {
+                tracing::warn!(error = %e, "failed to persist session usage");
+                return;
+            }
+            ctx.bus
+                .publish(crate::event::AppEvent::SessionUpdated { session });
+        }
+        Ok(None) => {}
+        Err(e) => tracing::warn!(error = %e, "failed to load session for usage accounting"),
+    }
+}
+
+/// opencode's exit condition: keep stepping while the last assistant turn asked for more
+/// tool calls; stop once it produced a final answer (or a fresh user message is waiting).
+async fn should_continue(store: &Store, session_id: &SessionId) -> Result<bool, ProviderError> {
+    let messages = store
+        .messages(session_id.clone())
+        .await
+        .map_err(|e| ProviderError::Decode(e.to_string()))?;
+    Ok(match messages.last() {
+        None => false,
+        Some(msg) if msg.role == Role::User => true,
+        Some(msg) => matches!(msg.finished, Some(FinishReason::ToolCalls)),
+    })
+}
+
+fn convert_message(message: &Message, parts: &[Part]) -> Vec<WireMessage> {
+    match message.role {
+        Role::User => {
+            let content: Vec<WireContent> = parts
+                .iter()
+                .filter_map(|p| match &p.body {
+                    PartBody::Text { text, .. } => Some(WireContent::Text { text: text.clone() }),
+                    _ => None,
+                })
+                .collect();
+            if content.is_empty() {
+                vec![]
+            } else {
+                vec![WireMessage {
+                    role: WireRole::User,
+                    content,
+                }]
+            }
+        }
+        Role::Assistant => {
+            let mut assistant_content = Vec::new();
+            let mut tool_results = Vec::new();
+            for part in parts {
+                match &part.body {
+                    PartBody::Text { text, .. } => {
+                        assistant_content.push(WireContent::Text { text: text.clone() })
+                    }
+                    PartBody::Tool {
+                        call_id,
+                        name,
+                        state,
+                    } => match state {
+                        ToolState::Completed { input, output, .. } => {
+                            assistant_content.push(WireContent::ToolCall {
+                                call_id: call_id.clone(),
+                                name: name.clone(),
+                                input: input.clone(),
+                            });
+                            tool_results.push(WireContent::ToolResult {
+                                call_id: call_id.clone(),
+                                output: output.clone(),
+                                is_error: false,
+                            });
+                        }
+                        ToolState::Error { input, error } => {
+                            assistant_content.push(WireContent::ToolCall {
+                                call_id: call_id.clone(),
+                                name: name.clone(),
+                                input: input.clone(),
+                            });
+                            tool_results.push(WireContent::ToolResult {
+                                call_id: call_id.clone(),
+                                output: error.clone(),
+                                is_error: true,
+                            });
+                        }
+                        ToolState::Running { input, .. } => {
+                            assistant_content.push(WireContent::ToolCall {
+                                call_id: call_id.clone(),
+                                name: name.clone(),
+                                input: input.clone(),
+                            });
+                        }
+                        ToolState::Pending { .. } => {}
+                    },
+                    _ => {}
+                }
+            }
+            let mut out = Vec::new();
+            if !assistant_content.is_empty() {
+                out.push(WireMessage {
+                    role: WireRole::Assistant,
+                    content: assistant_content,
+                });
+            }
+            if !tool_results.is_empty() {
+                out.push(WireMessage {
+                    role: WireRole::Tool,
+                    content: tool_results,
+                });
+            }
+            out
+        }
+    }
+}
+
+/// The outer loop: one call per session turn until the model stops asking for tool calls.
+/// `now_fn` supplies `created_at`/`StepContext::now` stamps (kept out of the loop body so
+/// tests can drive deterministic timestamps).
+pub async fn run_session(
+    provider: Arc<dyn Provider>,
+    mut ctx: StepContext,
+    run_config: &RunConfig,
+    now_fn: impl Fn() -> i64,
+) -> RunOutcome {
+    let mut doomloop = DoomLoopGuard::new();
+    let mut steps = 0u32;
+
+    loop {
+        match should_continue(&ctx.store, &ctx.session_id).await {
+            Ok(true) => {}
+            Ok(false) => return RunOutcome::Stopped,
+            Err(e) => {
+                return RunOutcome::Errored {
+                    message: e.to_string(),
+                }
+            }
+        }
+        if steps >= run_config.max_steps {
+            return RunOutcome::Stopped;
+        }
+        steps += 1;
+        ctx.now = now_fn();
+
+        let messages = match ctx.store.messages(ctx.session_id.clone()).await {
+            Ok(m) => m,
+            Err(e) => {
+                return RunOutcome::Errored {
+                    message: e.to_string(),
+                }
+            }
+        };
+        let mut wire_messages = Vec::new();
+        for message in &messages {
+            let parts = match ctx.store.parts(message.id.clone()).await {
+                Ok(p) => p,
+                Err(e) => {
+                    return RunOutcome::Errored {
+                        message: e.to_string(),
+                    }
+                }
+            };
+            wire_messages.extend(convert_message(message, &parts));
+        }
+
+        let system_blocks = system::assemble(
+            system::env_header(&ctx.cwd),
+            &run_config.agent_prompt,
+            &run_config.instructions,
+        );
+        let tools: Vec<ToolSchema> = ctx
+            .tools
+            .all()
+            .iter()
+            .map(|t| ToolSchema {
+                name: t.name().to_string(),
+                description: t.description().to_string(),
+                parameters: t.parameters(),
+            })
+            .collect();
+
+        let req = LlmRequest {
+            model: run_config.model.model_id.clone(),
+            system: system_blocks,
+            messages: wire_messages,
+            tools,
+            temperature: run_config.temperature,
+            max_tokens: None,
+            reasoning: None,
+            initiator: Initiator::User,
+        };
+
+        let stream_result = retry::with_retry(&ctx.cancel, || {
+            let provider = provider.clone();
+            let req = req.clone();
+            let cancel = ctx.cancel.child_token();
+            async move { provider.stream(req, cancel).await }
+        })
+        .await;
+
+        let stream = match stream_result {
+            Ok(s) => s,
+            Err(ProviderError::Cancelled) => return RunOutcome::Aborted,
+            Err(e) => {
+                return RunOutcome::Errored {
+                    message: e.to_string(),
+                }
+            }
+        };
+
+        let step = process_step(
+            stream,
+            &ctx,
+            run_config.model.clone(),
+            &run_config.agent_name,
+            run_config.cost,
+            &mut doomloop,
+        )
+        .await;
+
+        match step {
+            Ok(outcome) if outcome.aborted => {
+                accumulate_session_usage(&ctx, &outcome.usage, outcome.cost, now_fn()).await;
+                return RunOutcome::Aborted;
+            }
+            Ok(outcome) => {
+                accumulate_session_usage(&ctx, &outcome.usage, outcome.cost, now_fn()).await;
+                match outcome.result {
+                    StepResult::Continue => continue,
+                    StepResult::Stop => return RunOutcome::Stopped,
+                    StepResult::Compact => return RunOutcome::Stopped, // stub until M6
+                }
+            }
+            Err(step_err) if matches!(step_err.source, ProviderError::Cancelled) => {
+                return RunOutcome::Aborted;
+            }
+            Err(step_err) => {
+                return RunOutcome::Errored {
+                    message: step_err.source.to_string(),
+                };
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::VecDeque;
+    use std::sync::Mutex as StdMutex;
+
+    use async_trait::async_trait;
+    use tokio_util::sync::CancellationToken;
+
+    use super::*;
+    use crate::event::EventBus;
+    use crate::llm::{LlmEvent, LlmEventStream};
+    use crate::permission::{spawn_auto_approve, PermissionService};
+    use crate::store::Store;
+    use crate::tool::{Tool, ToolCtx, ToolError, ToolOutput, ToolRegistry};
+    use crate::types::{ModelInfo, ModelRef, Session, TokenUsage};
+
+    struct MockProvider {
+        steps: StdMutex<VecDeque<Vec<Result<LlmEvent, ProviderError>>>>,
+    }
+
+    impl MockProvider {
+        fn scripted(steps: Vec<Vec<Result<LlmEvent, ProviderError>>>) -> Self {
+            Self {
+                steps: StdMutex::new(steps.into_iter().collect()),
+            }
+        }
+    }
+
+    #[async_trait]
+    impl Provider for MockProvider {
+        fn id(&self) -> &str {
+            "mock"
+        }
+        async fn list_models(&self) -> Result<Vec<ModelInfo>, ProviderError> {
+            Ok(vec![])
+        }
+        async fn stream(
+            &self,
+            _req: LlmRequest,
+            _cancel: CancellationToken,
+        ) -> Result<LlmEventStream, ProviderError> {
+            let events = self
+                .steps
+                .lock()
+                .unwrap()
+                .pop_front()
+                .ok_or_else(|| ProviderError::Decode("no more scripted steps".into()))?;
+            Ok(Box::pin(futures::stream::iter(events)))
+        }
+    }
+
+    struct MockReadTool;
+
+    #[async_trait]
+    impl Tool for MockReadTool {
+        fn name(&self) -> &str {
+            "read"
+        }
+        fn description(&self) -> &str {
+            "reads a file"
+        }
+        fn parameters(&self) -> serde_json::Value {
+            serde_json::json!({"type": "object", "properties": {"file_path": {"type": "string"}}})
+        }
+        async fn execute(
+            &self,
+            _input: serde_json::Value,
+            _ctx: ToolCtx,
+        ) -> Result<ToolOutput, ToolError> {
+            Ok(ToolOutput::new("read foo.txt", "mock file content"))
+        }
+    }
+
+    fn usage(input: u64, output: u64) -> TokenUsage {
+        TokenUsage {
+            input,
+            output,
+            ..Default::default()
+        }
+    }
+
+    async fn make_ctx(
+        store: Store,
+        bus: EventBus,
+        session_id: SessionId,
+        cwd: std::path::PathBuf,
+    ) -> StepContext {
+        let mut tools = ToolRegistry::new();
+        tools.register(Arc::new(MockReadTool));
+        let permissions = Arc::new(PermissionService::new(bus.clone()));
+        spawn_auto_approve(bus.clone(), permissions.clone());
+        StepContext {
+            store,
+            bus,
+            tools,
+            permissions,
+            static_rules: Vec::new(),
+            extra_rules: Arc::new(std::sync::Mutex::new(Vec::new())),
+            session_id,
+            cwd: cwd.clone(),
+            data_dir: cwd.join("tool-output"),
+            cancel: CancellationToken::new(),
+            now: 1,
+        }
+    }
+
+    #[tokio::test]
+    async fn text_then_tool_call_then_final_text_persists_correctly() {
+        let store = Store::open_in_memory().unwrap();
+        let bus = EventBus::new();
+        let model = ModelRef::new("mock", "mock-model");
+        let session = Session::new_root("orchestrator", model.clone(), 1);
+        let session_id = session.id.clone();
+        store.upsert_session(session).await.unwrap();
+
+        let user_message = Message::new_user(session_id.clone(), 1);
+        store.upsert_message(user_message.clone()).await.unwrap();
+        store
+            .upsert_part(Part {
+                id: crate::types::PartId::new(),
+                message_id: user_message.id.clone(),
+                session_id: session_id.clone(),
+                idx: 0,
+                body: PartBody::Text {
+                    text: "please read foo.txt".into(),
+                    synthetic: false,
+                },
+            })
+            .await
+            .unwrap();
+
+        let provider = MockProvider::scripted(vec![
+            vec![
+                Ok(LlmEvent::TextStart { id: "t1".into() }),
+                Ok(LlmEvent::TextDelta {
+                    id: "t1".into(),
+                    text: "Let me check the file.".into(),
+                }),
+                Ok(LlmEvent::TextEnd { id: "t1".into() }),
+                Ok(LlmEvent::ToolCall {
+                    call_id: "call_1".into(),
+                    name: "read".into(),
+                    input: serde_json::json!({"file_path": "foo.txt"}),
+                }),
+                Ok(LlmEvent::Finish {
+                    reason: FinishReason::ToolCalls,
+                    usage: usage(10, 5),
+                }),
+            ],
+            vec![
+                Ok(LlmEvent::TextStart { id: "t2".into() }),
+                Ok(LlmEvent::TextDelta {
+                    id: "t2".into(),
+                    text: "The file contains: mock file content".into(),
+                }),
+                Ok(LlmEvent::TextEnd { id: "t2".into() }),
+                Ok(LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: usage(20, 8),
+                }),
+            ],
+        ]);
+
+        let cwd = tempfile::tempdir().unwrap();
+        let ctx = make_ctx(
+            store.clone(),
+            bus,
+            session_id.clone(),
+            cwd.path().to_path_buf(),
+        )
+        .await;
+        let run_config = RunConfig {
+            agent_name: "orchestrator".into(),
+            agent_prompt: "You are a helpful assistant.".into(),
+            model,
+            temperature: None,
+            max_steps: 10,
+            instructions: Vec::new(),
+            // $3/1M input, $15/1M output.
+            cost: Some(crate::types::ModelCost {
+                input: 3.0,
+                output: 15.0,
+                ..Default::default()
+            }),
+        };
+
+        let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;
+        assert!(matches!(outcome, RunOutcome::Stopped));
+
+        // Both steps' usage and cost accumulate onto the session.
+        let session = store.session(session_id.clone()).await.unwrap().unwrap();
+        assert_eq!(session.usage.input, 30);
+        assert_eq!(session.usage.output, 13);
+        // (10*3 + 5*15)/1e6 + (20*3 + 8*15)/1e6 = 0.000105 + 0.00018
+        assert!(
+            (session.cost - 0.000_285).abs() < 1e-9,
+            "cost = {}",
+            session.cost
+        );
+
+        let messages = store.messages(session_id.clone()).await.unwrap();
+        assert_eq!(
+            messages.len(),
+            3,
+            "user + tool-call assistant turn + final assistant turn"
+        );
+        assert_eq!(messages[0].role, Role::User);
+        assert_eq!(messages[1].role, Role::Assistant);
+        assert_eq!(messages[1].finished, Some(FinishReason::ToolCalls));
+        assert_eq!(messages[2].role, Role::Assistant);
+        assert_eq!(messages[2].finished, Some(FinishReason::Stop));
+
+        let first_parts = store.parts(messages[1].id.clone()).await.unwrap();
+        let tool_part = first_parts
+            .iter()
+            .find(|p| matches!(p.body, PartBody::Tool { .. }))
+            .expect("tool part persisted");
+        match &tool_part.body {
+            PartBody::Tool { name, state, .. } => {
+                assert_eq!(name, "read");
+                match state {
+                    ToolState::Completed { output, .. } => assert_eq!(output, "mock file content"),
+                    other => panic!("expected Completed tool state, got {other:?}"),
+                }
+            }
+            _ => unreachable!(),
+        }
+        let first_text = first_parts
+            .iter()
+            .find_map(|p| match &p.body {
+                PartBody::Text { text, .. } => Some(text.clone()),
+                _ => None,
+            })
+            .expect("text part persisted");
+        assert_eq!(first_text, "Let me check the file.");
+
+        let final_parts = store.parts(messages[2].id.clone()).await.unwrap();
+        let final_text = final_parts
+            .iter()
+            .find_map(|p| match &p.body {
+                PartBody::Text { text, .. } => Some(text.clone()),
+                _ => None,
+            })
+            .expect("final text part persisted");
+        assert_eq!(final_text, "The file contains: mock file content");
+    }
+
+    #[tokio::test]
+    async fn provider_error_on_first_event_is_reported_and_run_errors() {
+        let store = Store::open_in_memory().unwrap();
+        let bus = EventBus::new();
+        let model = ModelRef::new("mock", "mock-model");
+        let session = Session::new_root("orchestrator", model.clone(), 1);
+        let session_id = session.id.clone();
+        store.upsert_session(session).await.unwrap();
+
+        let user_message = Message::new_user(session_id.clone(), 1);
+        store.upsert_message(user_message.clone()).await.unwrap();
+        store
+            .upsert_part(Part {
+                id: crate::types::PartId::new(),
+                message_id: user_message.id.clone(),
+                session_id: session_id.clone(),
+                idx: 0,
+                body: PartBody::Text {
+                    text: "hi".into(),
+                    synthetic: false,
+                },
+            })
+            .await
+            .unwrap();
+
+        // ContextOverflow is never retried, so this should surface immediately as Errored.
+        let provider = MockProvider::scripted(vec![vec![Err(ProviderError::ContextOverflow)]]);
+
+        let cwd = tempfile::tempdir().unwrap();
+        let ctx = make_ctx(
+            store.clone(),
+            bus,
+            session_id.clone(),
+            cwd.path().to_path_buf(),
+        )
+        .await;
+        let run_config = RunConfig {
+            agent_name: "orchestrator".into(),
+            agent_prompt: "You are a helpful assistant.".into(),
+            model,
+            temperature: None,
+            max_steps: 10,
+            instructions: Vec::new(),
+            cost: None,
+        };
+
+        let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;
+        assert!(matches!(outcome, RunOutcome::Errored { .. }));
+
+        // No assistant message should have been created — the error hit before any event.
+        let messages = store.messages(session_id).await.unwrap();
+        assert_eq!(messages.len(), 1);
+    }
+}

crates/harness-core/src/engine/mod.rs

@@ -0,0 +1,10 @@
+pub mod doomloop;
+pub mod processor;
+pub mod retry;
+#[path = "loop.rs"]
+pub mod session_loop;
+pub mod system;
+
+pub use doomloop::DoomLoopGuard;
+pub use processor::{process_step, StepContext, StepError, StepOutcome, StepResult};
+pub use session_loop::{run_session, RunConfig};

crates/harness-core/src/engine/processor.rs

@@ -0,0 +1,665 @@
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+
+use futures::StreamExt;
+use tokio_util::sync::CancellationToken;
+
+use crate::event::{AppEvent, EventBus};
+use crate::llm::{FinishReason, LlmEvent, LlmEventStream, ProviderError};
+use crate::permission::{PermissionService, Ruleset};
+use crate::store::Store;
+use crate::tool::{MetadataSink, PermissionHandle, Tool, ToolCtx, ToolError, ToolRegistry};
+use crate::types::{Message, MessageId, Part, PartBody, PartId, SessionId, TokenUsage, ToolState};
+
+use super::doomloop::DoomLoopGuard;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum StepResult {
+    Continue,
+    Stop,
+    Compact,
+}
+
+pub struct StepOutcome {
+    pub result: StepResult,
+    pub message_id: Option<MessageId>,
+    pub usage: TokenUsage,
+    /// Dollar cost of this step's usage (0.0 when no pricing is available).
+    pub cost: f64,
+    pub aborted: bool,
+}
+
+pub struct StepError {
+    pub source: ProviderError,
+    /// If `true`, the assistant message/parts were already persisted for this step —
+    /// the outer loop must surface this as a terminal error, not retry.
+    pub any_persisted: bool,
+}
+
+/// Everything a step needs that stays constant across the run (cheap to construct per step;
+/// owns clones/handles, not the session itself, so the loop keeps ownership of `Session`).
+pub struct StepContext {
+    pub store: Store,
+    pub bus: EventBus,
+    pub tools: ToolRegistry,
+    pub permissions: Arc<PermissionService>,
+    pub static_rules: Ruleset,
+    pub extra_rules: Arc<Mutex<Ruleset>>,
+    pub session_id: SessionId,
+    pub cwd: PathBuf,
+    /// Session's `tool-output` spill directory (see `tool::truncate`).
+    pub data_dir: PathBuf,
+    pub cancel: CancellationToken,
+    /// Wall-clock for `created_at` stamps — passed in so tests stay deterministic.
+    pub now: i64,
+}
+
+struct FlushTracker {
+    last_flush: Instant,
+    bytes_since_flush: usize,
+}
+
+impl FlushTracker {
+    fn new() -> Self {
+        Self {
+            last_flush: Instant::now(),
+            bytes_since_flush: 0,
+        }
+    }
+
+    /// docs/02-engine.md: flush to store every 50ms or 2KB of buffered delta text.
+    fn should_flush(&mut self, new_bytes: usize) -> bool {
+        self.bytes_since_flush += new_bytes;
+        if self.bytes_since_flush >= 2048 || self.last_flush.elapsed() >= Duration::from_millis(50)
+        {
+            self.bytes_since_flush = 0;
+            self.last_flush = Instant::now();
+            true
+        } else {
+            false
+        }
+    }
+}
+
+struct Run<'a> {
+    ctx: &'a StepContext,
+    assistant: Option<Message>,
+    next_idx: u32,
+    active_text: Option<PartId>,
+    active_reasoning: Option<PartId>,
+    pending_tools: HashMap<String, PartId>,
+    text_buf: HashMap<PartId, String>,
+    reasoning_buf: HashMap<PartId, String>,
+    tool_input_buf: HashMap<String, String>,
+    flushers: HashMap<PartId, FlushTracker>,
+    parts_by_id: HashMap<PartId, Part>,
+}
+
+impl<'a> Run<'a> {
+    fn new(ctx: &'a StepContext) -> Self {
+        Self {
+            ctx,
+            assistant: None,
+            next_idx: 0,
+            active_text: None,
+            active_reasoning: None,
+            pending_tools: HashMap::new(),
+            text_buf: HashMap::new(),
+            reasoning_buf: HashMap::new(),
+            tool_input_buf: HashMap::new(),
+            flushers: HashMap::new(),
+            parts_by_id: HashMap::new(),
+        }
+    }
+
+    fn message_id(&self) -> MessageId {
+        self.assistant
+            .as_ref()
+            .expect("assistant created")
+            .id
+            .clone()
+    }
+
+    async fn ensure_assistant(
+        &mut self,
+        model: crate::types::ModelRef,
+        agent: &str,
+    ) -> Result<(), ProviderError> {
+        if self.assistant.is_some() {
+            return Ok(());
+        }
+        let message =
+            Message::new_assistant(self.ctx.session_id.clone(), model, agent, self.ctx.now);
+        self.ctx
+            .store
+            .upsert_message(message.clone())
+            .await
+            .map_err(|e| ProviderError::Decode(e.to_string()))?;
+        self.ctx.bus.publish(AppEvent::MessageCreated {
+            message: message.clone(),
+        });
+
+        let step_start = Part {
+            id: PartId::new(),
+            message_id: message.id.clone(),
+            session_id: self.ctx.session_id.clone(),
+            idx: self.next_idx,
+            body: PartBody::StepStart,
+        };
+        self.next_idx += 1;
+        self.persist_part(step_start).await?;
+
+        self.assistant = Some(message);
+        Ok(())
+    }
+
+    async fn persist_part(&mut self, part: Part) -> Result<(), ProviderError> {
+        self.ctx
+            .store
+            .upsert_part(part.clone())
+            .await
+            .map_err(|e| ProviderError::Decode(e.to_string()))?;
+        self.ctx
+            .bus
+            .publish(AppEvent::PartUpdated { part: part.clone() });
+        self.parts_by_id.insert(part.id.clone(), part);
+        Ok(())
+    }
+
+    async fn on_text_start(&mut self, id: String) -> Result<(), ProviderError> {
+        let part_id = PartId::from(id);
+        let part = Part {
+            id: part_id.clone(),
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: self.next_idx,
+            body: PartBody::Text {
+                text: String::new(),
+                synthetic: false,
+            },
+        };
+        self.next_idx += 1;
+        self.text_buf.insert(part_id.clone(), String::new());
+        self.flushers.insert(part_id.clone(), FlushTracker::new());
+        self.active_text = Some(part_id);
+        self.persist_part(part).await
+    }
+
+    async fn on_text_delta(&mut self, id: String, text: String) -> Result<(), ProviderError> {
+        let part_id = PartId::from(id);
+        let buf = self.text_buf.entry(part_id.clone()).or_default();
+        buf.push_str(&text);
+        let full_text = buf.clone();
+
+        self.ctx.bus.publish(AppEvent::PartDelta {
+            part_id: part_id.clone(),
+            message_id: self.message_id(),
+            delta: text.clone(),
+        });
+
+        let should_flush = self
+            .flushers
+            .entry(part_id.clone())
+            .or_insert_with(FlushTracker::new)
+            .should_flush(text.len());
+        if should_flush {
+            let part = Part {
+                id: part_id,
+                message_id: self.message_id(),
+                session_id: self.ctx.session_id.clone(),
+                idx: 0, // overwritten below from cached idx
+                body: PartBody::Text {
+                    text: full_text,
+                    synthetic: false,
+                },
+            };
+            self.persist_existing(part).await?;
+        }
+        Ok(())
+    }
+
+    async fn on_text_end(&mut self, id: String) -> Result<(), ProviderError> {
+        let part_id = PartId::from(id);
+        let full_text = self.text_buf.get(&part_id).cloned().unwrap_or_default();
+        let part = Part {
+            id: part_id,
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: 0,
+            body: PartBody::Text {
+                text: full_text,
+                synthetic: false,
+            },
+        };
+        self.active_text = None;
+        self.persist_existing(part).await
+    }
+
+    async fn on_reasoning_start(&mut self, id: String) -> Result<(), ProviderError> {
+        let part_id = PartId::from(id);
+        let part = Part {
+            id: part_id.clone(),
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: self.next_idx,
+            body: PartBody::Reasoning {
+                text: String::new(),
+                signature: None,
+            },
+        };
+        self.next_idx += 1;
+        self.reasoning_buf.insert(part_id.clone(), String::new());
+        self.flushers.insert(part_id.clone(), FlushTracker::new());
+        self.active_reasoning = Some(part_id);
+        self.persist_part(part).await
+    }
+
+    async fn on_reasoning_delta(&mut self, id: String, text: String) -> Result<(), ProviderError> {
+        let part_id = PartId::from(id);
+        let buf = self.reasoning_buf.entry(part_id.clone()).or_default();
+        buf.push_str(&text);
+        let full_text = buf.clone();
+
+        self.ctx.bus.publish(AppEvent::PartDelta {
+            part_id: part_id.clone(),
+            message_id: self.message_id(),
+            delta: text.clone(),
+        });
+
+        let should_flush = self
+            .flushers
+            .entry(part_id.clone())
+            .or_insert_with(FlushTracker::new)
+            .should_flush(text.len());
+        if should_flush {
+            let part = Part {
+                id: part_id,
+                message_id: self.message_id(),
+                session_id: self.ctx.session_id.clone(),
+                idx: 0,
+                body: PartBody::Reasoning {
+                    text: full_text,
+                    signature: None,
+                },
+            };
+            self.persist_existing(part).await?;
+        }
+        Ok(())
+    }
+
+    async fn on_reasoning_end(
+        &mut self,
+        id: String,
+        signature: Option<String>,
+    ) -> Result<(), ProviderError> {
+        let part_id = PartId::from(id);
+        let full_text = self
+            .reasoning_buf
+            .get(&part_id)
+            .cloned()
+            .unwrap_or_default();
+        let part = Part {
+            id: part_id,
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: 0,
+            body: PartBody::Reasoning {
+                text: full_text,
+                signature,
+            },
+        };
+        self.active_reasoning = None;
+        self.persist_existing(part).await
+    }
+
+    /// Re-persists a part that already exists, preserving its original `idx`.
+    async fn persist_existing(&mut self, mut part: Part) -> Result<(), ProviderError> {
+        if let Some(existing) = self.parts_by_id.get(&part.id) {
+            part.idx = existing.idx;
+        }
+        self.persist_part(part).await
+    }
+
+    async fn on_tool_input_start(
+        &mut self,
+        call_id: String,
+        name: String,
+    ) -> Result<(), ProviderError> {
+        let part_id = PartId::new();
+        self.pending_tools.insert(call_id.clone(), part_id.clone());
+        self.tool_input_buf.insert(call_id.clone(), String::new());
+        let part = Part {
+            id: part_id,
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: self.next_idx,
+            body: PartBody::Tool {
+                call_id,
+                name,
+                state: ToolState::Pending {
+                    partial_input: String::new(),
+                },
+            },
+        };
+        self.next_idx += 1;
+        self.persist_part(part).await
+    }
+
+    async fn on_tool_input_delta(
+        &mut self,
+        call_id: String,
+        json: String,
+    ) -> Result<(), ProviderError> {
+        let buf = self.tool_input_buf.entry(call_id.clone()).or_default();
+        buf.push_str(&json);
+        if let Some(part_id) = self.pending_tools.get(&call_id).cloned() {
+            self.ctx.bus.publish(AppEvent::PartDelta {
+                part_id,
+                message_id: self.message_id(),
+                delta: json,
+            });
+        }
+        Ok(())
+    }
+
+    async fn on_tool_call(
+        &mut self,
+        call_id: String,
+        name: String,
+        input: serde_json::Value,
+        doomloop: &mut DoomLoopGuard,
+    ) -> Result<(), ProviderError> {
+        let part_id = self.pending_tools.remove(&call_id).unwrap_or_default();
+        let running = Part {
+            id: part_id.clone(),
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: self.next_idx_for(&part_id),
+            body: PartBody::Tool {
+                call_id: call_id.clone(),
+                name: name.clone(),
+                state: ToolState::Running {
+                    input: input.clone(),
+                    title: None,
+                    metadata: serde_json::Value::Null,
+                },
+            },
+        };
+        self.persist_existing(running).await?;
+
+        let mut doom_blocked = None;
+        if doomloop.record_and_check(&name, &input) {
+            let cancel = self.ctx.cancel.child_token();
+            if let Err(e) = self
+                .ctx
+                .permissions
+                .force_ask(
+                    &self.ctx.session_id,
+                    crate::permission::AskInput {
+                        permission: "doom_loop".into(),
+                        pattern: name.clone(),
+                        always_pattern: name.clone(),
+                        metadata: serde_json::json!({"tool": name, "input": input}),
+                    },
+                    &cancel,
+                )
+                .await
+            {
+                doom_blocked = Some(format!("model appears stuck repeating {name}: {e}"));
+            }
+        }
+
+        let tool = self.ctx.tools.get(&name);
+        let (title, output, metadata, is_error) = if let Some(reason) = doom_blocked {
+            ("error".to_string(), reason, serde_json::Value::Null, true)
+        } else {
+            match tool {
+                None => (
+                    "error".to_string(),
+                    format!("unknown tool: {name}"),
+                    serde_json::Value::Null,
+                    true,
+                ),
+                Some(tool) => self.run_tool(tool, call_id.clone(), input.clone()).await,
+            }
+        };
+
+        // docs/05-tools.md: cap tool output at 30k chars regardless of which tool produced it.
+        let output = crate::tool::truncate::truncate(&output, &self.ctx.data_dir, &call_id)
+            .map(|t| t.text)
+            .unwrap_or(output);
+
+        let final_state = if is_error {
+            ToolState::Error {
+                input,
+                error: output.clone(),
+            }
+        } else {
+            ToolState::Completed {
+                input,
+                title: title.clone(),
+                output: output.clone(),
+                metadata,
+                duration_ms: 0,
+            }
+        };
+        let final_part = Part {
+            id: part_id,
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: 0,
+            body: PartBody::Tool {
+                call_id,
+                name,
+                state: final_state,
+            },
+        };
+        self.persist_existing(final_part).await
+    }
+
+    fn next_idx_for(&mut self, part_id: &PartId) -> u32 {
+        if let Some(existing) = self.parts_by_id.get(part_id) {
+            existing.idx
+        } else {
+            let idx = self.next_idx;
+            self.next_idx += 1;
+            idx
+        }
+    }
+
+    async fn run_tool(
+        &self,
+        tool: Arc<dyn Tool>,
+        call_id: String,
+        input: serde_json::Value,
+    ) -> (String, String, serde_json::Value, bool) {
+        let call_cancel = self.ctx.cancel.child_token();
+        let (metadata_sink, _metadata_rx) = MetadataSink::channel();
+        let ask = PermissionHandle::new(
+            self.ctx.permissions.clone(),
+            self.ctx.session_id.clone(),
+            self.ctx.static_rules.clone(),
+            self.ctx.extra_rules.clone(),
+            call_cancel.clone(),
+        );
+        let tool_ctx = ToolCtx {
+            session_id: self.ctx.session_id.clone(),
+            message_id: self.message_id(),
+            call_id: call_id.clone(),
+            cwd: self.ctx.cwd.clone(),
+            data_dir: self.ctx.data_dir.clone(),
+            cancel: call_cancel.clone(),
+            ask,
+            metadata: metadata_sink,
+        };
+
+        let result = tokio::select! {
+            res = tool.execute(input, tool_ctx) => res,
+            _ = call_cancel.cancelled() => Err(ToolError::Cancelled),
+        };
+
+        match result {
+            Ok(output) => (output.title, output.output, output.metadata, false),
+            Err(err) => (
+                "error".to_string(),
+                err.to_string(),
+                serde_json::Value::Null,
+                true,
+            ),
+        }
+    }
+
+    async fn on_finish(
+        &mut self,
+        reason: FinishReason,
+        usage: TokenUsage,
+        cost: f64,
+    ) -> Result<StepResult, ProviderError> {
+        let part = Part {
+            id: PartId::new(),
+            message_id: self.message_id(),
+            session_id: self.ctx.session_id.clone(),
+            idx: self.next_idx,
+            body: PartBody::StepFinish {
+                usage,
+                cost,
+                reason: reason.clone(),
+            },
+        };
+        self.next_idx += 1;
+        self.persist_part(part).await?;
+
+        let result = match reason {
+            FinishReason::ToolCalls => StepResult::Continue,
+            _ => StepResult::Stop,
+        };
+
+        if let Some(message) = self.assistant.as_mut() {
+            message.finished = Some(reason);
+            message.usage = usage;
+            self.ctx
+                .store
+                .upsert_message(message.clone())
+                .await
+                .map_err(|e| ProviderError::Decode(e.to_string()))?;
+            self.ctx.bus.publish(AppEvent::MessageUpdated {
+                message: message.clone(),
+            });
+        }
+        Ok(result)
+    }
+
+    async fn mark_errored(&mut self, source: &ProviderError) {
+        if let Some(message) = self.assistant.as_mut() {
+            message.error = Some(crate::types::MessageError::Provider(source.to_string()));
+            if self.ctx.store.upsert_message(message.clone()).await.is_ok() {
+                self.ctx.bus.publish(AppEvent::MessageUpdated {
+                    message: message.clone(),
+                });
+            }
+        }
+    }
+}
+
+/// Consumes one provider stream end-to-end: persists parts/messages as events arrive and
+/// executes tool calls inline, in stream order (opencode's processor.ts semantics).
+pub async fn process_step(
+    mut stream: LlmEventStream,
+    ctx: &StepContext,
+    model: crate::types::ModelRef,
+    agent: &str,
+    cost: Option<crate::types::ModelCost>,
+    doomloop: &mut DoomLoopGuard,
+) -> Result<StepOutcome, StepError> {
+    let mut run = Run::new(ctx);
+    let mut usage = TokenUsage::default();
+    let mut step_cost = 0.0;
+    let mut result = StepResult::Stop;
+
+    loop {
+        let item = tokio::select! {
+            item = stream.next() => item,
+            _ = ctx.cancel.cancelled() => {
+                run.mark_errored(&ProviderError::Cancelled).await;
+                return Ok(StepOutcome {
+                    result: StepResult::Stop,
+                    message_id: run.assistant.as_ref().map(|m| m.id.clone()),
+                    usage,
+                    cost: step_cost,
+                    aborted: true,
+                });
+            }
+        };
+        let Some(item) = item else { break };
+
+        let event = match item {
+            Ok(event) => event,
+            Err(source) => {
+                let any_persisted = run.assistant.is_some();
+                run.mark_errored(&source).await;
+                return Err(StepError {
+                    source,
+                    any_persisted,
+                });
+            }
+        };
+
+        run.ensure_assistant(model.clone(), agent)
+            .await
+            .map_err(|source| StepError {
+                source,
+                any_persisted: false,
+            })?;
+
+        let outcome = match event {
+            LlmEvent::TextStart { id } => run.on_text_start(id).await,
+            LlmEvent::TextDelta { id, text } => run.on_text_delta(id, text).await,
+            LlmEvent::TextEnd { id } => run.on_text_end(id).await,
+            LlmEvent::ReasoningStart { id } => run.on_reasoning_start(id).await,
+            LlmEvent::ReasoningDelta { id, text } => run.on_reasoning_delta(id, text).await,
+            LlmEvent::ReasoningEnd { id, signature } => run.on_reasoning_end(id, signature).await,
+            LlmEvent::ToolInputStart { call_id, name } => {
+                run.on_tool_input_start(call_id, name).await
+            }
+            LlmEvent::ToolInputDelta { call_id, json } => {
+                run.on_tool_input_delta(call_id, json).await
+            }
+            LlmEvent::ToolCall {
+                call_id,
+                name,
+                input,
+            } => run.on_tool_call(call_id, name, input, doomloop).await,
+            LlmEvent::Finish {
+                reason,
+                usage: finish_usage,
+            } => {
+                usage = finish_usage;
+                step_cost = cost.map(|c| c.cost_of(&finish_usage)).unwrap_or(0.0);
+                match run.on_finish(reason, finish_usage, step_cost).await {
+                    Ok(r) => {
+                        result = r;
+                        Ok(())
+                    }
+                    Err(e) => Err(e),
+                }
+            }
+        };
+
+        if let Err(source) = outcome {
+            return Err(StepError {
+                source,
+                any_persisted: true,
+            });
+        }
+    }
+
+    Ok(StepOutcome {
+        result,
+        message_id: run.assistant.as_ref().map(|m| m.id.clone()),
+        usage,
+        cost: step_cost,
+        aborted: false,
+    })
+}

crates/harness-core/src/engine/retry.rs

@@ -0,0 +1,113 @@
+use std::future::Future;
+use std::time::Duration;
+
+use tokio_util::sync::CancellationToken;
+
+use crate::llm::ProviderError;
+
+const MAX_ATTEMPTS: u32 = 8;
+
+fn delay_for(err: &ProviderError, attempt: u32) -> Duration {
+    if let ProviderError::RateLimited {
+        retry_after: Some(d),
+    } = err
+    {
+        return *d;
+    }
+    let secs = 2u64.saturating_mul(1u64 << attempt.min(4));
+    Duration::from_secs(secs.min(30))
+}
+
+/// Wraps only the *start* of a step (acquiring the provider stream) in retry — once any
+/// event has been consumed from the stream we surface the error instead (opencode behavior).
+///
+/// `ContextOverflow` is never retried; the caller reacts with `StepResult::Compact`.
+pub async fn with_retry<F, Fut, T>(
+    cancel: &CancellationToken,
+    mut attempt_fn: F,
+) -> Result<T, ProviderError>
+where
+    F: FnMut() -> Fut,
+    Fut: Future<Output = Result<T, ProviderError>>,
+{
+    let mut attempt = 0u32;
+    loop {
+        match attempt_fn().await {
+            Ok(v) => return Ok(v),
+            Err(e) if e.is_retryable() && attempt + 1 < MAX_ATTEMPTS => {
+                let delay = delay_for(&e, attempt);
+                attempt += 1;
+                tokio::select! {
+                    _ = tokio::time::sleep(delay) => {}
+                    _ = cancel.cancelled() => return Err(ProviderError::Cancelled),
+                }
+            }
+            Err(e) => return Err(e),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::sync::atomic::{AtomicU32, Ordering};
+
+    #[tokio::test(start_paused = true)]
+    async fn retries_rate_limited_until_success() {
+        let calls = AtomicU32::new(0);
+        let cancel = CancellationToken::new();
+        let result = with_retry(&cancel, || {
+            let n = calls.fetch_add(1, Ordering::SeqCst);
+            async move {
+                if n < 2 {
+                    Err(ProviderError::RateLimited { retry_after: None })
+                } else {
+                    Ok(42)
+                }
+            }
+        })
+        .await;
+        assert_eq!(result.unwrap(), 42);
+        assert_eq!(calls.load(Ordering::SeqCst), 3);
+    }
+
+    #[tokio::test]
+    async fn does_not_retry_context_overflow() {
+        let calls = AtomicU32::new(0);
+        let cancel = CancellationToken::new();
+        let result: Result<(), _> = with_retry(&cancel, || {
+            calls.fetch_add(1, Ordering::SeqCst);
+            async move { Err(ProviderError::ContextOverflow) }
+        })
+        .await;
+        assert!(matches!(result, Err(ProviderError::ContextOverflow)));
+        assert_eq!(calls.load(Ordering::SeqCst), 1);
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn gives_up_after_max_attempts() {
+        let calls = AtomicU32::new(0);
+        let cancel = CancellationToken::new();
+        let result: Result<(), _> = with_retry(&cancel, || {
+            calls.fetch_add(1, Ordering::SeqCst);
+            async move { Err(ProviderError::Overloaded) }
+        })
+        .await;
+        assert!(result.is_err());
+        assert_eq!(calls.load(Ordering::SeqCst), MAX_ATTEMPTS);
+    }
+
+    #[tokio::test]
+    async fn cancellation_aborts_the_wait_between_retries() {
+        let calls = AtomicU32::new(0);
+        let cancel = CancellationToken::new();
+        cancel.cancel();
+        let result: Result<(), _> = with_retry(&cancel, || {
+            calls.fetch_add(1, Ordering::SeqCst);
+            async move { Err(ProviderError::Overloaded) }
+        })
+        .await;
+        assert!(matches!(result, Err(ProviderError::Cancelled)));
+        assert_eq!(calls.load(Ordering::SeqCst), 1);
+    }
+}

crates/harness-core/src/engine/system.rs

@@ -0,0 +1,56 @@
+use std::path::Path;
+
+/// cwd, platform, date, best-effort git status. Kept as its own block (not merged into the
+/// agent prompt) because Anthropic cache breakpoints need the system prompt structured as
+/// separate blocks (see `LlmRequest.system: Vec<String>`).
+pub fn env_header(cwd: &Path) -> String {
+    let git_status = std::process::Command::new("git")
+        .arg("status")
+        .arg("--short")
+        .current_dir(cwd)
+        .output()
+        .ok()
+        .filter(|o| o.status.success())
+        .map(|o| String::from_utf8_lossy(&o.stdout).trim().to_string())
+        .filter(|s| !s.is_empty());
+
+    let mut header = format!(
+        "Working directory: {}\nPlatform: {}\n",
+        cwd.display(),
+        std::env::consts::OS
+    );
+    if let Some(status) = git_status {
+        header.push_str(&format!("Git status:\n{status}\n"));
+    }
+    header
+}
+
+/// Ordered system prompt blocks: environment header, agent prompt, then project
+/// instructions (e.g. AGENTS.md contents). Order matches `02-engine.md`.
+pub fn assemble(env_header: String, agent_prompt: &str, instructions: &[String]) -> Vec<String> {
+    let mut blocks = vec![env_header, agent_prompt.to_string()];
+    blocks.extend(instructions.iter().cloned());
+    blocks
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn env_header_includes_cwd_and_platform() {
+        let header = env_header(Path::new("."));
+        assert!(header.contains("Working directory:"));
+        assert!(header.contains(std::env::consts::OS));
+    }
+
+    #[test]
+    fn assemble_orders_env_agent_then_instructions() {
+        let blocks = assemble(
+            "ENV".to_string(),
+            "AGENT",
+            &["AGENTS.md contents".to_string()],
+        );
+        assert_eq!(blocks, vec!["ENV", "AGENT", "AGENTS.md contents"]);
+    }
+}

crates/harness-core/src/event.rs

@@ -24,7 +24,8 @@ pub struct PermissionRequest {
     pub id: String,
     pub session_id: SessionId,
     pub permission: String,
-    pub patterns: Vec<String>,
+    pub pattern: String,
+    pub always_pattern: String,
     pub metadata: serde_json::Value,
 }
 

crates/harness-core/src/lib.rs

@@ -1,7 +1,10 @@
+pub mod config;
+pub mod engine;
 pub mod event;
 pub mod llm;
 pub mod permission;
 pub mod store;
+pub mod tool;
 pub mod types;
 
 pub use event::{AppEvent, EventBus};

crates/harness-core/src/llm.rs

@@ -1,7 +1,13 @@
+use std::pin::Pin;
+use std::time::Duration;
+
+use async_trait::async_trait;
+use futures::Stream;
 use serde::{Deserialize, Serialize};
+use tokio_util::sync::CancellationToken;
+
+use crate::types::{ModelInfo, TokenUsage};
 
-// Full LlmEvent/LlmRequest/Provider trait land in M1 alongside harness-providers.
-// FinishReason lives here (not types/) because it's part of the provider-facing vocabulary.
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub enum FinishReason {
     Stop,
@@ -10,3 +16,185 @@ pub enum FinishReason {
     ContentFilter,
     Unknown(String),
 }
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub enum LlmEvent {
+    TextStart {
+        id: String,
+    },
+    TextDelta {
+        id: String,
+        text: String,
+    },
+    TextEnd {
+        id: String,
+    },
+    ReasoningStart {
+        id: String,
+    },
+    ReasoningDelta {
+        id: String,
+        text: String,
+    },
+    ReasoningEnd {
+        id: String,
+        signature: Option<String>,
+    },
+    ToolInputStart {
+        call_id: String,
+        name: String,
+    },
+    ToolInputDelta {
+        call_id: String,
+        json: String,
+    },
+    ToolCall {
+        call_id: String,
+        name: String,
+        input: serde_json::Value,
+    },
+    Finish {
+        reason: FinishReason,
+        usage: TokenUsage,
+    },
+}
+
+#[derive(Debug, Clone)]
+pub enum ProviderError {
+    RateLimited {
+        retry_after: Option<Duration>,
+    },
+    Overloaded,
+    /// Never retried — the engine reacts with `StepResult::Compact`.
+    ContextOverflow,
+    Auth(String),
+    Http {
+        status: u16,
+        body: String,
+    },
+    Network(String),
+    Decode(String),
+    Cancelled,
+}
+
+impl std::fmt::Display for ProviderError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            ProviderError::RateLimited { retry_after } => {
+                write!(f, "rate limited (retry_after={retry_after:?})")
+            }
+            ProviderError::Overloaded => write!(f, "provider overloaded"),
+            ProviderError::ContextOverflow => write!(f, "context window exceeded"),
+            ProviderError::Auth(msg) => write!(f, "auth error: {msg}"),
+            ProviderError::Http { status, body } => write!(f, "http {status}: {body}"),
+            ProviderError::Network(msg) => write!(f, "network error: {msg}"),
+            ProviderError::Decode(msg) => write!(f, "decode error: {msg}"),
+            ProviderError::Cancelled => write!(f, "cancelled"),
+        }
+    }
+}
+
+impl std::error::Error for ProviderError {}
+
+impl ProviderError {
+    /// Retry policy classification (`engine/retry.rs`): 5xx counts as `Http`, so check status.
+    pub fn is_retryable(&self) -> bool {
+        matches!(
+            self,
+            ProviderError::RateLimited { .. }
+                | ProviderError::Overloaded
+                | ProviderError::Network(_)
+        ) || matches!(self, ProviderError::Http { status, .. } if *status >= 500)
+    }
+}
+
+pub type LlmEventStream = Pin<Box<dyn Stream<Item = Result<LlmEvent, ProviderError>> + Send>>;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum Role {
+    System,
+    User,
+    Assistant,
+    Tool,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum WireContent {
+    Text {
+        text: String,
+    },
+    ToolCall {
+        call_id: String,
+        name: String,
+        input: serde_json::Value,
+    },
+    ToolResult {
+        call_id: String,
+        output: String,
+        is_error: bool,
+    },
+    Image {
+        mime_type: String,
+        data: String,
+    },
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct WireMessage {
+    pub role: Role,
+    pub content: Vec<WireContent>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ToolSchema {
+    pub name: String,
+    pub description: String,
+    pub parameters: serde_json::Value,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum ReasoningEffort {
+    Low,
+    Medium,
+    High,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ReasoningOpts {
+    pub effort: Option<ReasoningEffort>,
+    /// Anthropic extended-thinking token budget.
+    pub budget_tokens: Option<u32>,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum Initiator {
+    User,
+    Agent,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct LlmRequest {
+    pub model: String,
+    pub system: Vec<String>,
+    pub messages: Vec<WireMessage>,
+    pub tools: Vec<ToolSchema>,
+    pub temperature: Option<f32>,
+    pub max_tokens: Option<u32>,
+    pub reasoning: Option<ReasoningOpts>,
+    pub initiator: Initiator,
+}
+
+#[async_trait]
+pub trait Provider: Send + Sync {
+    fn id(&self) -> &str;
+    async fn list_models(&self) -> Result<Vec<ModelInfo>, ProviderError>;
+    async fn stream(
+        &self,
+        req: LlmRequest,
+        cancel: CancellationToken,
+    ) -> Result<LlmEventStream, ProviderError>;
+}

crates/harness-core/src/permission/mod.rs

@@ -1,3 +1,7 @@
 pub mod rule;
+pub mod service;
 
 pub use rule::{evaluate, Action, Rule, Ruleset};
+pub use service::{
+    spawn_auto_approve, AskDecision, AskError, AskInput, PermissionReply, PermissionService,
+};

crates/harness-core/src/permission/service.rs

@@ -0,0 +1,359 @@
+use std::collections::HashMap;
+use std::sync::Mutex;
+
+use tokio::sync::oneshot;
+use tokio_util::sync::CancellationToken;
+use ulid::Ulid;
+
+use crate::event::{AppEvent, EventBus, PermissionRequest};
+use crate::types::SessionId;
+
+use super::rule::{evaluate, Action, Rule, Ruleset};
+
+pub struct AskInput {
+    pub permission: String,
+    /// Pattern evaluated against the ruleset stack, and granted on a `Once` reply.
+    pub pattern: String,
+    /// Coarser pattern granted (as an `Allow` rule on `session.extra_rules`) on an `Always` reply.
+    pub always_pattern: String,
+    pub metadata: serde_json::Value,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum PermissionReply {
+    Once,
+    Always,
+    Reject,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum AskError {
+    #[error("permission denied by ruleset")]
+    Denied,
+    #[error("permission rejected by user")]
+    Rejected,
+    #[error("cancelled")]
+    Cancelled,
+}
+
+#[derive(Debug)]
+pub enum AskDecision {
+    /// Ruleset allowed it outright, or the user replied `Once`.
+    Allowed,
+    /// User replied `Always` — caller persists this rule onto `session.extra_rules`.
+    AllowedAlways(Rule),
+}
+
+/// Port of opencode's permission/index.ts ask flow: evaluate the ruleset stack first;
+/// only fall through to a blocking, event-published ask when the verdict is `Ask`.
+pub struct PermissionService {
+    bus: EventBus,
+    pending: Mutex<HashMap<String, oneshot::Sender<PermissionReply>>>,
+}
+
+impl PermissionService {
+    pub fn new(bus: EventBus) -> Self {
+        Self {
+            bus,
+            pending: Mutex::new(HashMap::new()),
+        }
+    }
+
+    pub async fn ask(
+        &self,
+        session_id: &SessionId,
+        stack: &[&Ruleset],
+        input: AskInput,
+        cancel: &CancellationToken,
+    ) -> Result<AskDecision, AskError> {
+        match evaluate(stack, &input.permission, &input.pattern) {
+            Action::Allow => Ok(AskDecision::Allowed),
+            Action::Deny => Err(AskError::Denied),
+            Action::Ask => self.ask_user(session_id, input, cancel).await,
+        }
+    }
+
+    /// Bypasses ruleset evaluation entirely — used by the doom-loop guard, which must ask
+    /// regardless of any `Allow` rule.
+    pub async fn force_ask(
+        &self,
+        session_id: &SessionId,
+        input: AskInput,
+        cancel: &CancellationToken,
+    ) -> Result<AskDecision, AskError> {
+        self.ask_user(session_id, input, cancel).await
+    }
+
+    async fn ask_user(
+        &self,
+        session_id: &SessionId,
+        input: AskInput,
+        cancel: &CancellationToken,
+    ) -> Result<AskDecision, AskError> {
+        let id = Ulid::new().to_string();
+        let (tx, rx) = oneshot::channel();
+        self.pending.lock().unwrap().insert(id.clone(), tx);
+
+        self.bus.publish(AppEvent::PermissionAsked {
+            request: PermissionRequest {
+                id: id.clone(),
+                session_id: session_id.clone(),
+                permission: input.permission.clone(),
+                pattern: input.pattern.clone(),
+                always_pattern: input.always_pattern.clone(),
+                metadata: input.metadata.clone(),
+            },
+        });
+
+        let reply = tokio::select! {
+            reply = rx => reply.map_err(|_| AskError::Cancelled)?,
+            _ = cancel.cancelled() => {
+                self.pending.lock().unwrap().remove(&id);
+                return Err(AskError::Cancelled);
+            }
+        };
+
+        self.bus.publish(AppEvent::PermissionResolved { id });
+
+        match reply {
+            PermissionReply::Once => Ok(AskDecision::Allowed),
+            PermissionReply::Always => Ok(AskDecision::AllowedAlways(Rule {
+                permission: input.permission,
+                pattern: input.always_pattern,
+                action: Action::Allow,
+            })),
+            PermissionReply::Reject => Err(AskError::Rejected),
+        }
+    }
+
+    /// Resolve a pending ask (called by a frontend: TUI keypress, auto-approve stub, ...).
+    /// Returns `false` if `id` had already been resolved or never existed.
+    pub fn reply(&self, id: &str, reply: PermissionReply) -> bool {
+        match self.pending.lock().unwrap().remove(id) {
+            Some(tx) => tx.send(reply).is_ok(),
+            None => false,
+        }
+    }
+}
+
+/// Auto-approves every ask as `Once` — the M1..M2 stand-in for a real TUI permission modal.
+pub fn spawn_auto_approve(bus: EventBus, service: std::sync::Arc<PermissionService>) {
+    let mut rx = bus.subscribe();
+    tokio::spawn(async move {
+        while let Ok(event) = rx.recv().await {
+            if let AppEvent::PermissionAsked { request } = event {
+                service.reply(&request.id, PermissionReply::Once);
+            }
+        }
+    });
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn rule(permission: &str, pattern: &str, action: Action) -> Rule {
+        Rule {
+            permission: permission.to_string(),
+            pattern: pattern.to_string(),
+            action,
+        }
+    }
+
+    #[tokio::test]
+    async fn allow_rule_short_circuits_without_asking() {
+        let service = PermissionService::new(EventBus::new());
+        let rules: Ruleset = vec![rule("bash", "ls*", Action::Allow)];
+        let outcome = service
+            .ask(
+                &SessionId::new(),
+                &[&rules],
+                AskInput {
+                    permission: "bash".into(),
+                    pattern: "ls -la".into(),
+                    always_pattern: "ls *".into(),
+                    metadata: serde_json::Value::Null,
+                },
+                &CancellationToken::new(),
+            )
+            .await
+            .unwrap();
+        assert!(matches!(outcome, AskDecision::Allowed));
+    }
+
+    #[tokio::test]
+    async fn deny_rule_short_circuits_to_denied_error() {
+        let service = PermissionService::new(EventBus::new());
+        let rules: Ruleset = vec![rule("edit", "*.lock", Action::Deny)];
+        let err = service
+            .ask(
+                &SessionId::new(),
+                &[&rules],
+                AskInput {
+                    permission: "edit".into(),
+                    pattern: "Cargo.lock".into(),
+                    always_pattern: "Cargo.lock".into(),
+                    metadata: serde_json::Value::Null,
+                },
+                &CancellationToken::new(),
+            )
+            .await
+            .unwrap_err();
+        assert_eq!(err, AskError::Denied);
+    }
+
+    #[tokio::test]
+    async fn ask_publishes_event_and_waits_for_reply() {
+        let bus = EventBus::new();
+        let mut sub = bus.subscribe();
+        let service = std::sync::Arc::new(PermissionService::new(bus));
+        let svc = service.clone();
+
+        let handle = tokio::spawn(async move {
+            svc.ask(
+                &SessionId::new(),
+                &[],
+                AskInput {
+                    permission: "bash".into(),
+                    pattern: "rm -rf /".into(),
+                    always_pattern: "rm *".into(),
+                    metadata: serde_json::Value::Null,
+                },
+                &CancellationToken::new(),
+            )
+            .await
+        });
+
+        let event = sub.recv().await.unwrap();
+        let id = match event {
+            AppEvent::PermissionAsked { request } => request.id,
+            _ => panic!("expected PermissionAsked"),
+        };
+        assert!(service.reply(&id, PermissionReply::Once));
+
+        let outcome = handle.await.unwrap().unwrap();
+        assert!(matches!(outcome, AskDecision::Allowed));
+    }
+
+    #[tokio::test]
+    async fn always_reply_yields_rule_to_persist() {
+        let bus = EventBus::new();
+        let mut sub = bus.subscribe();
+        let service = std::sync::Arc::new(PermissionService::new(bus));
+        let svc = service.clone();
+
+        let handle = tokio::spawn(async move {
+            svc.ask(
+                &SessionId::new(),
+                &[],
+                AskInput {
+                    permission: "bash".into(),
+                    pattern: "git push origin main".into(),
+                    always_pattern: "git push*".into(),
+                    metadata: serde_json::Value::Null,
+                },
+                &CancellationToken::new(),
+            )
+            .await
+        });
+
+        let request = match sub.recv().await.unwrap() {
+            AppEvent::PermissionAsked { request } => request,
+            _ => panic!("expected PermissionAsked"),
+        };
+        service.reply(&request.id, PermissionReply::Always);
+
+        let outcome = handle.await.unwrap().unwrap();
+        match outcome {
+            AskDecision::AllowedAlways(rule) => {
+                assert_eq!(rule.permission, "bash");
+                assert_eq!(rule.pattern, "git push*");
+                assert_eq!(rule.action, Action::Allow);
+            }
+            _ => panic!("expected AllowedAlways"),
+        }
+    }
+
+    #[tokio::test]
+    async fn reject_reply_yields_rejected_error() {
+        let bus = EventBus::new();
+        let mut sub = bus.subscribe();
+        let service = std::sync::Arc::new(PermissionService::new(bus));
+        let svc = service.clone();
+
+        let handle = tokio::spawn(async move {
+            svc.ask(
+                &SessionId::new(),
+                &[],
+                AskInput {
+                    permission: "bash".into(),
+                    pattern: "rm -rf /".into(),
+                    always_pattern: "rm *".into(),
+                    metadata: serde_json::Value::Null,
+                },
+                &CancellationToken::new(),
+            )
+            .await
+        });
+
+        let request = match sub.recv().await.unwrap() {
+            AppEvent::PermissionAsked { request } => request,
+            _ => panic!("expected PermissionAsked"),
+        };
+        service.reply(&request.id, PermissionReply::Reject);
+
+        let err = handle.await.unwrap().unwrap_err();
+        assert_eq!(err, AskError::Rejected);
+    }
+
+    #[tokio::test]
+    async fn cancellation_aborts_a_pending_ask() {
+        let bus = EventBus::new();
+        let service = std::sync::Arc::new(PermissionService::new(bus));
+        let cancel = CancellationToken::new();
+        let cancel2 = cancel.clone();
+        let svc = service.clone();
+
+        let handle = tokio::spawn(async move {
+            svc.ask(
+                &SessionId::new(),
+                &[],
+                AskInput {
+                    permission: "bash".into(),
+                    pattern: "rm -rf /".into(),
+                    always_pattern: "rm *".into(),
+                    metadata: serde_json::Value::Null,
+                },
+                &cancel2,
+            )
+            .await
+        });
+
+        cancel.cancel();
+        let err = handle.await.unwrap().unwrap_err();
+        assert_eq!(err, AskError::Cancelled);
+    }
+
+    #[tokio::test]
+    async fn auto_approve_resolves_asks_as_once() {
+        let bus = EventBus::new();
+        let service = std::sync::Arc::new(PermissionService::new(bus.clone()));
+        spawn_auto_approve(bus, service.clone());
+
+        let outcome = service
+            .ask(
+                &SessionId::new(),
+                &[],
+                AskInput {
+                    permission: "bash".into(),
+                    pattern: "ls".into(),
+                    always_pattern: "ls *".into(),
+                    metadata: serde_json::Value::Null,
+                },
+                &CancellationToken::new(),
+            )
+            .await
+            .unwrap();
+        assert!(matches!(outcome, AskDecision::Allowed));
+    }
+}

crates/harness-core/src/store/actor.rs

@@ -13,6 +13,7 @@ pub enum StoreCmd {
     UpsertSession(Session, Reply<()>),
     UpsertMessage(Message, Reply<()>),
     UpsertPart(Part, Reply<()>),
+    Session(SessionId, Reply<Option<Session>>),
     Sessions(Reply<Vec<Session>>),
     Messages(SessionId, Reply<Vec<Message>>),
     Parts(MessageId, Reply<Vec<Part>>),
@@ -69,6 +70,15 @@ fn upsert_part(conn: &Connection, part: &Part) -> Result<(), StoreError> {
     Ok(())
 }
 
+fn get_session(conn: &Connection, id: &SessionId) -> Result<Option<Session>, StoreError> {
+    let mut stmt = conn.prepare("SELECT data FROM session WHERE id = ?1")?;
+    let mut rows = stmt.query_map(params![id.as_ref()], |row| row.get::<_, String>(0))?;
+    match rows.next() {
+        Some(data) => Ok(Some(serde_json::from_str(&data?)?)),
+        None => Ok(None),
+    }
+}
+
 fn list_sessions(conn: &Connection) -> Result<Vec<Session>, StoreError> {
     let mut stmt = conn.prepare("SELECT data FROM session ORDER BY id")?;
     let rows = stmt
@@ -116,6 +126,9 @@ pub fn run(conn: Connection, mut rx: tokio::sync::mpsc::Receiver<StoreCmd>) {
             StoreCmd::UpsertPart(part, reply) => {
                 let _ = reply.send(upsert_part(&conn, &part));
             }
+            StoreCmd::Session(id, reply) => {
+                let _ = reply.send(get_session(&conn, &id));
+            }
             StoreCmd::Sessions(reply) => {
                 let _ = reply.send(list_sessions(&conn));
             }

crates/harness-core/src/store/api.rs

@@ -77,6 +77,11 @@ impl Store {
         self.call(|reply| StoreCmd::UpsertPart(part, reply)).await
     }
 
+    pub async fn session(&self, session_id: SessionId) -> Result<Option<Session>, StoreError> {
+        self.call(|reply| StoreCmd::Session(session_id, reply))
+            .await
+    }
+
     pub async fn sessions(&self) -> Result<Vec<Session>, StoreError> {
         self.call(StoreCmd::Sessions).await
     }

crates/harness-core/src/tool/mod.rs

@@ -0,0 +1,296 @@
+pub mod truncate;
+
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::{Arc, Mutex};
+
+use async_trait::async_trait;
+use tokio_util::sync::CancellationToken;
+
+use crate::permission::{AskDecision, AskError, AskInput, PermissionService, Ruleset};
+use crate::types::{MessageId, SessionId};
+
+#[derive(Debug, thiserror::Error)]
+pub enum ToolError {
+    #[error("permission denied")]
+    Denied,
+    #[error("rejected by user")]
+    Rejected,
+    #[error("cancelled")]
+    Cancelled,
+    #[error("invalid input: {0}")]
+    Invalid(String),
+    #[error("{0}")]
+    Other(String),
+}
+
+impl From<AskError> for ToolError {
+    fn from(err: AskError) -> Self {
+        match err {
+            AskError::Denied => ToolError::Denied,
+            AskError::Rejected => ToolError::Rejected,
+            AskError::Cancelled => ToolError::Cancelled,
+        }
+    }
+}
+
+/// Live metadata updates for a running tool call — consumed by the processor to
+/// upsert the part's `ToolState::Running.metadata` and emit `AppEvent::PartUpdated`.
+#[derive(Clone)]
+pub struct MetadataSink(tokio::sync::mpsc::UnboundedSender<serde_json::Value>);
+
+impl MetadataSink {
+    pub fn channel() -> (
+        Self,
+        tokio::sync::mpsc::UnboundedReceiver<serde_json::Value>,
+    ) {
+        let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
+        (Self(tx), rx)
+    }
+
+    pub fn set(&self, value: serde_json::Value) {
+        let _ = self.0.send(value);
+    }
+}
+
+/// Bound to one tool call: session + a config/agent ruleset snapshot + the session's
+/// live `extra_rules` (mutated in place when the user replies "Always").
+pub struct PermissionHandle {
+    service: Arc<PermissionService>,
+    session_id: SessionId,
+    static_rules: Ruleset,
+    extra_rules: Arc<Mutex<Ruleset>>,
+    cancel: CancellationToken,
+}
+
+impl PermissionHandle {
+    pub fn new(
+        service: Arc<PermissionService>,
+        session_id: SessionId,
+        static_rules: Ruleset,
+        extra_rules: Arc<Mutex<Ruleset>>,
+        cancel: CancellationToken,
+    ) -> Self {
+        Self {
+            service,
+            session_id,
+            static_rules,
+            extra_rules,
+            cancel,
+        }
+    }
+
+    pub async fn ask(
+        &self,
+        permission: impl Into<String>,
+        pattern: impl Into<String>,
+        always_pattern: impl Into<String>,
+        metadata: serde_json::Value,
+    ) -> Result<(), ToolError> {
+        let extra_snapshot = self.extra_rules.lock().unwrap().clone();
+        let stack: [&Ruleset; 2] = [&self.static_rules, &extra_snapshot];
+        let decision = self
+            .service
+            .ask(
+                &self.session_id,
+                &stack,
+                AskInput {
+                    permission: permission.into(),
+                    pattern: pattern.into(),
+                    always_pattern: always_pattern.into(),
+                    metadata,
+                },
+                &self.cancel,
+            )
+            .await?;
+        if let AskDecision::AllowedAlways(rule) = decision {
+            self.extra_rules.lock().unwrap().push(rule);
+        }
+        Ok(())
+    }
+}
+
+pub struct ToolCtx {
+    pub session_id: SessionId,
+    pub message_id: MessageId,
+    pub call_id: String,
+    pub cwd: PathBuf,
+    /// Per-session directory tool output spills to when it exceeds `truncate::MAX_CHARS`.
+    pub data_dir: PathBuf,
+    pub cancel: CancellationToken,
+    pub ask: PermissionHandle,
+    pub metadata: MetadataSink,
+}
+
+#[derive(Debug)]
+pub struct ToolOutput {
+    pub title: String,
+    pub output: String,
+    pub metadata: serde_json::Value,
+}
+
+impl ToolOutput {
+    pub fn new(title: impl Into<String>, output: impl Into<String>) -> Self {
+        Self {
+            title: title.into(),
+            output: output.into(),
+            metadata: serde_json::Value::Null,
+        }
+    }
+}
+
+#[async_trait]
+pub trait Tool: Send + Sync {
+    fn name(&self) -> &str;
+    fn description(&self) -> &str;
+    fn parameters(&self) -> serde_json::Value;
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError>;
+}
+
+/// Also exposed for tests/tool-error metadata (opencode's `invalid` pattern: echo the schema back).
+pub fn invalid_input(tool: &dyn Tool, message: impl std::fmt::Display) -> ToolError {
+    ToolError::Invalid(format!("{message}\nExpected schema: {}", tool.parameters()))
+}
+
+#[derive(Default, Clone)]
+pub struct ToolRegistry {
+    tools: HashMap<String, Arc<dyn Tool>>,
+}
+
+impl ToolRegistry {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn register(&mut self, tool: Arc<dyn Tool>) {
+        self.tools.insert(tool.name().to_string(), tool);
+    }
+
+    pub fn get(&self, name: &str) -> Option<Arc<dyn Tool>> {
+        self.tools.get(name).cloned()
+    }
+
+    pub fn all(&self) -> Vec<Arc<dyn Tool>> {
+        self.tools.values().cloned().collect()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::event::EventBus;
+    use crate::permission::Rule;
+
+    struct Echo;
+
+    #[async_trait]
+    impl Tool for Echo {
+        fn name(&self) -> &str {
+            "echo"
+        }
+        fn description(&self) -> &str {
+            "echoes input"
+        }
+        fn parameters(&self) -> serde_json::Value {
+            serde_json::json!({"type": "object"})
+        }
+        async fn execute(
+            &self,
+            input: serde_json::Value,
+            _ctx: ToolCtx,
+        ) -> Result<ToolOutput, ToolError> {
+            Ok(ToolOutput::new("echo", input.to_string()))
+        }
+    }
+
+    #[test]
+    fn registry_registers_and_looks_up_by_name() {
+        let mut registry = ToolRegistry::new();
+        registry.register(Arc::new(Echo));
+        assert!(registry.get("echo").is_some());
+        assert!(registry.get("missing").is_none());
+        assert_eq!(registry.all().len(), 1);
+    }
+
+    #[tokio::test]
+    async fn permission_handle_allows_when_ruleset_grants() {
+        let service = Arc::new(PermissionService::new(EventBus::new()));
+        let handle = PermissionHandle::new(
+            service,
+            SessionId::new(),
+            vec![Rule {
+                permission: "bash".into(),
+                pattern: "ls*".into(),
+                action: crate::permission::Action::Allow,
+            }],
+            Arc::new(Mutex::new(Vec::new())),
+            CancellationToken::new(),
+        );
+        handle
+            .ask("bash", "ls -la", "ls *", serde_json::Value::Null)
+            .await
+            .unwrap();
+    }
+
+    #[tokio::test]
+    async fn permission_handle_denies_and_maps_to_tool_error() {
+        let service = Arc::new(PermissionService::new(EventBus::new()));
+        let handle = PermissionHandle::new(
+            service,
+            SessionId::new(),
+            vec![Rule {
+                permission: "bash".into(),
+                pattern: "rm*".into(),
+                action: crate::permission::Action::Deny,
+            }],
+            Arc::new(Mutex::new(Vec::new())),
+            CancellationToken::new(),
+        );
+        let err = handle
+            .ask("bash", "rm -rf /", "rm *", serde_json::Value::Null)
+            .await
+            .unwrap_err();
+        assert!(matches!(err, ToolError::Denied));
+    }
+
+    #[tokio::test]
+    async fn permission_handle_always_mutates_shared_extra_rules() {
+        let bus = EventBus::new();
+        let mut sub = bus.subscribe();
+        let service = Arc::new(PermissionService::new(bus));
+        let extra = Arc::new(Mutex::new(Vec::new()));
+        let handle = PermissionHandle::new(
+            service.clone(),
+            SessionId::new(),
+            Vec::new(),
+            extra.clone(),
+            CancellationToken::new(),
+        );
+
+        let ask = tokio::spawn(async move {
+            handle
+                .ask(
+                    "bash",
+                    "git push origin",
+                    "git push*",
+                    serde_json::Value::Null,
+                )
+                .await
+        });
+
+        let request = match sub.recv().await.unwrap() {
+            crate::event::AppEvent::PermissionAsked { request } => request,
+            _ => panic!("expected PermissionAsked"),
+        };
+        service.reply(&request.id, crate::permission::PermissionReply::Always);
+        ask.await.unwrap().unwrap();
+
+        let rules = extra.lock().unwrap();
+        assert_eq!(rules.len(), 1);
+        assert_eq!(rules[0].pattern, "git push*");
+    }
+}

crates/harness-core/src/tool/truncate.rs

@@ -0,0 +1,75 @@
+use std::path::Path;
+
+/// docs/05-tools.md: cap tool output at 30,000 chars, keep head + tail, spill the full
+/// text to `<data_dir>/tool-output/<call_id>.txt` so a subagent can Grep/Read it later.
+pub const MAX_CHARS: usize = 30_000;
+const HEAD_CHARS: usize = MAX_CHARS / 2;
+const TAIL_CHARS: usize = MAX_CHARS - HEAD_CHARS;
+
+pub struct Truncated {
+    pub text: String,
+    pub truncated: bool,
+}
+
+/// `data_dir` is the session's tool-output directory; `call_id` names the spill file.
+pub fn truncate(text: &str, data_dir: &Path, call_id: &str) -> std::io::Result<Truncated> {
+    if text.chars().count() <= MAX_CHARS {
+        return Ok(Truncated {
+            text: text.to_string(),
+            truncated: false,
+        });
+    }
+
+    let chars: Vec<char> = text.chars().collect();
+    let head: String = chars[..HEAD_CHARS].iter().collect();
+    let tail: String = chars[chars.len() - TAIL_CHARS..].iter().collect();
+    let hidden = chars.len() - HEAD_CHARS - TAIL_CHARS;
+
+    std::fs::create_dir_all(data_dir)?;
+    let spill_path = data_dir.join(format!("{call_id}.txt"));
+    std::fs::write(&spill_path, text)?;
+
+    let message = format!(
+        "{head}\n... [{hidden} chars truncated; full output: {} ] ...\n{tail}",
+        spill_path.display()
+    );
+    Ok(Truncated {
+        text: message,
+        truncated: true,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn short_output_is_unchanged() {
+        let dir = tempfile::tempdir().unwrap();
+        let result = truncate("hello world", dir.path(), "call_1").unwrap();
+        assert!(!result.truncated);
+        assert_eq!(result.text, "hello world");
+        assert!(!dir.path().join("call_1.txt").exists());
+    }
+
+    #[test]
+    fn long_output_is_truncated_and_spilled() {
+        let dir = tempfile::tempdir().unwrap();
+        let long = "a".repeat(MAX_CHARS + 1000);
+        let result = truncate(&long, dir.path(), "call_2").unwrap();
+        assert!(result.truncated);
+        assert!(result.text.contains("chars truncated"));
+        assert!(result.text.len() < long.len());
+
+        let spilled = std::fs::read_to_string(dir.path().join("call_2.txt")).unwrap();
+        assert_eq!(spilled, long);
+    }
+
+    #[test]
+    fn boundary_exactly_at_max_is_not_truncated() {
+        let dir = tempfile::tempdir().unwrap();
+        let text = "a".repeat(MAX_CHARS);
+        let result = truncate(&text, dir.path(), "call_3").unwrap();
+        assert!(!result.truncated);
+    }
+}

crates/harness-core/src/types/mod.rs

@@ -6,6 +6,6 @@ pub mod session;
 
 pub use ids::{MessageId, PartId, SessionId};
 pub use message::{Message, MessageError, Role};
-pub use model::{ModelInfo, ModelRef, TokenUsage};
+pub use model::{ModelCost, ModelInfo, ModelRef, TokenUsage};
 pub use part::{Part, PartBody, ToolState};
 pub use session::Session;

crates/harness-core/src/types/model.rs

@@ -34,11 +34,71 @@ impl TokenUsage {
     }
 }
 
-// Full cost/context-limit metadata is populated by harness-providers (models.dev) in M1/M3;
+/// Per-model pricing in USD per **one million** tokens. Populated from models.dev metadata
-// this placeholder only carries what harness-core needs to key on.
+/// by `harness-providers`.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Serialize, Deserialize)]
+pub struct ModelCost {
+    pub input: f64,
+    pub output: f64,
+    pub cache_read: f64,
+    pub cache_write: f64,
+}
+
+impl ModelCost {
+    /// Dollar cost of a usage sample. Reasoning tokens are billed within `output` by the
+    /// providers we support (OpenAI reports them as a subset of `output_tokens`; Anthropic
+    /// counts thinking in output), so they are intentionally not charged separately here.
+    pub fn cost_of(&self, usage: &TokenUsage) -> f64 {
+        let per_million = |tokens: u64, rate: f64| (tokens as f64) * rate / 1_000_000.0;
+        per_million(usage.input, self.input)
+            + per_million(usage.output, self.output)
+            + per_million(usage.cache_read, self.cache_read)
+            + per_million(usage.cache_write, self.cache_write)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn cost_of_sums_components_per_million_excluding_reasoning() {
+        let cost = ModelCost {
+            input: 3.0,
+            output: 15.0,
+            cache_read: 0.30,
+            cache_write: 3.75,
+        };
+        let usage = TokenUsage {
+            input: 1_000_000,
+            output: 1_000_000,
+            reasoning: 500_000, // billed within output — must not add extra cost
+            cache_read: 1_000_000,
+            cache_write: 1_000_000,
+        };
+        // 3 + 15 + 0.30 + 3.75, with reasoning contributing nothing.
+        assert!((cost.cost_of(&usage) - 22.05).abs() < 1e-9);
+    }
+
+    #[test]
+    fn default_cost_is_zero() {
+        assert_eq!(ModelCost::default().cost_of(&TokenUsage::default()), 0.0);
+    }
+}
+
+/// Model metadata, keyed on `model`. Cost/limits are populated from models.dev by
+/// `harness-providers`; `reasoning`/`tool_call`/`attachment` are capability flags.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 pub struct ModelInfo {
     pub model: ModelRef,
     pub context_limit: u64,
     pub output_limit: u64,
+    #[serde(default)]
+    pub cost: ModelCost,
+    #[serde(default)]
+    pub reasoning: bool,
+    #[serde(default)]
+    pub tool_call: bool,
+    #[serde(default)]
+    pub attachment: bool,
 }

crates/harness-providers/Cargo.toml

@@ -6,6 +6,23 @@ license.workspace = true
 
 [dependencies]
 harness-core = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+futures = { workspace = true }
+async-stream = { workspace = true }
+async-trait = { workspace = true }
+reqwest = { workspace = true }
+eventsource-stream = { workspace = true }
+bytes = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+thiserror = { workspace = true }
+tracing = { workspace = true }
+dirs = { workspace = true }
+
+[dev-dependencies]
+tokio = { workspace = true, features = ["test-util", "macros"] }
+tempfile = { workspace = true }
 
 [lints]
 workspace = true

crates/harness-providers/assets/models-snapshot.json

@@ -0,0 +1,67 @@
+{
+  "anthropic": {
+    "id": "anthropic",
+    "name": "Anthropic",
+    "models": {
+      "claude-sonnet-4-5": {
+        "id": "claude-sonnet-4-5",
+        "reasoning": true,
+        "tool_call": true,
+        "attachment": true,
+        "cost": { "input": 3, "output": 15, "cache_read": 0.3, "cache_write": 3.75 },
+        "limit": { "context": 200000, "output": 64000 }
+      },
+      "claude-opus-4-1": {
+        "id": "claude-opus-4-1",
+        "reasoning": true,
+        "tool_call": true,
+        "attachment": true,
+        "cost": { "input": 15, "output": 75, "cache_read": 1.5, "cache_write": 18.75 },
+        "limit": { "context": 200000, "output": 32000 }
+      },
+      "claude-haiku-4-5": {
+        "id": "claude-haiku-4-5",
+        "reasoning": true,
+        "tool_call": true,
+        "attachment": true,
+        "cost": { "input": 1, "output": 5, "cache_read": 0.1, "cache_write": 1.25 },
+        "limit": { "context": 200000, "output": 64000 }
+      }
+    }
+  },
+  "openai": {
+    "id": "openai",
+    "name": "OpenAI",
+    "models": {
+      "gpt-4o": {
+        "id": "gpt-4o",
+        "tool_call": true,
+        "attachment": true,
+        "cost": { "input": 2.5, "output": 10, "cache_read": 1.25 },
+        "limit": { "context": 128000, "output": 16384 }
+      },
+      "gpt-4o-mini": {
+        "id": "gpt-4o-mini",
+        "tool_call": true,
+        "attachment": true,
+        "cost": { "input": 0.15, "output": 0.6, "cache_read": 0.075 },
+        "limit": { "context": 128000, "output": 16384 }
+      },
+      "gpt-5": {
+        "id": "gpt-5",
+        "reasoning": true,
+        "tool_call": true,
+        "attachment": true,
+        "cost": { "input": 1.25, "output": 10, "cache_read": 0.125 },
+        "limit": { "context": 400000, "output": 128000 }
+      },
+      "o3": {
+        "id": "o3",
+        "reasoning": true,
+        "tool_call": true,
+        "cost": { "input": 2, "output": 8, "cache_read": 0.5 },
+        "limit": { "context": 200000, "output": 100000 }
+      }
+    }
+  }
+}

crates/harness-providers/src/anthropic.rs

@@ -0,0 +1,157 @@
+use std::time::Duration;
+
+use async_trait::async_trait;
+use harness_core::llm::{LlmEventStream, LlmRequest, Provider, ProviderError};
+use harness_core::types::ModelInfo;
+use tokio_util::sync::CancellationToken;
+
+use crate::codec::anthropic::{build_request, decode, initiator_header};
+
+const DEFAULT_BASE_URL: &str = "https://api.anthropic.com";
+const ANTHROPIC_VERSION: &str = "2023-06-01";
+const ANTHROPIC_BETA: &str =
+    "interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14";
+
+pub struct AnthropicProvider {
+    api_key: String,
+    base_url: String,
+    client: reqwest::Client,
+}
+
+impl AnthropicProvider {
+    pub fn new(api_key: impl Into<String>) -> Self {
+        Self {
+            api_key: api_key.into(),
+            base_url: DEFAULT_BASE_URL.to_string(),
+            client: reqwest::Client::new(),
+        }
+    }
+
+    #[cfg(test)]
+    fn with_base_url(api_key: impl Into<String>, base_url: impl Into<String>) -> Self {
+        Self {
+            api_key: api_key.into(),
+            base_url: base_url.into(),
+            client: reqwest::Client::new(),
+        }
+    }
+
+    fn classify_error(
+        status: reqwest::StatusCode,
+        body: String,
+        retry_after: Option<Duration>,
+    ) -> ProviderError {
+        match status.as_u16() {
+            401 | 403 => ProviderError::Auth(body),
+            429 => ProviderError::RateLimited { retry_after },
+            529 => ProviderError::Overloaded,
+            s if (500..600).contains(&s) => ProviderError::Overloaded,
+            s => ProviderError::Http { status: s, body },
+        }
+    }
+}
+
+#[async_trait]
+impl Provider for AnthropicProvider {
+    fn id(&self) -> &str {
+        "anthropic"
+    }
+
+    async fn list_models(&self) -> Result<Vec<ModelInfo>, ProviderError> {
+        // Full models.dev metadata integration lands in M3; callers fall back to
+        // config-specified model strings until then.
+        Ok(Vec::new())
+    }
+
+    async fn stream(
+        &self,
+        req: LlmRequest,
+        cancel: CancellationToken,
+    ) -> Result<LlmEventStream, ProviderError> {
+        let body = build_request(&req);
+        let url = format!("{}/v1/messages", self.base_url);
+        let initiator = initiator_header(req.initiator);
+
+        let send = self
+            .client
+            .post(&url)
+            .header("x-api-key", &self.api_key)
+            .header("anthropic-version", ANTHROPIC_VERSION)
+            .header("anthropic-beta", ANTHROPIC_BETA)
+            .header("x-initiator", initiator)
+            .json(&body)
+            .send();
+
+        let response = tokio::select! {
+            result = send => result.map_err(|e| ProviderError::Network(e.to_string()))?,
+            _ = cancel.cancelled() => return Err(ProviderError::Cancelled),
+        };
+
+        if !response.status().is_success() {
+            let status = response.status();
+            let retry_after = response
+                .headers()
+                .get("retry-after")
+                .and_then(|v| v.to_str().ok())
+                .and_then(|s| s.parse::<u64>().ok())
+                .map(Duration::from_secs);
+            let body = response.text().await.unwrap_or_default();
+            return Err(Self::classify_error(status, body, retry_after));
+        }
+
+        Ok(decode(response.bytes_stream()))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn classifies_status_codes() {
+        assert!(matches!(
+            AnthropicProvider::classify_error(
+                reqwest::StatusCode::TOO_MANY_REQUESTS,
+                String::new(),
+                None
+            ),
+            ProviderError::RateLimited { .. }
+        ));
+        assert!(matches!(
+            AnthropicProvider::classify_error(
+                reqwest::StatusCode::UNAUTHORIZED,
+                String::new(),
+                None
+            ),
+            ProviderError::Auth(_)
+        ));
+        assert!(matches!(
+            AnthropicProvider::classify_error(
+                reqwest::StatusCode::SERVICE_UNAVAILABLE,
+                String::new(),
+                None
+            ),
+            ProviderError::Overloaded
+        ));
+        assert!(matches!(
+            AnthropicProvider::classify_error(
+                reqwest::StatusCode::BAD_REQUEST,
+                String::new(),
+                None
+            ),
+            ProviderError::Http { status: 400, .. }
+        ));
+    }
+
+    #[test]
+    fn id_is_anthropic() {
+        let provider = AnthropicProvider::new("test-key");
+        assert_eq!(provider.id(), "anthropic");
+    }
+
+    #[test]
+    fn constructs_with_custom_base_url() {
+        let provider = AnthropicProvider::with_base_url("test-key", "http://localhost:1234");
+        assert_eq!(provider.base_url, "http://localhost:1234");
+    }
+}

crates/harness-providers/src/auth.rs

@@ -0,0 +1,228 @@
+//! Credential storage for providers that authenticate outside the config file.
+//!
+//! A single JSON file (`~/.local/share/ai-harness/auth.json`, mode `0600`) keyed by provider
+//! id. OAuth providers (Copilot) store the token triple and refresh it in place; API-key
+//! providers store the bare key. Config-supplied keys take precedence over this file — this
+//! is only for interactively-obtained credentials.
+
+use std::collections::HashMap;
+use std::io;
+use std::path::{Path, PathBuf};
+
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum AuthRecord {
+    /// OAuth credentials. `expires` is a unix-ms timestamp; `0` means the token never expires
+    /// (opencode's direct-Bearer Copilot mode).
+    OAuth {
+        access: String,
+        refresh: String,
+        expires: i64,
+    },
+    Api {
+        key: String,
+    },
+}
+
+impl AuthRecord {
+    /// True when an OAuth token is at or past `expires` (with a safety skew), given `now_ms`.
+    /// API keys and never-expiring OAuth tokens (`expires == 0`) are never considered expired.
+    pub fn is_expired(&self, now_ms: i64, skew_ms: i64) -> bool {
+        match self {
+            AuthRecord::OAuth { expires, .. } if *expires > 0 => now_ms + skew_ms >= *expires,
+            _ => false,
+        }
+    }
+}
+
+/// Read/modify/write access to the auth file. Cheap to construct; every operation reloads so
+/// concurrent writers (a refresh in one provider, a login in another) don't clobber each other.
+#[derive(Debug, Clone)]
+pub struct AuthStorage {
+    path: PathBuf,
+}
+
+impl AuthStorage {
+    pub fn new(path: PathBuf) -> Self {
+        Self { path }
+    }
+
+    /// `~/.local/share/ai-harness/auth.json` (falling back to the temp dir if there is no
+    /// data directory).
+    pub fn default_path() -> PathBuf {
+        dirs::data_dir()
+            .unwrap_or_else(std::env::temp_dir)
+            .join("ai-harness")
+            .join("auth.json")
+    }
+
+    pub fn with_default_path() -> Self {
+        Self::new(Self::default_path())
+    }
+
+    /// Loads all records. A missing file yields an empty map; a corrupt file is an error.
+    pub fn load(&self) -> io::Result<HashMap<String, AuthRecord>> {
+        match std::fs::read(&self.path) {
+            Ok(bytes) => serde_json::from_slice(&bytes)
+                .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e)),
+            Err(e) if e.kind() == io::ErrorKind::NotFound => Ok(HashMap::new()),
+            Err(e) => Err(e),
+        }
+    }
+
+    pub fn get(&self, provider: &str) -> io::Result<Option<AuthRecord>> {
+        Ok(self.load()?.remove(provider))
+    }
+
+    pub fn set(&self, provider: &str, record: AuthRecord) -> io::Result<()> {
+        let mut records = self.load()?;
+        records.insert(provider.to_string(), record);
+        self.write(&records)
+    }
+
+    pub fn remove(&self, provider: &str) -> io::Result<()> {
+        let mut records = self.load()?;
+        if records.remove(provider).is_some() {
+            self.write(&records)?;
+        }
+        Ok(())
+    }
+
+    /// Serializes `records` and writes them with `0600` permissions via a temp-file rename so
+    /// a partial write can never leave a truncated auth file behind.
+    fn write(&self, records: &HashMap<String, AuthRecord>) -> io::Result<()> {
+        if let Some(parent) = self.path.parent() {
+            std::fs::create_dir_all(parent)?;
+        }
+        let json = serde_json::to_vec_pretty(records)
+            .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
+
+        let tmp = self.path.with_extension("json.tmp");
+        std::fs::write(&tmp, &json)?;
+        set_owner_only(&tmp)?;
+        std::fs::rename(&tmp, &self.path)?;
+        Ok(())
+    }
+}
+
+#[cfg(unix)]
+fn set_owner_only(path: &Path) -> io::Result<()> {
+    use std::os::unix::fs::PermissionsExt;
+    std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o600))
+}
+
+#[cfg(not(unix))]
+fn set_owner_only(_path: &Path) -> io::Result<()> {
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn storage() -> (tempfile::TempDir, AuthStorage) {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("nested").join("auth.json");
+        (dir, AuthStorage::new(path))
+    }
+
+    #[test]
+    fn missing_file_loads_empty() {
+        let (_dir, store) = storage();
+        assert!(store.load().unwrap().is_empty());
+        assert_eq!(store.get("anthropic").unwrap(), None);
+    }
+
+    #[test]
+    fn set_get_roundtrip_and_overwrite() {
+        let (_dir, store) = storage();
+        store
+            .set("anthropic", AuthRecord::Api { key: "k1".into() })
+            .unwrap();
+        assert_eq!(
+            store.get("anthropic").unwrap(),
+            Some(AuthRecord::Api { key: "k1".into() })
+        );
+
+        store
+            .set("anthropic", AuthRecord::Api { key: "k2".into() })
+            .unwrap();
+        assert_eq!(
+            store.get("anthropic").unwrap(),
+            Some(AuthRecord::Api { key: "k2".into() })
+        );
+    }
+
+    #[test]
+    fn multiple_providers_coexist() {
+        let (_dir, store) = storage();
+        store
+            .set("openai", AuthRecord::Api { key: "sk".into() })
+            .unwrap();
+        store
+            .set(
+                "github-copilot",
+                AuthRecord::OAuth {
+                    access: "a".into(),
+                    refresh: "r".into(),
+                    expires: 0,
+                },
+            )
+            .unwrap();
+        let all = store.load().unwrap();
+        assert_eq!(all.len(), 2);
+        assert!(all.contains_key("openai"));
+        assert!(all.contains_key("github-copilot"));
+    }
+
+    #[test]
+    fn remove_deletes_only_named_provider() {
+        let (_dir, store) = storage();
+        store
+            .set("openai", AuthRecord::Api { key: "sk".into() })
+            .unwrap();
+        store
+            .set("anthropic", AuthRecord::Api { key: "an".into() })
+            .unwrap();
+        store.remove("openai").unwrap();
+        assert_eq!(store.get("openai").unwrap(), None);
+        assert!(store.get("anthropic").unwrap().is_some());
+        // Removing an absent provider is a no-op, not an error.
+        store.remove("openai").unwrap();
+    }
+
+    #[test]
+    fn oauth_expiry_respects_skew_and_never_expires() {
+        let never = AuthRecord::OAuth {
+            access: "a".into(),
+            refresh: "r".into(),
+            expires: 0,
+        };
+        assert!(!never.is_expired(i64::MAX, 0));
+
+        let expiring = AuthRecord::OAuth {
+            access: "a".into(),
+            refresh: "r".into(),
+            expires: 1_000,
+        };
+        assert!(!expiring.is_expired(800, 120));
+        assert!(expiring.is_expired(880, 120)); // within skew window
+        assert!(expiring.is_expired(1_000, 0));
+
+        assert!(!AuthRecord::Api { key: "k".into() }.is_expired(i64::MAX, 0));
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn file_is_written_owner_only() {
+        use std::os::unix::fs::PermissionsExt;
+        let (_dir, store) = storage();
+        store
+            .set("openai", AuthRecord::Api { key: "sk".into() })
+            .unwrap();
+        let mode = std::fs::metadata(&store.path).unwrap().permissions().mode();
+        assert_eq!(mode & 0o777, 0o600);
+    }
+}

crates/harness-providers/src/codec/anthropic.rs

@@ -0,0 +1,499 @@
+//! Request builder + SSE decoder for Anthropic's `/v1/messages` streaming API.
+
+use std::collections::HashMap;
+
+use async_stream::try_stream;
+use eventsource_stream::Eventsource;
+use futures::Stream;
+use harness_core::llm::{
+    FinishReason, Initiator, LlmEvent, LlmEventStream, LlmRequest, ProviderError, Role, WireContent,
+};
+use harness_core::types::TokenUsage;
+use serde_json::{json, Value};
+
+const DEFAULT_MAX_TOKENS: u32 = 8192;
+/// opencode's `applyCaching`: cache breakpoints on the first 2 system blocks + last 2
+/// non-system messages.
+const CACHE_BREAKPOINTS: usize = 2;
+
+fn cache_control() -> Value {
+    json!({"type": "ephemeral"})
+}
+
+fn build_system(system: &[String]) -> Vec<Value> {
+    system
+        .iter()
+        .enumerate()
+        .map(|(i, block)| {
+            let mut b = json!({"type": "text", "text": block});
+            if i < CACHE_BREAKPOINTS {
+                b["cache_control"] = cache_control();
+            }
+            b
+        })
+        .collect()
+}
+
+fn wire_role_to_anthropic(role: Role) -> &'static str {
+    match role {
+        Role::User | Role::Tool => "user",
+        Role::Assistant => "assistant",
+        Role::System => "user", // system blocks are carried separately in `system`, not here
+    }
+}
+
+fn build_messages(messages: &[harness_core::llm::WireMessage]) -> Vec<Value> {
+    let mut built: Vec<Value> = messages
+        .iter()
+        .map(|m| {
+            let content: Vec<Value> = m
+                .content
+                .iter()
+                .map(|c| match c {
+                    WireContent::Text { text } => json!({"type": "text", "text": text}),
+                    WireContent::ToolCall { call_id, name, input } => {
+                        json!({"type": "tool_use", "id": call_id, "name": name, "input": input})
+                    }
+                    WireContent::ToolResult { call_id, output, is_error } => {
+                        json!({"type": "tool_result", "tool_use_id": call_id, "content": output, "is_error": is_error})
+                    }
+                    WireContent::Image { mime_type, data } => {
+                        json!({"type": "image", "source": {"type": "base64", "media_type": mime_type, "data": data}})
+                    }
+                })
+                .collect();
+            json!({"role": wire_role_to_anthropic(m.role), "content": content})
+        })
+        .collect();
+
+    let n = built.len();
+    for msg in built.iter_mut().skip(n.saturating_sub(CACHE_BREAKPOINTS)) {
+        if let Some(content) = msg["content"].as_array_mut() {
+            if let Some(last) = content.last_mut() {
+                last["cache_control"] = cache_control();
+            }
+        }
+    }
+    built
+}
+
+pub fn build_request(req: &LlmRequest) -> Value {
+    let mut body = json!({
+        "model": req.model,
+        "system": build_system(&req.system),
+        "messages": build_messages(&req.messages),
+        "max_tokens": req.max_tokens.unwrap_or(DEFAULT_MAX_TOKENS),
+        "stream": true,
+    });
+
+    if !req.tools.is_empty() {
+        body["tools"] = Value::Array(
+            req.tools
+                .iter()
+                .map(|t| json!({"name": t.name, "description": t.description, "input_schema": t.parameters}))
+                .collect(),
+        );
+    }
+    if let Some(temp) = req.temperature {
+        body["temperature"] = json!(temp);
+    }
+    if let Some(reasoning) = &req.reasoning {
+        if let Some(budget) = reasoning.budget_tokens {
+            body["thinking"] = json!({"type": "enabled", "budget_tokens": budget});
+        }
+    }
+    body
+}
+
+pub fn initiator_header(initiator: Initiator) -> &'static str {
+    match initiator {
+        Initiator::User => "user",
+        Initiator::Agent => "agent",
+    }
+}
+
+fn map_stop_reason(reason: Option<&str>) -> FinishReason {
+    match reason {
+        Some("end_turn") | Some("stop_sequence") => FinishReason::Stop,
+        Some("tool_use") => FinishReason::ToolCalls,
+        Some("max_tokens") => FinishReason::Length,
+        Some(other) => FinishReason::Unknown(other.to_string()),
+        None => FinishReason::Unknown("none".to_string()),
+    }
+}
+
+#[derive(Default)]
+struct BlockState {
+    kind: BlockKind,
+    call_id: String,
+    name: String,
+    partial_json: String,
+    signature: Option<String>,
+}
+
+#[derive(Default, PartialEq, Eq)]
+enum BlockKind {
+    #[default]
+    Text,
+    Thinking,
+    ToolUse,
+}
+
+/// Decodes a raw SSE byte stream into our normalized `LlmEvent` stream. Errors from the
+/// underlying HTTP stream and any Anthropic `error` event both surface as `Err`.
+pub fn decode<S, E>(byte_stream: S) -> LlmEventStream
+where
+    S: Stream<Item = Result<bytes::Bytes, E>> + Send + 'static,
+    E: std::error::Error + Send + Sync + 'static,
+{
+    let events = byte_stream.eventsource();
+    Box::pin(try_stream! {
+        futures::pin_mut!(events);
+        let mut blocks: HashMap<u64, BlockState> = HashMap::new();
+        let mut usage = TokenUsage::default();
+
+        while let Some(item) = futures::StreamExt::next(&mut events).await {
+            let event = item.map_err(|e| ProviderError::Decode(e.to_string()))?;
+            if event.data.is_empty() {
+                continue;
+            }
+            let value: Value = serde_json::from_str(&event.data)
+                .map_err(|e| ProviderError::Decode(format!("{e}: {}", event.data)))?;
+            let kind = value["type"].as_str().unwrap_or_default();
+
+            match kind {
+                "message_start" => {
+                    let u = &value["message"]["usage"];
+                    usage.input = u["input_tokens"].as_u64().unwrap_or(0);
+                    usage.cache_write = u["cache_creation_input_tokens"].as_u64().unwrap_or(0);
+                    usage.cache_read = u["cache_read_input_tokens"].as_u64().unwrap_or(0);
+                }
+                "content_block_start" => {
+                    let index = value["index"].as_u64().unwrap_or(0);
+                    let block = &value["content_block"];
+                    match block["type"].as_str().unwrap_or_default() {
+                        "text" => {
+                            blocks.insert(index, BlockState { kind: BlockKind::Text, ..Default::default() });
+                            yield LlmEvent::TextStart { id: index.to_string() };
+                        }
+                        "thinking" => {
+                            blocks.insert(index, BlockState { kind: BlockKind::Thinking, ..Default::default() });
+                            yield LlmEvent::ReasoningStart { id: index.to_string() };
+                        }
+                        "tool_use" => {
+                            let call_id = block["id"].as_str().unwrap_or_default().to_string();
+                            let name = block["name"].as_str().unwrap_or_default().to_string();
+                            blocks.insert(index, BlockState {
+                                kind: BlockKind::ToolUse,
+                                call_id: call_id.clone(),
+                                name: name.clone(),
+                                ..Default::default()
+                            });
+                            yield LlmEvent::ToolInputStart { call_id, name };
+                        }
+                        _ => {}
+                    }
+                }
+                "content_block_delta" => {
+                    let index = value["index"].as_u64().unwrap_or(0);
+                    let delta = &value["delta"];
+                    match delta["type"].as_str().unwrap_or_default() {
+                        "text_delta" => {
+                            let text = delta["text"].as_str().unwrap_or_default().to_string();
+                            yield LlmEvent::TextDelta { id: index.to_string(), text };
+                        }
+                        "thinking_delta" => {
+                            let text = delta["thinking"].as_str().unwrap_or_default().to_string();
+                            yield LlmEvent::ReasoningDelta { id: index.to_string(), text };
+                        }
+                        "signature_delta" => {
+                            if let Some(state) = blocks.get_mut(&index) {
+                                state.signature = Some(delta["signature"].as_str().unwrap_or_default().to_string());
+                            }
+                        }
+                        "input_json_delta" => {
+                            let partial = delta["partial_json"].as_str().unwrap_or_default();
+                            if let Some(state) = blocks.get_mut(&index) {
+                                state.partial_json.push_str(partial);
+                                yield LlmEvent::ToolInputDelta { call_id: state.call_id.clone(), json: partial.to_string() };
+                            }
+                        }
+                        _ => {}
+                    }
+                }
+                "content_block_stop" => {
+                    let index = value["index"].as_u64().unwrap_or(0);
+                    if let Some(state) = blocks.remove(&index) {
+                        match state.kind {
+                            BlockKind::Text => yield LlmEvent::TextEnd { id: index.to_string() },
+                            BlockKind::Thinking => {
+                                yield LlmEvent::ReasoningEnd { id: index.to_string(), signature: state.signature };
+                            }
+                            BlockKind::ToolUse => {
+                                let input: Value = if state.partial_json.trim().is_empty() {
+                                    json!({})
+                                } else {
+                                    serde_json::from_str(&state.partial_json).unwrap_or(Value::Null)
+                                };
+                                yield LlmEvent::ToolCall { call_id: state.call_id, name: state.name, input };
+                            }
+                        }
+                    }
+                }
+                "message_delta" => {
+                    if let Some(out) = value["usage"]["output_tokens"].as_u64() {
+                        usage.output = out;
+                    }
+                    let stop_reason = value["delta"]["stop_reason"].as_str();
+                    yield LlmEvent::Finish { reason: map_stop_reason(stop_reason), usage };
+                }
+                "error" => {
+                    let message = value["error"]["message"].as_str().unwrap_or("unknown error").to_string();
+                    let err_type = value["error"]["type"].as_str().unwrap_or("");
+                    Err(match err_type {
+                        "overloaded_error" => ProviderError::Overloaded,
+                        "rate_limit_error" => ProviderError::RateLimited { retry_after: None },
+                        "authentication_error" | "permission_error" => ProviderError::Auth(message),
+                        _ => ProviderError::Http { status: 0, body: message },
+                    })?;
+                }
+                _ => {} // ping, message_stop: nothing to emit
+            }
+        }
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use futures::StreamExt;
+    use harness_core::llm::{ReasoningOpts, ToolSchema, WireMessage};
+
+    fn sse_stream(raw: &'static str) -> LlmEventStream {
+        let chunks: Vec<Result<bytes::Bytes, std::io::Error>> =
+            vec![Ok(bytes::Bytes::from_static(raw.as_bytes()))];
+        decode(futures::stream::iter(chunks))
+    }
+
+    #[tokio::test]
+    async fn decodes_text_only_response() {
+        let raw = concat!(
+            "event: message_start\n",
+            "data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"input_tokens\":10}}}\n\n",
+            "event: content_block_start\n",
+            "data: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"text\",\"text\":\"\"}}\n\n",
+            "event: content_block_delta\n",
+            "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"text_delta\",\"text\":\"Hello\"}}\n\n",
+            "event: content_block_stop\n",
+            "data: {\"type\":\"content_block_stop\",\"index\":0}\n\n",
+            "event: message_delta\n",
+            "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\"},\"usage\":{\"output_tokens\":5}}\n\n",
+            "event: message_stop\n",
+            "data: {\"type\":\"message_stop\"}\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::TextStart { id: "0".into() },
+                LlmEvent::TextDelta {
+                    id: "0".into(),
+                    text: "Hello".into()
+                },
+                LlmEvent::TextEnd { id: "0".into() },
+                LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: TokenUsage {
+                        input: 10,
+                        output: 5,
+                        ..Default::default()
+                    },
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn decodes_tool_call_with_streamed_json_input() {
+        let raw = concat!(
+            "data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"input_tokens\":1}}}\n\n",
+            "data: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"tool_use\",\"id\":\"call_1\",\"name\":\"read\"}}\n\n",
+            "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\"{\\\"file\\\"\"}}\n\n",
+            "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"input_json_delta\",\"partial_json\":\":\\\"a.txt\\\"}\"}}\n\n",
+            "data: {\"type\":\"content_block_stop\",\"index\":0}\n\n",
+            "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"tool_use\"},\"usage\":{\"output_tokens\":8}}\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::ToolInputStart {
+                    call_id: "call_1".into(),
+                    name: "read".into()
+                },
+                LlmEvent::ToolInputDelta {
+                    call_id: "call_1".into(),
+                    json: "{\"file\"".into()
+                },
+                LlmEvent::ToolInputDelta {
+                    call_id: "call_1".into(),
+                    json: ":\"a.txt\"}".into()
+                },
+                LlmEvent::ToolCall {
+                    call_id: "call_1".into(),
+                    name: "read".into(),
+                    input: json!({"file": "a.txt"}),
+                },
+                LlmEvent::Finish {
+                    reason: FinishReason::ToolCalls,
+                    usage: TokenUsage {
+                        input: 1,
+                        output: 8,
+                        ..Default::default()
+                    },
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn decodes_thinking_block_with_signature() {
+        let raw = concat!(
+            "data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"input_tokens\":1}}}\n\n",
+            "data: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"thinking\"}}\n\n",
+            "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"thinking_delta\",\"thinking\":\"pondering\"}}\n\n",
+            "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"signature_delta\",\"signature\":\"sig123\"}}\n\n",
+            "data: {\"type\":\"content_block_stop\",\"index\":0}\n\n",
+            "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\"},\"usage\":{\"output_tokens\":2}}\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::ReasoningStart { id: "0".into() },
+                LlmEvent::ReasoningDelta {
+                    id: "0".into(),
+                    text: "pondering".into()
+                },
+                LlmEvent::ReasoningEnd {
+                    id: "0".into(),
+                    signature: Some("sig123".into())
+                },
+                LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: TokenUsage {
+                        input: 1,
+                        output: 2,
+                        ..Default::default()
+                    },
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn mid_stream_error_event_surfaces_as_err() {
+        let raw = concat!(
+            "data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"input_tokens\":1}}}\n\n",
+            "data: {\"type\":\"error\",\"error\":{\"type\":\"overloaded_error\",\"message\":\"overloaded\"}}\n\n",
+        );
+        let events: Vec<Result<LlmEvent, ProviderError>> = sse_stream(raw).collect().await;
+        assert!(matches!(
+            events.last(),
+            Some(Err(ProviderError::Overloaded))
+        ));
+    }
+
+    #[test]
+    fn build_request_applies_cache_control_to_first_two_system_blocks() {
+        let req = LlmRequest {
+            model: "claude-sonnet".into(),
+            system: vec!["env".into(), "agent".into(), "instructions".into()],
+            messages: vec![],
+            tools: vec![],
+            temperature: None,
+            max_tokens: None,
+            reasoning: None,
+            initiator: Initiator::User,
+        };
+        let body = build_request(&req);
+        let system = body["system"].as_array().unwrap();
+        assert_eq!(system.len(), 3);
+        assert!(system[0]["cache_control"].is_object());
+        assert!(system[1]["cache_control"].is_object());
+        assert!(system[2].get("cache_control").is_none());
+    }
+
+    #[test]
+    fn build_request_applies_cache_control_to_last_two_messages() {
+        let messages = vec![
+            WireMessage {
+                role: Role::User,
+                content: vec![WireContent::Text { text: "1".into() }],
+            },
+            WireMessage {
+                role: Role::Assistant,
+                content: vec![WireContent::Text { text: "2".into() }],
+            },
+            WireMessage {
+                role: Role::User,
+                content: vec![WireContent::Text { text: "3".into() }],
+            },
+        ];
+        let req = LlmRequest {
+            model: "claude-sonnet".into(),
+            system: vec![],
+            messages,
+            tools: vec![],
+            temperature: None,
+            max_tokens: None,
+            reasoning: None,
+            initiator: Initiator::User,
+        };
+        let body = build_request(&req);
+        let msgs = body["messages"].as_array().unwrap();
+        assert!(msgs[0]["content"][0].get("cache_control").is_none());
+        assert!(msgs[1]["content"][0]["cache_control"].is_object());
+        assert!(msgs[2]["content"][0]["cache_control"].is_object());
+    }
+
+    #[test]
+    fn build_request_maps_tool_schema_to_input_schema_key() {
+        let req = LlmRequest {
+            model: "claude-sonnet".into(),
+            system: vec![],
+            messages: vec![],
+            tools: vec![ToolSchema {
+                name: "read".into(),
+                description: "reads a file".into(),
+                parameters: json!({"type": "object"}),
+            }],
+            temperature: None,
+            max_tokens: None,
+            reasoning: None,
+            initiator: Initiator::User,
+        };
+        let body = build_request(&req);
+        assert_eq!(body["tools"][0]["name"], "read");
+        assert_eq!(body["tools"][0]["input_schema"], json!({"type": "object"}));
+    }
+
+    #[test]
+    fn build_request_includes_thinking_budget_when_reasoning_set() {
+        let req = LlmRequest {
+            model: "claude-sonnet".into(),
+            system: vec![],
+            messages: vec![],
+            tools: vec![],
+            temperature: None,
+            max_tokens: None,
+            reasoning: Some(ReasoningOpts {
+                effort: None,
+                budget_tokens: Some(2048),
+            }),
+            initiator: Initiator::User,
+        };
+        let body = build_request(&req);
+        assert_eq!(body["thinking"]["budget_tokens"], 2048);
+    }
+}

crates/harness-providers/src/codec/mod.rs

@@ -0,0 +1,3 @@
+pub mod anthropic;
+pub mod openai_chat;
+pub mod openai_responses;

crates/harness-providers/src/codec/openai_chat.rs

@@ -0,0 +1,502 @@
+//! Request builder + SSE decoder for OpenAI's `/chat/completions` streaming API.
+//!
+//! Also used as the fallback codec for OpenAI-compatible endpoints (including Copilot models
+//! whose `supported_endpoints` list `/chat/completions`).
+
+use std::collections::HashMap;
+
+use async_stream::try_stream;
+use eventsource_stream::Eventsource;
+use futures::Stream;
+use harness_core::llm::{
+    FinishReason, LlmEvent, LlmEventStream, LlmRequest, ProviderError, ReasoningEffort, Role,
+    WireContent,
+};
+use harness_core::types::TokenUsage;
+use serde_json::{json, Value};
+
+fn effort_str(effort: ReasoningEffort) -> &'static str {
+    match effort {
+        ReasoningEffort::Low => "low",
+        ReasoningEffort::Medium => "medium",
+        ReasoningEffort::High => "high",
+    }
+}
+
+fn role_str(role: Role) -> &'static str {
+    match role {
+        Role::System => "system",
+        Role::User => "user",
+        Role::Assistant => "assistant",
+        Role::Tool => "tool",
+    }
+}
+
+/// Flattens our grouped `WireMessage`s into the flat chat-completions message list. Tool
+/// results become their own `role: "tool"` messages (one per result), which is what the API
+/// expects regardless of how the engine grouped them.
+fn build_messages(system: &[String], messages: &[harness_core::llm::WireMessage]) -> Vec<Value> {
+    let mut out: Vec<Value> = Vec::new();
+
+    if !system.is_empty() {
+        out.push(json!({"role": "system", "content": system.join("\n\n")}));
+    }
+
+    for m in messages {
+        // Tool results are always emitted as standalone `tool` messages.
+        for c in &m.content {
+            if let WireContent::ToolResult {
+                call_id, output, ..
+            } = c
+            {
+                out.push(json!({
+                    "role": "tool",
+                    "tool_call_id": call_id,
+                    "content": output,
+                }));
+            }
+        }
+
+        let mut text = String::new();
+        let mut images: Vec<Value> = Vec::new();
+        let mut tool_calls: Vec<Value> = Vec::new();
+        for c in &m.content {
+            match c {
+                WireContent::Text { text: t } => {
+                    if !text.is_empty() {
+                        text.push('\n');
+                    }
+                    text.push_str(t);
+                }
+                WireContent::ToolCall {
+                    call_id,
+                    name,
+                    input,
+                } => tool_calls.push(json!({
+                    "id": call_id,
+                    "type": "function",
+                    "function": {"name": name, "arguments": input.to_string()},
+                })),
+                WireContent::Image { mime_type, data } => images.push(json!({
+                    "type": "image_url",
+                    "image_url": {"url": format!("data:{mime_type};base64,{data}")},
+                })),
+                WireContent::ToolResult { .. } => {} // handled above
+            }
+        }
+
+        // A message that carried nothing but tool results contributes no further entry.
+        if text.is_empty() && images.is_empty() && tool_calls.is_empty() {
+            continue;
+        }
+
+        let mut msg = json!({"role": role_str(m.role)});
+        if images.is_empty() {
+            msg["content"] = json!(text);
+        } else {
+            let mut parts = vec![json!({"type": "text", "text": text})];
+            parts.extend(images);
+            msg["content"] = json!(parts);
+        }
+        if !tool_calls.is_empty() {
+            msg["tool_calls"] = json!(tool_calls);
+        }
+        out.push(msg);
+    }
+
+    out
+}
+
+pub fn build_request(req: &LlmRequest) -> Value {
+    let mut body = json!({
+        "model": req.model,
+        "messages": build_messages(&req.system, &req.messages),
+        "stream": true,
+        "stream_options": {"include_usage": true},
+    });
+
+    if !req.tools.is_empty() {
+        body["tools"] = Value::Array(
+            req.tools
+                .iter()
+                .map(|t| {
+                    json!({
+                        "type": "function",
+                        "function": {
+                            "name": t.name,
+                            "description": t.description,
+                            "parameters": t.parameters,
+                        },
+                    })
+                })
+                .collect(),
+        );
+    }
+    if let Some(temp) = req.temperature {
+        body["temperature"] = json!(temp);
+    }
+    if let Some(max) = req.max_tokens {
+        body["max_completion_tokens"] = json!(max);
+    }
+    if let Some(effort) = req.reasoning.as_ref().and_then(|r| r.effort) {
+        body["reasoning_effort"] = json!(effort_str(effort));
+    }
+    body
+}
+
+fn map_finish_reason(reason: &str) -> FinishReason {
+    match reason {
+        "stop" => FinishReason::Stop,
+        "tool_calls" | "function_call" => FinishReason::ToolCalls,
+        "length" => FinishReason::Length,
+        "content_filter" => FinishReason::ContentFilter,
+        other => FinishReason::Unknown(other.to_string()),
+    }
+}
+
+#[derive(Default)]
+struct ToolAccum {
+    call_id: String,
+    name: String,
+    args: String,
+}
+
+/// Decodes a chat-completions SSE byte stream into normalized `LlmEvent`s. Tool-call deltas
+/// are accumulated by their `index` and flushed as `ToolCall`s once the stream ends.
+pub fn decode<S, E>(byte_stream: S) -> LlmEventStream
+where
+    S: Stream<Item = Result<bytes::Bytes, E>> + Send + 'static,
+    E: std::error::Error + Send + Sync + 'static,
+{
+    let events = byte_stream.eventsource();
+    Box::pin(try_stream! {
+        futures::pin_mut!(events);
+        let mut usage = TokenUsage::default();
+        let mut reason = FinishReason::Stop;
+        let mut text_open = false;
+        let mut reasoning_open = false;
+        let mut tools: HashMap<u64, ToolAccum> = HashMap::new();
+        let mut tool_order: Vec<u64> = Vec::new();
+
+        while let Some(item) = futures::StreamExt::next(&mut events).await {
+            let event = item.map_err(|e| ProviderError::Decode(e.to_string()))?;
+            let data = event.data.trim();
+            if data.is_empty() {
+                continue;
+            }
+            if data == "[DONE]" {
+                break;
+            }
+            let value: Value = serde_json::from_str(data)
+                .map_err(|e| ProviderError::Decode(format!("{e}: {data}")))?;
+
+            if let Some(u) = value.get("usage").filter(|u| u.is_object()) {
+                usage.input = u["prompt_tokens"].as_u64().unwrap_or(usage.input);
+                usage.output = u["completion_tokens"].as_u64().unwrap_or(usage.output);
+                usage.reasoning = u["completion_tokens_details"]["reasoning_tokens"]
+                    .as_u64()
+                    .unwrap_or(usage.reasoning);
+                usage.cache_read = u["prompt_tokens_details"]["cached_tokens"]
+                    .as_u64()
+                    .unwrap_or(usage.cache_read);
+            }
+
+            let choice = &value["choices"][0];
+            let delta = &choice["delta"];
+
+            if let Some(rc) = delta["reasoning_content"].as_str().filter(|s| !s.is_empty()) {
+                if !reasoning_open {
+                    reasoning_open = true;
+                    yield LlmEvent::ReasoningStart { id: "reasoning".into() };
+                }
+                yield LlmEvent::ReasoningDelta { id: "reasoning".into(), text: rc.to_string() };
+            }
+
+            if let Some(text) = delta["content"].as_str().filter(|s| !s.is_empty()) {
+                if reasoning_open {
+                    reasoning_open = false;
+                    yield LlmEvent::ReasoningEnd { id: "reasoning".into(), signature: None };
+                }
+                if !text_open {
+                    text_open = true;
+                    yield LlmEvent::TextStart { id: "0".into() };
+                }
+                yield LlmEvent::TextDelta { id: "0".into(), text: text.to_string() };
+            }
+
+            if let Some(calls) = delta["tool_calls"].as_array() {
+                for call in calls {
+                    let index = call["index"].as_u64().unwrap_or(0);
+                    let entry = tools.entry(index).or_insert_with(|| {
+                        tool_order.push(index);
+                        ToolAccum::default()
+                    });
+                    if let Some(id) = call["id"].as_str().filter(|s| !s.is_empty()) {
+                        entry.call_id = id.to_string();
+                    }
+                    if let Some(name) = call["function"]["name"].as_str().filter(|s| !s.is_empty()) {
+                        entry.name = name.to_string();
+                        yield LlmEvent::ToolInputStart {
+                            call_id: entry.call_id.clone(),
+                            name: entry.name.clone(),
+                        };
+                    }
+                    if let Some(args) = call["function"]["arguments"].as_str().filter(|s| !s.is_empty()) {
+                        entry.args.push_str(args);
+                        yield LlmEvent::ToolInputDelta {
+                            call_id: entry.call_id.clone(),
+                            json: args.to_string(),
+                        };
+                    }
+                }
+            }
+
+            if let Some(fr) = choice["finish_reason"].as_str() {
+                reason = map_finish_reason(fr);
+            }
+        }
+
+        if reasoning_open {
+            yield LlmEvent::ReasoningEnd { id: "reasoning".into(), signature: None };
+        }
+        if text_open {
+            yield LlmEvent::TextEnd { id: "0".into() };
+        }
+        for index in tool_order {
+            if let Some(entry) = tools.remove(&index) {
+                let input: Value = if entry.args.trim().is_empty() {
+                    json!({})
+                } else {
+                    serde_json::from_str(&entry.args).unwrap_or(Value::Null)
+                };
+                yield LlmEvent::ToolCall { call_id: entry.call_id, name: entry.name, input };
+            }
+        }
+        yield LlmEvent::Finish { reason, usage };
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use futures::StreamExt;
+    use harness_core::llm::{Initiator, ReasoningOpts, ToolSchema, WireMessage};
+
+    fn sse_stream(raw: &'static str) -> LlmEventStream {
+        let chunks: Vec<Result<bytes::Bytes, std::io::Error>> =
+            vec![Ok(bytes::Bytes::from_static(raw.as_bytes()))];
+        decode(futures::stream::iter(chunks))
+    }
+
+    fn req(model: &str) -> LlmRequest {
+        LlmRequest {
+            model: model.into(),
+            system: vec![],
+            messages: vec![],
+            tools: vec![],
+            temperature: None,
+            max_tokens: None,
+            reasoning: None,
+            initiator: Initiator::User,
+        }
+    }
+
+    #[tokio::test]
+    async fn decodes_text_only_response() {
+        let raw = concat!(
+            "data: {\"choices\":[{\"delta\":{\"content\":\"Hel\"},\"finish_reason\":null}]}\n\n",
+            "data: {\"choices\":[{\"delta\":{\"content\":\"lo\"},\"finish_reason\":null}]}\n\n",
+            "data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n",
+            "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":10,\"completion_tokens\":5}}\n\n",
+            "data: [DONE]\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::TextStart { id: "0".into() },
+                LlmEvent::TextDelta {
+                    id: "0".into(),
+                    text: "Hel".into()
+                },
+                LlmEvent::TextDelta {
+                    id: "0".into(),
+                    text: "lo".into()
+                },
+                LlmEvent::TextEnd { id: "0".into() },
+                LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: TokenUsage {
+                        input: 10,
+                        output: 5,
+                        ..Default::default()
+                    },
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn decodes_tool_call_accumulated_by_index() {
+        let raw = concat!(
+            "data: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"index\":0,\"id\":\"call_1\",\"function\":{\"name\":\"read\",\"arguments\":\"\"}}]}}]}\n\n",
+            "data: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\"{\\\"file\\\"\"}}]}}]}\n\n",
+            "data: {\"choices\":[{\"delta\":{\"tool_calls\":[{\"index\":0,\"function\":{\"arguments\":\":\\\"a.txt\\\"}\"}}]}}]}\n\n",
+            "data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\"}],\"usage\":{\"prompt_tokens\":1,\"completion_tokens\":8}}\n\n",
+            "data: [DONE]\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::ToolInputStart {
+                    call_id: "call_1".into(),
+                    name: "read".into()
+                },
+                LlmEvent::ToolInputDelta {
+                    call_id: "call_1".into(),
+                    json: "{\"file\"".into()
+                },
+                LlmEvent::ToolInputDelta {
+                    call_id: "call_1".into(),
+                    json: ":\"a.txt\"}".into()
+                },
+                LlmEvent::ToolCall {
+                    call_id: "call_1".into(),
+                    name: "read".into(),
+                    input: json!({"file": "a.txt"}),
+                },
+                LlmEvent::Finish {
+                    reason: FinishReason::ToolCalls,
+                    usage: TokenUsage {
+                        input: 1,
+                        output: 8,
+                        ..Default::default()
+                    },
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn decodes_reasoning_content_before_text() {
+        let raw = concat!(
+            "data: {\"choices\":[{\"delta\":{\"reasoning_content\":\"hmm\"}}]}\n\n",
+            "data: {\"choices\":[{\"delta\":{\"content\":\"answer\"}}]}\n\n",
+            "data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n",
+            "data: [DONE]\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::ReasoningStart {
+                    id: "reasoning".into()
+                },
+                LlmEvent::ReasoningDelta {
+                    id: "reasoning".into(),
+                    text: "hmm".into()
+                },
+                LlmEvent::ReasoningEnd {
+                    id: "reasoning".into(),
+                    signature: None
+                },
+                LlmEvent::TextStart { id: "0".into() },
+                LlmEvent::TextDelta {
+                    id: "0".into(),
+                    text: "answer".into()
+                },
+                LlmEvent::TextEnd { id: "0".into() },
+                LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: TokenUsage::default()
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn build_request_wraps_tools_in_function_envelope() {
+        let mut r = req("gpt-4o");
+        r.tools = vec![ToolSchema {
+            name: "read".into(),
+            description: "reads a file".into(),
+            parameters: json!({"type": "object"}),
+        }];
+        let body = build_request(&r);
+        assert_eq!(body["tools"][0]["type"], "function");
+        assert_eq!(body["tools"][0]["function"]["name"], "read");
+        assert_eq!(
+            body["tools"][0]["function"]["parameters"],
+            json!({"type": "object"})
+        );
+        assert_eq!(body["stream_options"]["include_usage"], true);
+    }
+
+    #[test]
+    fn build_request_prepends_system_message() {
+        let mut r = req("gpt-4o");
+        r.system = vec!["env".into(), "agent".into()];
+        r.messages = vec![WireMessage {
+            role: Role::User,
+            content: vec![WireContent::Text { text: "hi".into() }],
+        }];
+        let body = build_request(&r);
+        let msgs = body["messages"].as_array().unwrap();
+        assert_eq!(msgs[0]["role"], "system");
+        assert_eq!(msgs[0]["content"], "env\n\nagent");
+        assert_eq!(msgs[1]["role"], "user");
+        assert_eq!(msgs[1]["content"], "hi");
+    }
+
+    #[test]
+    fn build_request_expands_tool_results_to_tool_messages() {
+        let mut r = req("gpt-4o");
+        r.messages = vec![WireMessage {
+            role: Role::Tool,
+            content: vec![WireContent::ToolResult {
+                call_id: "call_1".into(),
+                output: "contents".into(),
+                is_error: false,
+            }],
+        }];
+        let body = build_request(&r);
+        let msgs = body["messages"].as_array().unwrap();
+        assert_eq!(msgs.len(), 1);
+        assert_eq!(msgs[0]["role"], "tool");
+        assert_eq!(msgs[0]["tool_call_id"], "call_1");
+        assert_eq!(msgs[0]["content"], "contents");
+    }
+
+    #[test]
+    fn build_request_emits_assistant_tool_calls() {
+        let mut r = req("gpt-4o");
+        r.messages = vec![WireMessage {
+            role: Role::Assistant,
+            content: vec![WireContent::ToolCall {
+                call_id: "call_1".into(),
+                name: "read".into(),
+                input: json!({"file": "a.txt"}),
+            }],
+        }];
+        let body = build_request(&r);
+        let msgs = body["messages"].as_array().unwrap();
+        assert_eq!(msgs[0]["tool_calls"][0]["id"], "call_1");
+        assert_eq!(msgs[0]["tool_calls"][0]["function"]["name"], "read");
+        assert_eq!(
+            msgs[0]["tool_calls"][0]["function"]["arguments"],
+            "{\"file\":\"a.txt\"}"
+        );
+    }
+
+    #[test]
+    fn build_request_sets_reasoning_effort() {
+        let mut r = req("o3");
+        r.reasoning = Some(ReasoningOpts {
+            effort: Some(ReasoningEffort::High),
+            budget_tokens: None,
+        });
+        let body = build_request(&r);
+        assert_eq!(body["reasoning_effort"], "high");
+    }
+}

crates/harness-providers/src/codec/openai_responses.rs

@@ -0,0 +1,402 @@
+//! Request builder + SSE decoder for OpenAI's `/responses` streaming API.
+//!
+//! The responses API is the preferred surface for `gpt-*` / `o-*` models: it carries
+//! reasoning items natively and reports reasoning-token usage separately.
+
+use std::collections::HashMap;
+
+use async_stream::try_stream;
+use eventsource_stream::Eventsource;
+use futures::Stream;
+use harness_core::llm::{
+    FinishReason, LlmEvent, LlmEventStream, LlmRequest, ProviderError, ReasoningEffort, Role,
+    WireContent,
+};
+use harness_core::types::TokenUsage;
+use serde_json::{json, Value};
+
+fn effort_str(effort: ReasoningEffort) -> &'static str {
+    match effort {
+        ReasoningEffort::Low => "low",
+        ReasoningEffort::Medium => "medium",
+        ReasoningEffort::High => "high",
+    }
+}
+
+/// Builds the `input` item list. Text/image content become `message` items; tool calls and
+/// their results become `function_call` / `function_call_output` items (the responses API
+/// keeps these as top-level items rather than nesting them inside messages).
+fn build_input(messages: &[harness_core::llm::WireMessage]) -> Vec<Value> {
+    let mut out: Vec<Value> = Vec::new();
+    for m in messages {
+        let (role, text_type) = match m.role {
+            Role::Assistant => ("assistant", "output_text"),
+            _ => ("user", "input_text"),
+        };
+        let mut content_parts: Vec<Value> = Vec::new();
+        for c in &m.content {
+            match c {
+                WireContent::Text { text } => {
+                    content_parts.push(json!({"type": text_type, "text": text}));
+                }
+                WireContent::Image { mime_type, data } => {
+                    content_parts.push(json!({
+                        "type": "input_image",
+                        "image_url": format!("data:{mime_type};base64,{data}"),
+                    }));
+                }
+                WireContent::ToolCall {
+                    call_id,
+                    name,
+                    input,
+                } => out.push(json!({
+                    "type": "function_call",
+                    "call_id": call_id,
+                    "name": name,
+                    "arguments": input.to_string(),
+                })),
+                WireContent::ToolResult {
+                    call_id, output, ..
+                } => out.push(json!({
+                    "type": "function_call_output",
+                    "call_id": call_id,
+                    "output": output,
+                })),
+            }
+        }
+        if !content_parts.is_empty() {
+            out.push(json!({"type": "message", "role": role, "content": content_parts}));
+        }
+    }
+    out
+}
+
+pub fn build_request(req: &LlmRequest) -> Value {
+    let mut body = json!({
+        "model": req.model,
+        "input": build_input(&req.messages),
+        "stream": true,
+        "store": false,
+    });
+
+    if !req.system.is_empty() {
+        body["instructions"] = json!(req.system.join("\n\n"));
+    }
+    if !req.tools.is_empty() {
+        body["tools"] = Value::Array(
+            req.tools
+                .iter()
+                .map(|t| {
+                    json!({
+                        "type": "function",
+                        "name": t.name,
+                        "description": t.description,
+                        "parameters": t.parameters,
+                    })
+                })
+                .collect(),
+        );
+    }
+    if let Some(temp) = req.temperature {
+        body["temperature"] = json!(temp);
+    }
+    if let Some(max) = req.max_tokens {
+        body["max_output_tokens"] = json!(max);
+    }
+    if let Some(effort) = req.reasoning.as_ref().and_then(|r| r.effort) {
+        body["reasoning"] = json!({"effort": effort_str(effort), "summary": "auto"});
+    }
+    body
+}
+
+fn map_status(status: Option<&str>, had_tool_call: bool) -> FinishReason {
+    match status {
+        Some("completed") if had_tool_call => FinishReason::ToolCalls,
+        Some("completed") => FinishReason::Stop,
+        Some("incomplete") => FinishReason::Length,
+        Some(other) => FinishReason::Unknown(other.to_string()),
+        None if had_tool_call => FinishReason::ToolCalls,
+        None => FinishReason::Stop,
+    }
+}
+
+/// Tracks which normalized stream an `item_id` belongs to so deltas route correctly.
+enum ItemKind {
+    Text,
+    Reasoning,
+    FunctionCall { call_id: String, name: String },
+}
+
+/// Decodes a `/responses` SSE byte stream into normalized `LlmEvent`s.
+pub fn decode<S, E>(byte_stream: S) -> LlmEventStream
+where
+    S: Stream<Item = Result<bytes::Bytes, E>> + Send + 'static,
+    E: std::error::Error + Send + Sync + 'static,
+{
+    let events = byte_stream.eventsource();
+    Box::pin(try_stream! {
+        futures::pin_mut!(events);
+        let mut items: HashMap<String, ItemKind> = HashMap::new();
+        let mut fn_args: HashMap<String, String> = HashMap::new();
+        let mut usage = TokenUsage::default();
+        let mut had_tool_call = false;
+
+        while let Some(item) = futures::StreamExt::next(&mut events).await {
+            let event = item.map_err(|e| ProviderError::Decode(e.to_string()))?;
+            if event.data.trim().is_empty() {
+                continue;
+            }
+            let value: Value = serde_json::from_str(&event.data)
+                .map_err(|e| ProviderError::Decode(format!("{e}: {}", event.data)))?;
+            let kind = value["type"].as_str().unwrap_or_default();
+
+            match kind {
+                "response.output_item.added" => {
+                    let item = &value["item"];
+                    let id = item["id"].as_str().unwrap_or_default().to_string();
+                    match item["type"].as_str().unwrap_or_default() {
+                        "message" => {
+                            items.insert(id.clone(), ItemKind::Text);
+                            yield LlmEvent::TextStart { id };
+                        }
+                        "reasoning" => {
+                            items.insert(id.clone(), ItemKind::Reasoning);
+                            yield LlmEvent::ReasoningStart { id };
+                        }
+                        "function_call" => {
+                            let call_id = item["call_id"].as_str().unwrap_or_default().to_string();
+                            let name = item["name"].as_str().unwrap_or_default().to_string();
+                            fn_args.insert(id.clone(), String::new());
+                            items.insert(id, ItemKind::FunctionCall { call_id: call_id.clone(), name: name.clone() });
+                            had_tool_call = true;
+                            yield LlmEvent::ToolInputStart { call_id, name };
+                        }
+                        _ => {}
+                    }
+                }
+                "response.output_text.delta" => {
+                    let id = value["item_id"].as_str().unwrap_or_default().to_string();
+                    let text = value["delta"].as_str().unwrap_or_default().to_string();
+                    yield LlmEvent::TextDelta { id, text };
+                }
+                "response.reasoning_summary_text.delta" => {
+                    let id = value["item_id"].as_str().unwrap_or_default().to_string();
+                    let text = value["delta"].as_str().unwrap_or_default().to_string();
+                    yield LlmEvent::ReasoningDelta { id, text };
+                }
+                "response.function_call_arguments.delta" => {
+                    let id = value["item_id"].as_str().unwrap_or_default().to_string();
+                    let delta = value["delta"].as_str().unwrap_or_default();
+                    if let (Some(buf), Some(ItemKind::FunctionCall { call_id, .. })) =
+                        (fn_args.get_mut(&id), items.get(&id))
+                    {
+                        buf.push_str(delta);
+                        yield LlmEvent::ToolInputDelta { call_id: call_id.clone(), json: delta.to_string() };
+                    }
+                }
+                "response.output_item.done" => {
+                    let id = value["item"]["id"].as_str().unwrap_or_default().to_string();
+                    match items.remove(&id) {
+                        Some(ItemKind::Text) => yield LlmEvent::TextEnd { id },
+                        Some(ItemKind::Reasoning) => yield LlmEvent::ReasoningEnd { id, signature: None },
+                        Some(ItemKind::FunctionCall { call_id, name }) => {
+                            let raw = fn_args.remove(&id).unwrap_or_default();
+                            let input: Value = if raw.trim().is_empty() {
+                                json!({})
+                            } else {
+                                serde_json::from_str(&raw).unwrap_or(Value::Null)
+                            };
+                            yield LlmEvent::ToolCall { call_id, name, input };
+                        }
+                        None => {}
+                    }
+                }
+                "response.completed" | "response.incomplete" => {
+                    let response = &value["response"];
+                    let u = &response["usage"];
+                    usage.input = u["input_tokens"].as_u64().unwrap_or(0);
+                    usage.output = u["output_tokens"].as_u64().unwrap_or(0);
+                    usage.reasoning = u["output_tokens_details"]["reasoning_tokens"].as_u64().unwrap_or(0);
+                    usage.cache_read = u["input_tokens_details"]["cached_tokens"].as_u64().unwrap_or(0);
+                    let status = response["status"].as_str();
+                    yield LlmEvent::Finish { reason: map_status(status, had_tool_call), usage };
+                }
+                "error" | "response.failed" => {
+                    let message = value["response"]["error"]["message"]
+                        .as_str()
+                        .or_else(|| value["message"].as_str())
+                        .unwrap_or("unknown error")
+                        .to_string();
+                    Err(ProviderError::Http { status: 0, body: message })?;
+                }
+                _ => {}
+            }
+        }
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use futures::StreamExt;
+    use harness_core::llm::{Initiator, ReasoningOpts, ToolSchema, WireMessage};
+
+    fn sse_stream(raw: &'static str) -> LlmEventStream {
+        let chunks: Vec<Result<bytes::Bytes, std::io::Error>> =
+            vec![Ok(bytes::Bytes::from_static(raw.as_bytes()))];
+        decode(futures::stream::iter(chunks))
+    }
+
+    fn req(model: &str) -> LlmRequest {
+        LlmRequest {
+            model: model.into(),
+            system: vec![],
+            messages: vec![],
+            tools: vec![],
+            temperature: None,
+            max_tokens: None,
+            reasoning: None,
+            initiator: Initiator::User,
+        }
+    }
+
+    #[tokio::test]
+    async fn decodes_text_response() {
+        let raw = concat!(
+            "data: {\"type\":\"response.output_item.added\",\"item\":{\"id\":\"msg_1\",\"type\":\"message\"}}\n\n",
+            "data: {\"type\":\"response.output_text.delta\",\"item_id\":\"msg_1\",\"delta\":\"Hello\"}\n\n",
+            "data: {\"type\":\"response.output_item.done\",\"item\":{\"id\":\"msg_1\"}}\n\n",
+            "data: {\"type\":\"response.completed\",\"response\":{\"status\":\"completed\",\"usage\":{\"input_tokens\":10,\"output_tokens\":5}}}\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::TextStart { id: "msg_1".into() },
+                LlmEvent::TextDelta {
+                    id: "msg_1".into(),
+                    text: "Hello".into()
+                },
+                LlmEvent::TextEnd { id: "msg_1".into() },
+                LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: TokenUsage {
+                        input: 10,
+                        output: 5,
+                        ..Default::default()
+                    },
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn decodes_function_call_with_reasoning_tokens() {
+        let raw = concat!(
+            "data: {\"type\":\"response.output_item.added\",\"item\":{\"id\":\"fc_1\",\"type\":\"function_call\",\"call_id\":\"call_1\",\"name\":\"read\"}}\n\n",
+            "data: {\"type\":\"response.function_call_arguments.delta\",\"item_id\":\"fc_1\",\"delta\":\"{\\\"file\\\":\"}\n\n",
+            "data: {\"type\":\"response.function_call_arguments.delta\",\"item_id\":\"fc_1\",\"delta\":\"\\\"a.txt\\\"}\"}\n\n",
+            "data: {\"type\":\"response.output_item.done\",\"item\":{\"id\":\"fc_1\"}}\n\n",
+            "data: {\"type\":\"response.completed\",\"response\":{\"status\":\"completed\",\"usage\":{\"input_tokens\":3,\"output_tokens\":9,\"output_tokens_details\":{\"reasoning_tokens\":4}}}}\n\n",
+        );
+        let events: Vec<LlmEvent> = sse_stream(raw).map(|e| e.unwrap()).collect().await;
+        assert_eq!(
+            events,
+            vec![
+                LlmEvent::ToolInputStart {
+                    call_id: "call_1".into(),
+                    name: "read".into()
+                },
+                LlmEvent::ToolInputDelta {
+                    call_id: "call_1".into(),
+                    json: "{\"file\":".into()
+                },
+                LlmEvent::ToolInputDelta {
+                    call_id: "call_1".into(),
+                    json: "\"a.txt\"}".into()
+                },
+                LlmEvent::ToolCall {
+                    call_id: "call_1".into(),
+                    name: "read".into(),
+                    input: json!({"file": "a.txt"}),
+                },
+                LlmEvent::Finish {
+                    reason: FinishReason::ToolCalls,
+                    usage: TokenUsage {
+                        input: 3,
+                        output: 9,
+                        reasoning: 4,
+                        ..Default::default()
+                    },
+                },
+            ]
+        );
+    }
+
+    #[tokio::test]
+    async fn failed_response_surfaces_as_err() {
+        let raw = "data: {\"type\":\"response.failed\",\"response\":{\"error\":{\"message\":\"boom\"}}}\n\n";
+        let events: Vec<Result<LlmEvent, ProviderError>> = sse_stream(raw).collect().await;
+        assert!(
+            matches!(events.last(), Some(Err(ProviderError::Http { body, .. })) if body == "boom")
+        );
+    }
+
+    #[test]
+    fn build_request_puts_system_in_instructions() {
+        let mut r = req("gpt-5");
+        r.system = vec!["env".into(), "agent".into()];
+        let body = build_request(&r);
+        assert_eq!(body["instructions"], "env\n\nagent");
+        assert!(body.get("messages").is_none());
+    }
+
+    #[test]
+    fn build_request_flattens_tool_calls_and_results_to_items() {
+        let mut r = req("gpt-5");
+        r.messages = vec![
+            WireMessage {
+                role: Role::Assistant,
+                content: vec![WireContent::ToolCall {
+                    call_id: "call_1".into(),
+                    name: "read".into(),
+                    input: json!({"file": "a.txt"}),
+                }],
+            },
+            WireMessage {
+                role: Role::Tool,
+                content: vec![WireContent::ToolResult {
+                    call_id: "call_1".into(),
+                    output: "contents".into(),
+                    is_error: false,
+                }],
+            },
+        ];
+        let body = build_request(&r);
+        let input = body["input"].as_array().unwrap();
+        assert_eq!(input[0]["type"], "function_call");
+        assert_eq!(input[0]["call_id"], "call_1");
+        assert_eq!(input[0]["arguments"], "{\"file\":\"a.txt\"}");
+        assert_eq!(input[1]["type"], "function_call_output");
+        assert_eq!(input[1]["output"], "contents");
+    }
+
+    #[test]
+    fn build_request_uses_flat_function_tool_shape_and_reasoning() {
+        let mut r = req("o3");
+        r.tools = vec![ToolSchema {
+            name: "read".into(),
+            description: "reads a file".into(),
+            parameters: json!({"type": "object"}),
+        }];
+        r.reasoning = Some(ReasoningOpts {
+            effort: Some(ReasoningEffort::Medium),
+            budget_tokens: None,
+        });
+        let body = build_request(&r);
+        assert_eq!(body["tools"][0]["type"], "function");
+        assert_eq!(body["tools"][0]["name"], "read");
+        assert_eq!(body["reasoning"]["effort"], "medium");
+        assert_eq!(body["reasoning"]["summary"], "auto");
+    }
+}

crates/harness-providers/src/copilot/device_flow.rs

@@ -0,0 +1,209 @@
+//! GitHub OAuth device flow (RFC 8628) for Copilot login.
+//!
+//! ai-harness must register its own GitHub OAuth app and supply its client id (we do not
+//! hardcode opencode's). The client id is passed in by the caller — see [`request_device_code`].
+
+use std::time::Duration;
+
+use serde::Deserialize;
+use serde_json::Value;
+
+const DEVICE_CODE_URL: &str = "https://github.com/login/device/code";
+const ACCESS_TOKEN_URL: &str = "https://github.com/login/oauth/access_token";
+const GRANT_TYPE: &str = "urn:ietf:params:oauth:grant-type:device_code";
+/// Minimum scope needed to call the Copilot token-exchange endpoint.
+pub const SCOPE: &str = "read:user";
+
+#[derive(Debug, Clone, Deserialize)]
+pub struct DeviceCode {
+    pub device_code: String,
+    pub user_code: String,
+    pub verification_uri: String,
+    /// Seconds between polls; GitHub requires honoring this and any `slow_down` bumps.
+    #[serde(default = "default_interval")]
+    pub interval: u64,
+    #[serde(default)]
+    pub expires_in: u64,
+}
+
+fn default_interval() -> u64 {
+    5
+}
+
+/// Result of one poll of the access-token endpoint.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum PollOutcome {
+    /// The user hasn't authorized yet; keep polling at the current interval.
+    Pending,
+    /// GitHub asked us to slow down; add 5s to the interval (RFC 8628).
+    SlowDown,
+    /// Authorization complete.
+    Success { access_token: String },
+    /// Terminal failure (expired code, denied, unknown error) with a human-readable reason.
+    Failed(String),
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum DeviceFlowError {
+    #[error("network: {0}")]
+    Network(String),
+    #[error("unexpected response: {0}")]
+    Unexpected(String),
+}
+
+/// Pure mapping of an access-token poll response body to a [`PollOutcome`], so the state
+/// machine is testable without a live GitHub.
+pub fn parse_poll_response(body: &Value) -> PollOutcome {
+    if let Some(token) = body["access_token"].as_str() {
+        return PollOutcome::Success {
+            access_token: token.to_string(),
+        };
+    }
+    match body["error"].as_str() {
+        Some("authorization_pending") => PollOutcome::Pending,
+        Some("slow_down") => PollOutcome::SlowDown,
+        Some(other) => {
+            let desc = body["error_description"].as_str().unwrap_or(other);
+            PollOutcome::Failed(desc.to_string())
+        }
+        None => PollOutcome::Failed("no access_token and no error in response".to_string()),
+    }
+}
+
+/// Applies a `slow_down` to a poll interval per RFC 8628 (+5s).
+pub fn bump_interval(interval: u64) -> u64 {
+    interval + 5
+}
+
+/// Step 1: request a device + user code for `client_id`.
+pub async fn request_device_code(
+    client: &reqwest::Client,
+    client_id: &str,
+) -> Result<DeviceCode, DeviceFlowError> {
+    let resp = client
+        .post(DEVICE_CODE_URL)
+        .header("accept", "application/json")
+        .json(&serde_json::json!({"client_id": client_id, "scope": SCOPE}))
+        .send()
+        .await
+        .map_err(|e| DeviceFlowError::Network(e.to_string()))?;
+    let value: Value = resp
+        .json()
+        .await
+        .map_err(|e| DeviceFlowError::Network(e.to_string()))?;
+    serde_json::from_value(value.clone())
+        .map_err(|_| DeviceFlowError::Unexpected(value.to_string()))
+}
+
+/// Step 2 (single poll): exchange the device code for an access token, once.
+pub async fn poll_once(
+    client: &reqwest::Client,
+    client_id: &str,
+    device_code: &str,
+) -> Result<PollOutcome, DeviceFlowError> {
+    let resp = client
+        .post(ACCESS_TOKEN_URL)
+        .header("accept", "application/json")
+        .json(&serde_json::json!({
+            "client_id": client_id,
+            "device_code": device_code,
+            "grant_type": GRANT_TYPE,
+        }))
+        .send()
+        .await
+        .map_err(|e| DeviceFlowError::Network(e.to_string()))?;
+    let value: Value = resp
+        .json()
+        .await
+        .map_err(|e| DeviceFlowError::Network(e.to_string()))?;
+    Ok(parse_poll_response(&value))
+}
+
+/// Step 2 (full loop): polls until success or terminal failure, honoring `interval` and
+/// `slow_down`. `sleep` is injected so tests can drive it without real time.
+pub async fn poll_for_token<S, Fut>(
+    client: &reqwest::Client,
+    client_id: &str,
+    device: &DeviceCode,
+    sleep: S,
+) -> Result<String, DeviceFlowError>
+where
+    S: Fn(Duration) -> Fut,
+    Fut: std::future::Future<Output = ()>,
+{
+    let mut interval = device.interval;
+    loop {
+        sleep(Duration::from_secs(interval)).await;
+        match poll_once(client, client_id, &device.device_code).await? {
+            PollOutcome::Pending => {}
+            PollOutcome::SlowDown => interval = bump_interval(interval),
+            PollOutcome::Success { access_token } => return Ok(access_token),
+            PollOutcome::Failed(reason) => return Err(DeviceFlowError::Unexpected(reason)),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn parses_success() {
+        let outcome =
+            parse_poll_response(&json!({"access_token": "gho_abc", "token_type": "bearer"}));
+        assert_eq!(
+            outcome,
+            PollOutcome::Success {
+                access_token: "gho_abc".into()
+            }
+        );
+    }
+
+    #[test]
+    fn parses_pending_and_slow_down() {
+        assert_eq!(
+            parse_poll_response(&json!({"error": "authorization_pending"})),
+            PollOutcome::Pending
+        );
+        assert_eq!(
+            parse_poll_response(&json!({"error": "slow_down", "interval": 10})),
+            PollOutcome::SlowDown
+        );
+    }
+
+    #[test]
+    fn parses_terminal_errors_with_description() {
+        match parse_poll_response(
+            &json!({"error": "expired_token", "error_description": "code expired"}),
+        ) {
+            PollOutcome::Failed(msg) => assert_eq!(msg, "code expired"),
+            other => panic!("expected Failed, got {other:?}"),
+        }
+        assert!(matches!(
+            parse_poll_response(&json!({"error": "access_denied"})),
+            PollOutcome::Failed(_)
+        ));
+        assert!(matches!(
+            parse_poll_response(&json!({})),
+            PollOutcome::Failed(_)
+        ));
+    }
+
+    #[test]
+    fn slow_down_adds_five_seconds() {
+        assert_eq!(bump_interval(5), 10);
+        assert_eq!(bump_interval(10), 15);
+    }
+
+    #[test]
+    fn device_code_defaults_interval() {
+        let dc: DeviceCode = serde_json::from_value(json!({
+            "device_code": "d",
+            "user_code": "WXYZ-1234",
+            "verification_uri": "https://github.com/login/device"
+        }))
+        .unwrap();
+        assert_eq!(dc.interval, 5);
+    }
+}

crates/harness-providers/src/copilot/mod.rs

@@ -0,0 +1,10 @@
+//! GitHub Copilot: OAuth device-flow login, token exchange, and a provider that multiplexes
+//! the chat/responses/anthropic codecs behind `api.githubcopilot.com`.
+
+pub mod device_flow;
+pub mod provider;
+pub mod token;
+
+pub use device_flow::{DeviceCode, PollOutcome};
+pub use provider::{CopilotCodec, CopilotModel, CopilotProvider};
+pub use token::{CopilotToken, TokenProvider};

crates/harness-providers/src/copilot/provider.rs

@@ -0,0 +1,296 @@
+//! Copilot provider: multiplexes all three wire codecs behind `api.githubcopilot.com`.
+//!
+//! Each model's `supported_endpoints` (from `GET /models`) decides which codec/endpoint to
+//! use. The token comes from [`super::token::TokenProvider`]. Network paths here are the
+//! flagged live-verification risk — the routing/header logic is unit-tested; end-to-end
+//! streaming must be checked against the live API.
+
+use std::collections::HashMap;
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use async_trait::async_trait;
+use harness_core::llm::{
+    Initiator, LlmEventStream, LlmRequest, Provider, ProviderError, WireContent,
+};
+use harness_core::types::ModelInfo;
+use serde_json::Value;
+use tokio_util::sync::CancellationToken;
+
+use crate::codec::{anthropic, openai_chat, openai_responses};
+
+use super::token::TokenProvider;
+
+pub const DEFAULT_BASE_URL: &str = "https://api.githubcopilot.com";
+const API_VERSION: &str = "2026-06-01";
+const ANTHROPIC_BETA: &str = "interleaved-thinking-2025-05-14";
+const USER_AGENT: &str = concat!("ai-harness/", env!("CARGO_PKG_VERSION"));
+
+/// Which wire codec a Copilot model speaks.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum CopilotCodec {
+    Chat,
+    Responses,
+    Anthropic,
+}
+
+impl CopilotCodec {
+    fn path(self) -> &'static str {
+        match self {
+            CopilotCodec::Chat => "/chat/completions",
+            CopilotCodec::Responses => "/responses",
+            CopilotCodec::Anthropic => "/v1/messages",
+        }
+    }
+
+    fn build_request(self, req: &LlmRequest) -> Value {
+        match self {
+            CopilotCodec::Chat => openai_chat::build_request(req),
+            CopilotCodec::Responses => openai_responses::build_request(req),
+            CopilotCodec::Anthropic => anthropic::build_request(req),
+        }
+    }
+
+    fn decode(
+        self,
+        stream: impl futures::Stream<Item = reqwest::Result<bytes::Bytes>> + Send + 'static,
+    ) -> LlmEventStream {
+        match self {
+            CopilotCodec::Chat => openai_chat::decode(stream),
+            CopilotCodec::Responses => openai_responses::decode(stream),
+            CopilotCodec::Anthropic => anthropic::decode(stream),
+        }
+    }
+}
+
+/// A single model as advertised by `GET /models`.
+#[derive(Debug, Clone)]
+pub struct CopilotModel {
+    pub id: String,
+    pub codec: CopilotCodec,
+    /// Only `model_picker_enabled` models are offered in the UI picker.
+    pub picker_enabled: bool,
+}
+
+/// Picks a codec from a model's `supported_endpoints`, preferring the responses API, then the
+/// Anthropic messages API, then chat completions.
+pub fn codec_for_endpoints(endpoints: &[String]) -> CopilotCodec {
+    let has = |needle: &str| endpoints.iter().any(|e| e.contains(needle));
+    if has("/responses") {
+        CopilotCodec::Responses
+    } else if has("/v1/messages") || has("/messages") {
+        CopilotCodec::Anthropic
+    } else {
+        CopilotCodec::Chat
+    }
+}
+
+/// Parses a Copilot `GET /models` response body.
+pub fn parse_models(body: &Value) -> Vec<CopilotModel> {
+    body["data"]
+        .as_array()
+        .map(|arr| {
+            arr.iter()
+                .filter_map(|m| {
+                    let id = m["id"].as_str()?.to_string();
+                    let endpoints: Vec<String> = m["supported_endpoints"]
+                        .as_array()
+                        .map(|a| {
+                            a.iter()
+                                .filter_map(|e| e.as_str().map(String::from))
+                                .collect()
+                        })
+                        .unwrap_or_default();
+                    Some(CopilotModel {
+                        id,
+                        codec: codec_for_endpoints(&endpoints),
+                        picker_enabled: m["model_picker_enabled"].as_bool().unwrap_or(false),
+                    })
+                })
+                .collect()
+        })
+        .unwrap_or_default()
+}
+
+fn has_images(req: &LlmRequest) -> bool {
+    req.messages
+        .iter()
+        .flat_map(|m| &m.content)
+        .any(|c| matches!(c, WireContent::Image { .. }))
+}
+
+fn initiator_header(initiator: Initiator) -> &'static str {
+    match initiator {
+        Initiator::User => "user",
+        Initiator::Agent => "agent",
+    }
+}
+
+pub struct CopilotProvider {
+    tokens: TokenProvider,
+    base_url: String,
+    client: reqwest::Client,
+    /// model id → codec, from `GET /models`. Unknown models default to chat completions.
+    codecs: HashMap<String, CopilotCodec>,
+}
+
+impl CopilotProvider {
+    pub fn new(oauth_token: impl Into<String>, models: Vec<CopilotModel>) -> Self {
+        Self::with_base_url(oauth_token, DEFAULT_BASE_URL, models)
+    }
+
+    pub fn with_base_url(
+        oauth_token: impl Into<String>,
+        base_url: impl Into<String>,
+        models: Vec<CopilotModel>,
+    ) -> Self {
+        let client = reqwest::Client::new();
+        let codecs = models.into_iter().map(|m| (m.id, m.codec)).collect();
+        Self {
+            tokens: TokenProvider::new(oauth_token, client.clone()),
+            base_url: base_url.into(),
+            client,
+            codecs,
+        }
+    }
+
+    fn codec_for(&self, model: &str) -> CopilotCodec {
+        self.codecs
+            .get(model)
+            .copied()
+            .unwrap_or(CopilotCodec::Chat)
+    }
+
+    fn classify_error(status: reqwest::StatusCode, body: String) -> ProviderError {
+        match status.as_u16() {
+            401 | 403 => ProviderError::Auth(body),
+            429 => ProviderError::RateLimited { retry_after: None },
+            s if (500..600).contains(&s) => ProviderError::Overloaded,
+            s => ProviderError::Http { status: s, body },
+        }
+    }
+}
+
+#[async_trait]
+impl Provider for CopilotProvider {
+    fn id(&self) -> &str {
+        "github-copilot"
+    }
+
+    async fn list_models(&self) -> Result<Vec<ModelInfo>, ProviderError> {
+        // Cost/limit metadata is layered on by models.dev; routing metadata lives in `codecs`.
+        Ok(Vec::new())
+    }
+
+    async fn stream(
+        &self,
+        req: LlmRequest,
+        cancel: CancellationToken,
+    ) -> Result<LlmEventStream, ProviderError> {
+        let codec = self.codec_for(&req.model);
+        let body = codec.build_request(&req);
+        let url = format!("{}{}", self.base_url, codec.path());
+
+        let now_secs = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .unwrap_or_default()
+            .as_secs() as i64;
+        let token = self.tokens.token(now_secs).await;
+
+        let mut builder = self
+            .client
+            .post(&url)
+            .header("authorization", format!("Bearer {}", token.token))
+            .header("user-agent", USER_AGENT)
+            .header("x-github-api-version", API_VERSION)
+            .header("openai-intent", "conversation-edits")
+            .header("x-initiator", initiator_header(req.initiator));
+        if has_images(&req) {
+            builder = builder.header("copilot-vision-request", "true");
+        }
+        if codec == CopilotCodec::Anthropic {
+            builder = builder.header("anthropic-beta", ANTHROPIC_BETA);
+        }
+
+        let send = builder.json(&body).send();
+        let response = tokio::select! {
+            result = send => result.map_err(|e| ProviderError::Network(e.to_string()))?,
+            _ = cancel.cancelled() => return Err(ProviderError::Cancelled),
+        };
+
+        if !response.status().is_success() {
+            let status = response.status();
+            let body = response.text().await.unwrap_or_default();
+            return Err(Self::classify_error(status, body));
+        }
+
+        Ok(codec.decode(response.bytes_stream()))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn codec_routing_prefers_responses_then_anthropic_then_chat() {
+        assert_eq!(
+            codec_for_endpoints(&["/chat/completions".into(), "/responses".into()]),
+            CopilotCodec::Responses
+        );
+        assert_eq!(
+            codec_for_endpoints(&["/v1/messages".into()]),
+            CopilotCodec::Anthropic
+        );
+        assert_eq!(
+            codec_for_endpoints(&["/chat/completions".into()]),
+            CopilotCodec::Chat
+        );
+        assert_eq!(codec_for_endpoints(&[]), CopilotCodec::Chat);
+    }
+
+    #[test]
+    fn parses_models_with_endpoints_and_picker_flag() {
+        let body = json!({
+            "data": [
+                {"id": "gpt-4o", "supported_endpoints": ["/chat/completions"], "model_picker_enabled": true},
+                {"id": "claude-sonnet-4-5", "supported_endpoints": ["/v1/messages"], "model_picker_enabled": true},
+                {"id": "o3", "supported_endpoints": ["/responses"], "model_picker_enabled": false},
+                {"id": "no-endpoints"}
+            ]
+        });
+        let models = parse_models(&body);
+        assert_eq!(models.len(), 4);
+        assert_eq!(models[0].codec, CopilotCodec::Chat);
+        assert!(models[0].picker_enabled);
+        assert_eq!(models[1].codec, CopilotCodec::Anthropic);
+        assert_eq!(models[2].codec, CopilotCodec::Responses);
+        assert!(!models[2].picker_enabled);
+        // Missing supported_endpoints → default chat, picker false.
+        assert_eq!(models[3].codec, CopilotCodec::Chat);
+        assert!(!models[3].picker_enabled);
+    }
+
+    #[test]
+    fn unknown_model_defaults_to_chat_codec() {
+        let provider = CopilotProvider::new("oauth", vec![]);
+        assert_eq!(provider.codec_for("whatever"), CopilotCodec::Chat);
+        assert_eq!(provider.id(), "github-copilot");
+    }
+
+    #[test]
+    fn known_model_uses_mapped_codec() {
+        let provider = CopilotProvider::new(
+            "oauth",
+            vec![CopilotModel {
+                id: "claude-sonnet-4-5".into(),
+                codec: CopilotCodec::Anthropic,
+                picker_enabled: true,
+            }],
+        );
+        assert_eq!(
+            provider.codec_for("claude-sonnet-4-5"),
+            CopilotCodec::Anthropic
+        );
+    }
+}

crates/harness-providers/src/copilot/token.rs

@@ -0,0 +1,144 @@
+//! Copilot API token strategy (flagged-risk area — verify against the live API).
+//!
+//! opencode sends the GitHub OAuth token directly as the Bearer (`expires: 0`). The classic
+//! Copilot API instead wants a short-lived token from `copilot_internal/v2/token`. We try the
+//! exchange first and cache it until shortly before expiry; if the endpoint rejects us, we fall
+//! back to the direct-Bearer behavior. Refresh is single-flight under a `tokio::Mutex`.
+
+use serde::Deserialize;
+use tokio::sync::Mutex;
+
+const EXCHANGE_URL: &str = "https://api.github.com/copilot_internal/v2/token";
+/// Refresh this many seconds before the reported expiry.
+const EXPIRY_SKEW_SECS: i64 = 120;
+const USER_AGENT: &str = concat!("ai-harness/", env!("CARGO_PKG_VERSION"));
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CopilotToken {
+    pub token: String,
+    /// Unix seconds; `0` means "never expires" (direct-Bearer fallback).
+    pub expires_at: i64,
+}
+
+impl CopilotToken {
+    /// True when the token should be refreshed at `now_secs` (never, for `expires_at == 0`).
+    pub fn needs_refresh(&self, now_secs: i64) -> bool {
+        self.expires_at != 0 && now_secs + EXPIRY_SKEW_SECS >= self.expires_at
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum TokenError {
+    #[error("network: {0}")]
+    Network(String),
+    /// The exchange endpoint is unavailable/unauthorized — the caller should fall back to the
+    /// direct-Bearer strategy.
+    #[error("exchange unsupported (status {0})")]
+    ExchangeUnsupported(u16),
+    #[error("unexpected response: {0}")]
+    Unexpected(String),
+}
+
+#[derive(Deserialize)]
+struct ExchangeResponse {
+    token: String,
+    #[serde(default)]
+    expires_at: i64,
+}
+
+/// Performs the token exchange once. `ExchangeUnsupported` signals the caller to fall back.
+pub async fn exchange(
+    client: &reqwest::Client,
+    oauth_token: &str,
+) -> Result<CopilotToken, TokenError> {
+    let resp = client
+        .get(EXCHANGE_URL)
+        .header("authorization", format!("token {oauth_token}"))
+        .header("accept", "application/json")
+        .header("user-agent", USER_AGENT)
+        .send()
+        .await
+        .map_err(|e| TokenError::Network(e.to_string()))?;
+
+    let status = resp.status();
+    if !status.is_success() {
+        // 401/403/404 → this deployment doesn't support the exchange; fall back to direct Bearer.
+        if matches!(status.as_u16(), 401 | 403 | 404) {
+            return Err(TokenError::ExchangeUnsupported(status.as_u16()));
+        }
+        return Err(TokenError::Network(format!("status {}", status.as_u16())));
+    }
+
+    let parsed: ExchangeResponse = resp
+        .json()
+        .await
+        .map_err(|e| TokenError::Unexpected(e.to_string()))?;
+    Ok(CopilotToken {
+        token: parsed.token,
+        expires_at: parsed.expires_at,
+    })
+}
+
+/// Caches the exchanged Copilot token and refreshes it single-flight. Falls back to using the
+/// OAuth token directly (never-expiring) when the exchange endpoint rejects the request.
+pub struct TokenProvider {
+    oauth_token: String,
+    client: reqwest::Client,
+    cached: Mutex<Option<CopilotToken>>,
+}
+
+impl TokenProvider {
+    pub fn new(oauth_token: impl Into<String>, client: reqwest::Client) -> Self {
+        Self {
+            oauth_token: oauth_token.into(),
+            client,
+            cached: Mutex::new(None),
+        }
+    }
+
+    /// Returns a usable Bearer token, refreshing/exchanging as needed. `now_secs` is injected
+    /// for testability.
+    pub async fn token(&self, now_secs: i64) -> CopilotToken {
+        let mut guard = self.cached.lock().await;
+        if let Some(tok) = guard.as_ref() {
+            if !tok.needs_refresh(now_secs) {
+                return tok.clone();
+            }
+        }
+        let fresh = match exchange(&self.client, &self.oauth_token).await {
+            Ok(tok) => tok,
+            Err(_) => CopilotToken {
+                // Direct-Bearer fallback: use the OAuth token itself, never expiring.
+                token: self.oauth_token.clone(),
+                expires_at: 0,
+            },
+        };
+        *guard = Some(fresh.clone());
+        fresh
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn never_expiring_token_does_not_refresh() {
+        let tok = CopilotToken {
+            token: "t".into(),
+            expires_at: 0,
+        };
+        assert!(!tok.needs_refresh(i64::MAX));
+    }
+
+    #[test]
+    fn refresh_triggers_within_skew_window() {
+        let tok = CopilotToken {
+            token: "t".into(),
+            expires_at: 1_000,
+        };
+        assert!(!tok.needs_refresh(800)); // 800 + 120 < 1000
+        assert!(tok.needs_refresh(881)); // 881 + 120 >= 1000
+        assert!(tok.needs_refresh(1_000));
+    }
+}

crates/harness-providers/src/lib.rs

@@ -1 +1,14 @@
-// Provider implementations (Copilot, OpenAI, Anthropic) land here in M1/M3.
+pub mod anthropic;
+pub mod auth;
+pub mod codec;
+pub mod copilot;
+pub mod modelsdev;
+pub mod openai;
+pub mod registry;
+
+pub use anthropic::AnthropicProvider;
+pub use auth::{AuthRecord, AuthStorage};
+pub use copilot::CopilotProvider;
+pub use modelsdev::ModelCatalog;
+pub use openai::OpenAiProvider;
+pub use registry::ProviderRegistry;

crates/harness-providers/src/modelsdev.rs

@@ -0,0 +1,296 @@
+//! models.dev metadata: per-model context/output limits, pricing, and capability flags.
+//!
+//! At runtime we prefer a cached copy of `https://models.dev/api.json` (refreshed every 24h);
+//! if the cache is missing/stale and the network is unavailable, we fall back to a baked
+//! snapshot so cost display and limits still work offline.
+
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::time::{Duration, SystemTime};
+
+use harness_core::types::{ModelCost, ModelInfo, ModelRef};
+use serde::Deserialize;
+
+const API_URL: &str = "https://models.dev/api.json";
+const CACHE_TTL: Duration = Duration::from_secs(24 * 60 * 60);
+/// Baked fallback, refreshed manually. Keeps cost/limits working with no cache and no network.
+const SNAPSHOT: &str = include_str!("../assets/models-snapshot.json");
+
+/// models.dev top-level shape: `{ provider_id: { models: { model_id: {...} } } }`. Unknown
+/// keys (provider metadata, per-model extras) are ignored.
+#[derive(Deserialize)]
+struct ApiProvider {
+    #[serde(default)]
+    models: HashMap<String, ApiModel>,
+}
+
+#[derive(Deserialize)]
+struct ApiModel {
+    #[serde(default)]
+    reasoning: bool,
+    #[serde(default)]
+    tool_call: bool,
+    #[serde(default)]
+    attachment: bool,
+    #[serde(default)]
+    cost: ApiCost,
+    #[serde(default)]
+    limit: ApiLimit,
+}
+
+#[derive(Deserialize, Default)]
+struct ApiCost {
+    #[serde(default)]
+    input: f64,
+    #[serde(default)]
+    output: f64,
+    #[serde(default)]
+    cache_read: f64,
+    #[serde(default)]
+    cache_write: f64,
+}
+
+#[derive(Deserialize, Default)]
+struct ApiLimit {
+    #[serde(default)]
+    context: u64,
+    #[serde(default)]
+    output: u64,
+}
+
+/// Parsed metadata keyed by `(provider_id, model_id)`.
+#[derive(Debug, Clone, Default)]
+pub struct ModelCatalog {
+    models: HashMap<(String, String), ModelInfo>,
+}
+
+impl ModelCatalog {
+    /// Parses a models.dev `api.json` document.
+    pub fn from_api_json(bytes: &[u8]) -> Result<Self, serde_json::Error> {
+        let raw: HashMap<String, ApiProvider> = serde_json::from_slice(bytes)?;
+        let mut models = HashMap::new();
+        for (provider_id, provider) in raw {
+            for (model_id, m) in provider.models {
+                let info = ModelInfo {
+                    model: ModelRef::new(provider_id.clone(), model_id.clone()),
+                    context_limit: m.limit.context,
+                    output_limit: m.limit.output,
+                    cost: ModelCost {
+                        input: m.cost.input,
+                        output: m.cost.output,
+                        cache_read: m.cost.cache_read,
+                        cache_write: m.cost.cache_write,
+                    },
+                    reasoning: m.reasoning,
+                    tool_call: m.tool_call,
+                    attachment: m.attachment,
+                };
+                models.insert((provider_id.clone(), model_id), info);
+            }
+        }
+        Ok(Self { models })
+    }
+
+    /// The baked fallback snapshot — always available, never fails.
+    pub fn baked() -> Self {
+        Self::from_api_json(SNAPSHOT.as_bytes()).expect("baked models snapshot must parse")
+    }
+
+    pub fn get(&self, provider: &str, model: &str) -> Option<&ModelInfo> {
+        self.models.get(&(provider.to_string(), model.to_string()))
+    }
+
+    /// Pricing for a model, or zero-cost if unknown.
+    pub fn cost(&self, provider: &str, model: &str) -> ModelCost {
+        self.get(provider, model)
+            .map(|m| m.cost)
+            .unwrap_or_default()
+    }
+
+    pub fn len(&self) -> usize {
+        self.models.len()
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.models.is_empty()
+    }
+
+    /// Default cache location: `~/.cache/ai-harness/models.json`.
+    pub fn default_cache_path() -> PathBuf {
+        dirs::cache_dir()
+            .unwrap_or_else(std::env::temp_dir)
+            .join("ai-harness")
+            .join("models.json")
+    }
+
+    /// Loads from the cache if present and younger than [`CACHE_TTL`]; otherwise returns the
+    /// baked snapshot. Never touches the network — call [`ModelCatalog::refresh`] for that.
+    pub fn load_cached_or_baked(cache_path: &Path) -> Self {
+        if cache_fresh(cache_path, CACHE_TTL, SystemTime::now()) {
+            if let Ok(bytes) = std::fs::read(cache_path) {
+                if let Ok(catalog) = Self::from_api_json(&bytes) {
+                    if !catalog.is_empty() {
+                        return catalog;
+                    }
+                }
+            }
+        }
+        Self::baked()
+    }
+
+    /// Refreshes the default cache location with a fresh HTTP client. Convenience wrapper for
+    /// callers (the app) that don't want to depend on `reqwest` directly.
+    pub async fn refresh_default_cache() -> Result<Self, RefreshError> {
+        let client = reqwest::Client::new();
+        Self::refresh(&client, &Self::default_cache_path()).await
+    }
+
+    /// Fetches the latest metadata and writes it to `cache_path` (best-effort). Returns the
+    /// freshly-parsed catalog. Intended to run in the background so the *next* launch is current.
+    pub async fn refresh(
+        client: &reqwest::Client,
+        cache_path: &Path,
+    ) -> Result<Self, RefreshError> {
+        let bytes = client
+            .get(API_URL)
+            .send()
+            .await
+            .map_err(|e| RefreshError::Network(e.to_string()))?
+            .error_for_status()
+            .map_err(|e| RefreshError::Network(e.to_string()))?
+            .bytes()
+            .await
+            .map_err(|e| RefreshError::Network(e.to_string()))?;
+
+        let catalog =
+            Self::from_api_json(&bytes).map_err(|e| RefreshError::Parse(e.to_string()))?;
+
+        if let Some(parent) = cache_path.parent() {
+            let _ = std::fs::create_dir_all(parent);
+        }
+        if let Err(e) = std::fs::write(cache_path, &bytes) {
+            tracing::warn!(error = %e, "failed to cache models.dev metadata");
+        }
+        Ok(catalog)
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum RefreshError {
+    #[error("network: {0}")]
+    Network(String),
+    #[error("parse: {0}")]
+    Parse(String),
+}
+
+/// True when `path` exists and was modified within `ttl` of `now`.
+fn cache_fresh(path: &Path, ttl: Duration, now: SystemTime) -> bool {
+    let Ok(modified) = std::fs::metadata(path).and_then(|m| m.modified()) else {
+        return false;
+    };
+    match now.duration_since(modified) {
+        Ok(age) => age < ttl,
+        Err(_) => true, // modified in the future (clock skew) — treat as fresh
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const FIXTURE: &str = r#"{
+        "anthropic": {
+            "id": "anthropic",
+            "name": "Anthropic",
+            "models": {
+                "claude-sonnet-4-5": {
+                    "id": "claude-sonnet-4-5",
+                    "reasoning": true,
+                    "tool_call": true,
+                    "attachment": true,
+                    "cost": {"input": 3, "output": 15, "cache_read": 0.3, "cache_write": 3.75},
+                    "limit": {"context": 200000, "output": 64000}
+                }
+            }
+        },
+        "openai": {
+            "id": "openai",
+            "models": {
+                "gpt-4o": {
+                    "id": "gpt-4o",
+                    "tool_call": true,
+                    "cost": {"input": 2.5, "output": 10},
+                    "limit": {"context": 128000, "output": 16384}
+                }
+            }
+        }
+    }"#;
+
+    #[test]
+    fn parses_providers_and_models() {
+        let catalog = ModelCatalog::from_api_json(FIXTURE.as_bytes()).unwrap();
+        assert_eq!(catalog.len(), 2);
+
+        let sonnet = catalog.get("anthropic", "claude-sonnet-4-5").unwrap();
+        assert_eq!(sonnet.context_limit, 200_000);
+        assert_eq!(sonnet.output_limit, 64_000);
+        assert_eq!(sonnet.cost.input, 3.0);
+        assert_eq!(sonnet.cost.cache_write, 3.75);
+        assert!(sonnet.reasoning && sonnet.tool_call && sonnet.attachment);
+    }
+
+    #[test]
+    fn missing_cost_fields_default_to_zero() {
+        let catalog = ModelCatalog::from_api_json(FIXTURE.as_bytes()).unwrap();
+        let gpt = catalog.get("openai", "gpt-4o").unwrap();
+        assert_eq!(gpt.cost.cache_read, 0.0);
+        assert_eq!(gpt.cost.cache_write, 0.0);
+        assert!(!gpt.reasoning); // absent → false
+        assert!(gpt.tool_call);
+    }
+
+    #[test]
+    fn cost_helper_is_zero_for_unknown_model() {
+        let catalog = ModelCatalog::from_api_json(FIXTURE.as_bytes()).unwrap();
+        assert_eq!(catalog.cost("openai", "nonexistent"), ModelCost::default());
+        assert_eq!(catalog.cost("anthropic", "claude-sonnet-4-5").input, 3.0);
+    }
+
+    #[test]
+    fn baked_snapshot_parses_and_is_nonempty() {
+        let catalog = ModelCatalog::baked();
+        assert!(!catalog.is_empty());
+    }
+
+    #[test]
+    fn cache_freshness_respects_ttl() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("models.json");
+        assert!(!cache_fresh(&path, CACHE_TTL, SystemTime::now())); // missing
+
+        std::fs::write(&path, "{}").unwrap();
+        let now = SystemTime::now();
+        assert!(cache_fresh(&path, CACHE_TTL, now));
+        // A "now" far in the future makes the file look stale.
+        let future = now + Duration::from_secs(48 * 60 * 60);
+        assert!(!cache_fresh(&path, CACHE_TTL, future));
+    }
+
+    #[test]
+    fn load_cached_or_baked_falls_back_when_cache_absent() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("models.json");
+        let catalog = ModelCatalog::load_cached_or_baked(&path);
+        assert!(!catalog.is_empty()); // baked fallback
+    }
+
+    #[test]
+    fn load_cached_or_baked_reads_fresh_cache() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("models.json");
+        std::fs::write(&path, FIXTURE).unwrap();
+        let catalog = ModelCatalog::load_cached_or_baked(&path);
+        assert_eq!(catalog.len(), 2);
+        assert!(catalog.get("anthropic", "claude-sonnet-4-5").is_some());
+    }
+}

crates/harness-providers/src/openai.rs

@@ -0,0 +1,173 @@
+use std::time::Duration;
+
+use async_trait::async_trait;
+use harness_core::llm::{LlmEventStream, LlmRequest, Provider, ProviderError};
+use harness_core::types::ModelInfo;
+use tokio_util::sync::CancellationToken;
+
+use crate::codec::{openai_chat, openai_responses};
+
+const DEFAULT_BASE_URL: &str = "https://api.openai.com/v1";
+
+/// Which OpenAI wire format to speak for a given model.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum ApiFlavor {
+    Responses,
+    Chat,
+}
+
+/// `gpt-*` and `o-*` models use the responses API; everything else (including
+/// OpenAI-compatible third-party endpoints) falls back to chat completions.
+fn flavor_for(model: &str) -> ApiFlavor {
+    let is_native = model.starts_with("gpt-")
+        || model.starts_with("o1")
+        || model.starts_with("o3")
+        || model.starts_with("o4")
+        || model == "o1"
+        || model == "o3";
+    if is_native {
+        ApiFlavor::Responses
+    } else {
+        ApiFlavor::Chat
+    }
+}
+
+pub struct OpenAiProvider {
+    api_key: String,
+    base_url: String,
+    client: reqwest::Client,
+}
+
+impl OpenAiProvider {
+    pub fn new(api_key: impl Into<String>) -> Self {
+        Self {
+            api_key: api_key.into(),
+            base_url: DEFAULT_BASE_URL.to_string(),
+            client: reqwest::Client::new(),
+        }
+    }
+
+    pub fn with_base_url(api_key: impl Into<String>, base_url: impl Into<String>) -> Self {
+        Self {
+            api_key: api_key.into(),
+            base_url: base_url.into(),
+            client: reqwest::Client::new(),
+        }
+    }
+
+    fn classify_error(
+        status: reqwest::StatusCode,
+        body: String,
+        retry_after: Option<Duration>,
+    ) -> ProviderError {
+        match status.as_u16() {
+            400 if body.contains("context_length_exceeded")
+                || body.contains("maximum context length") =>
+            {
+                ProviderError::ContextOverflow
+            }
+            401 | 403 => ProviderError::Auth(body),
+            429 => ProviderError::RateLimited { retry_after },
+            s if (500..600).contains(&s) => ProviderError::Overloaded,
+            s => ProviderError::Http { status: s, body },
+        }
+    }
+}
+
+#[async_trait]
+impl Provider for OpenAiProvider {
+    fn id(&self) -> &str {
+        "openai"
+    }
+
+    async fn list_models(&self) -> Result<Vec<ModelInfo>, ProviderError> {
+        // models.dev metadata is layered on top by the caller (see `modelsdev.rs`).
+        Ok(Vec::new())
+    }
+
+    async fn stream(
+        &self,
+        req: LlmRequest,
+        cancel: CancellationToken,
+    ) -> Result<LlmEventStream, ProviderError> {
+        let (path, body) = match flavor_for(&req.model) {
+            ApiFlavor::Responses => ("/responses", openai_responses::build_request(&req)),
+            ApiFlavor::Chat => ("/chat/completions", openai_chat::build_request(&req)),
+        };
+        let url = format!("{}{}", self.base_url, path);
+
+        let send = self
+            .client
+            .post(&url)
+            .header("authorization", format!("Bearer {}", self.api_key))
+            .json(&body)
+            .send();
+
+        let response = tokio::select! {
+            result = send => result.map_err(|e| ProviderError::Network(e.to_string()))?,
+            _ = cancel.cancelled() => return Err(ProviderError::Cancelled),
+        };
+
+        if !response.status().is_success() {
+            let status = response.status();
+            let retry_after = response
+                .headers()
+                .get("retry-after")
+                .and_then(|v| v.to_str().ok())
+                .and_then(|s| s.parse::<u64>().ok())
+                .map(Duration::from_secs);
+            let body = response.text().await.unwrap_or_default();
+            return Err(Self::classify_error(status, body, retry_after));
+        }
+
+        Ok(match flavor_for(&req.model) {
+            ApiFlavor::Responses => openai_responses::decode(response.bytes_stream()),
+            ApiFlavor::Chat => openai_chat::decode(response.bytes_stream()),
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn routes_gpt_and_o_series_to_responses() {
+        assert_eq!(flavor_for("gpt-4o"), ApiFlavor::Responses);
+        assert_eq!(flavor_for("gpt-5"), ApiFlavor::Responses);
+        assert_eq!(flavor_for("o1"), ApiFlavor::Responses);
+        assert_eq!(flavor_for("o3-mini"), ApiFlavor::Responses);
+        assert_eq!(flavor_for("o4-mini"), ApiFlavor::Responses);
+    }
+
+    #[test]
+    fn routes_other_models_to_chat() {
+        assert_eq!(flavor_for("llama-3.1-70b"), ApiFlavor::Chat);
+        assert_eq!(flavor_for("deepseek-chat"), ApiFlavor::Chat);
+    }
+
+    #[test]
+    fn id_is_openai() {
+        assert_eq!(OpenAiProvider::new("k").id(), "openai");
+    }
+
+    #[test]
+    fn classifies_context_overflow_from_400_body() {
+        assert!(matches!(
+            OpenAiProvider::classify_error(
+                reqwest::StatusCode::BAD_REQUEST,
+                "context_length_exceeded".into(),
+                None,
+            ),
+            ProviderError::ContextOverflow
+        ));
+        assert!(matches!(
+            OpenAiProvider::classify_error(
+                reqwest::StatusCode::BAD_REQUEST,
+                "some other error".into(),
+                None,
+            ),
+            ProviderError::Http { status: 400, .. }
+        ));
+    }
+}

crates/harness-providers/src/registry.rs

@@ -0,0 +1,57 @@
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use harness_core::llm::Provider;
+
+#[derive(Default, Clone)]
+pub struct ProviderRegistry {
+    providers: HashMap<String, Arc<dyn Provider>>,
+}
+
+impl ProviderRegistry {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn register(&mut self, provider: Arc<dyn Provider>) {
+        self.providers.insert(provider.id().to_string(), provider);
+    }
+
+    pub fn get(&self, id: &str) -> Option<Arc<dyn Provider>> {
+        self.providers.get(id).cloned()
+    }
+
+    /// Resolves a `"provider/model"` string to its provider and bare model id.
+    pub fn resolve(&self, model_ref: &str) -> Option<(Arc<dyn Provider>, String)> {
+        let (provider_id, model_id) = model_ref.split_once('/')?;
+        self.get(provider_id).map(|p| (p, model_id.to_string()))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::anthropic::AnthropicProvider;
+
+    #[test]
+    fn resolves_provider_slash_model() {
+        let mut registry = ProviderRegistry::new();
+        registry.register(Arc::new(AnthropicProvider::new("key")));
+        let (provider, model) = registry.resolve("anthropic/claude-sonnet-4-5").unwrap();
+        assert_eq!(provider.id(), "anthropic");
+        assert_eq!(model, "claude-sonnet-4-5");
+    }
+
+    #[test]
+    fn unknown_provider_resolves_to_none() {
+        let registry = ProviderRegistry::new();
+        assert!(registry.resolve("openai/gpt-4o").is_none());
+    }
+
+    #[test]
+    fn malformed_model_ref_resolves_to_none() {
+        let mut registry = ProviderRegistry::new();
+        registry.register(Arc::new(AnthropicProvider::new("key")));
+        assert!(registry.resolve("no-slash-here").is_none());
+    }
+}

crates/harness-tools/Cargo.toml

@@ -6,6 +6,27 @@ license.workspace = true
 
 [dependencies]
 harness-core = { workspace = true }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+async-trait = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+schemars = { workspace = true }
+ignore = { workspace = true }
+globset = { workspace = true }
+grep-searcher = { workspace = true }
+grep-regex = { workspace = true }
+grep-matcher = { workspace = true }
+similar = { workspace = true }
+shell-words = { workspace = true }
+thiserror = { workspace = true }
+
+[target.'cfg(unix)'.dependencies]
+libc = { workspace = true }
+
+[dev-dependencies]
+tempfile = { workspace = true }
+tokio = { workspace = true, features = ["test-util", "macros"] }
 
 [lints]
 workspace = true

crates/harness-tools/src/bash.rs

@@ -0,0 +1,215 @@
+use std::process::Stdio;
+use std::time::Duration;
+
+use async_trait::async_trait;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+const DEFAULT_TIMEOUT_MS: u64 = 2 * 60 * 1000;
+const MAX_TIMEOUT_MS: u64 = 10 * 60 * 1000;
+
+#[derive(Debug, Deserialize, JsonSchema)]
+struct BashParams {
+    command: String,
+    timeout_ms: Option<u64>,
+    cwd: Option<String>,
+    #[allow(dead_code)]
+    description: String,
+}
+
+pub struct BashTool;
+
+/// The "always allow" pattern is coarser than the exact command: opencode/docs/05-tools.md
+/// grants `<first word> *` (e.g. `git *`), not the literal command string.
+fn always_pattern(command: &str) -> String {
+    shell_words::split(command)
+        .ok()
+        .and_then(|words| words.into_iter().next())
+        .map(|first| format!("{first} *"))
+        .unwrap_or_else(|| "*".to_string())
+}
+
+#[cfg(unix)]
+#[allow(unsafe_code)]
+fn kill_group(pid: u32) {
+    // Negative pid signals the whole process group (spawned with `process_group(0)`).
+    // Safety: `kill` is a plain libc syscall; passing a negative pid targets the group,
+    // which is exactly the process tree we spawned and want to tear down.
+    unsafe {
+        libc::kill(-(pid as i32), libc::SIGKILL);
+    }
+}
+
+#[cfg(not(unix))]
+fn kill_group(_pid: u32) {}
+
+#[async_trait]
+impl Tool for BashTool {
+    fn name(&self) -> &str {
+        "bash"
+    }
+
+    fn description(&self) -> &str {
+        "Runs a shell command and returns its combined stdout/stderr."
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        serde_json::to_value(schemars::schema_for!(BashParams)).unwrap()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let params: BashParams =
+            serde_json::from_value(input).map_err(|e| ToolError::Invalid(e.to_string()))?;
+        let timeout_ms = params
+            .timeout_ms
+            .unwrap_or(DEFAULT_TIMEOUT_MS)
+            .min(MAX_TIMEOUT_MS);
+
+        ctx.ask
+            .ask(
+                "bash",
+                params.command.clone(),
+                always_pattern(&params.command),
+                serde_json::json!({"command": params.command}),
+            )
+            .await?;
+
+        let cwd = params
+            .cwd
+            .as_ref()
+            .map(|c| crate::paths::resolve(&ctx.cwd, c))
+            .unwrap_or_else(|| ctx.cwd.clone());
+
+        let mut cmd = tokio::process::Command::new("sh");
+        cmd.arg("-c")
+            .arg(&params.command)
+            .current_dir(&cwd)
+            .stdout(Stdio::piped())
+            .stderr(Stdio::piped());
+        #[cfg(unix)]
+        cmd.process_group(0);
+
+        let child = cmd
+            .spawn()
+            .map_err(|e| ToolError::Other(format!("spawn failed: {e}")))?;
+        let pid = child.id();
+
+        tokio::select! {
+            result = child.wait_with_output() => {
+                let output = result.map_err(|e| ToolError::Other(format!("wait failed: {e}")))?;
+                let mut combined = String::from_utf8_lossy(&output.stdout).into_owned();
+                combined.push_str(&String::from_utf8_lossy(&output.stderr));
+                let title = format!("{} (exit {})", params.command, output.status.code().unwrap_or(-1));
+                Ok(ToolOutput::new(title, combined))
+            }
+            _ = tokio::time::sleep(Duration::from_millis(timeout_ms)) => {
+                if let Some(pid) = pid { kill_group(pid); }
+                Err(ToolError::Other(format!("command timed out after {timeout_ms}ms")))
+            }
+            _ = ctx.cancel.cancelled() => {
+                if let Some(pid) = pid { kill_group(pid); }
+                Err(ToolError::Cancelled)
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use harness_core::event::EventBus;
+    use harness_core::permission::{spawn_auto_approve, PermissionService};
+    use harness_core::tool::{MetadataSink, PermissionHandle};
+    use harness_core::types::SessionId;
+    use std::sync::{Arc, Mutex};
+    use tokio_util::sync::CancellationToken;
+
+    fn ctx(cwd: std::path::PathBuf) -> ToolCtx {
+        let bus = EventBus::new();
+        let service = Arc::new(PermissionService::new(bus.clone()));
+        spawn_auto_approve(bus, service.clone());
+        let (metadata, _rx) = MetadataSink::channel();
+        ToolCtx {
+            session_id: SessionId::new(),
+            message_id: harness_core::types::MessageId::new(),
+            call_id: "call_1".into(),
+            data_dir: cwd.join("tool-output"),
+            cwd,
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                service,
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+        }
+    }
+
+    #[test]
+    fn always_pattern_uses_first_word() {
+        assert_eq!(always_pattern("git push origin main"), "git *");
+        assert_eq!(always_pattern("ls -la"), "ls *");
+    }
+
+    #[tokio::test]
+    async fn runs_command_and_captures_stdout() {
+        let dir = tempfile::tempdir().unwrap();
+        let output = BashTool
+            .execute(
+                serde_json::json!({"command": "echo hello", "description": "say hi"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output.output.contains("hello"));
+    }
+
+    #[tokio::test]
+    async fn captures_stderr_too() {
+        let dir = tempfile::tempdir().unwrap();
+        let output = BashTool
+            .execute(
+                serde_json::json!({"command": "echo err 1>&2", "description": "stderr"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output.output.contains("err"));
+    }
+
+    #[tokio::test]
+    async fn times_out_long_running_commands() {
+        let dir = tempfile::tempdir().unwrap();
+        let err = BashTool
+            .execute(
+                serde_json::json!({"command": "sleep 5", "timeout_ms": 50, "description": "slow"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap_err();
+        assert!(matches!(err, ToolError::Other(_)));
+    }
+
+    #[tokio::test]
+    async fn respects_cwd_override() {
+        let dir = tempfile::tempdir().unwrap();
+        let output = BashTool
+            .execute(
+                serde_json::json!({"command": "pwd", "description": "where"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output
+            .output
+            .trim()
+            .ends_with(dir.path().file_name().unwrap().to_str().unwrap()));
+    }
+}

crates/harness-tools/src/edit/mod.rs

@@ -0,0 +1,467 @@
+mod replacers;
+
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::sync::{Arc, Mutex as StdMutex};
+
+use async_trait::async_trait;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use schemars::JsonSchema;
+use serde::Deserialize;
+use similar::{ChangeTag, TextDiff};
+use tokio::sync::Mutex as AsyncMutex;
+
+use crate::paths;
+
+#[derive(Debug, Deserialize, JsonSchema)]
+struct EditParams {
+    file_path: String,
+    old_string: String,
+    new_string: String,
+    replace_all: Option<bool>,
+}
+
+/// Per-file `tokio::Mutex` lock map so concurrent edits to different files proceed in
+/// parallel while edits to the *same* file serialize (docs/05-tools.md).
+pub struct EditTool {
+    locks: StdMutex<HashMap<PathBuf, Arc<AsyncMutex<()>>>>,
+}
+
+impl EditTool {
+    pub fn new() -> Self {
+        Self {
+            locks: StdMutex::new(HashMap::new()),
+        }
+    }
+
+    fn lock_for(&self, path: &Path) -> Arc<AsyncMutex<()>> {
+        self.locks
+            .lock()
+            .unwrap()
+            .entry(path.to_path_buf())
+            .or_insert_with(|| Arc::new(AsyncMutex::new(())))
+            .clone()
+    }
+}
+
+impl Default for EditTool {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+fn normalize_line_endings(text: &str) -> String {
+    text.replace("\r\n", "\n")
+}
+
+fn detect_line_ending(text: &str) -> &'static str {
+    if text.contains("\r\n") {
+        "\r\n"
+    } else {
+        "\n"
+    }
+}
+
+fn convert_line_ending(text: &str, ending: &str) -> String {
+    if ending == "\n" {
+        text.to_string()
+    } else {
+        text.replace('\n', "\r\n")
+    }
+}
+
+const UTF8_BOM: [u8; 3] = [0xEF, 0xBB, 0xBF];
+
+fn strip_bom(bytes: &[u8]) -> (bool, &[u8]) {
+    if bytes.starts_with(&UTF8_BOM) {
+        (true, &bytes[3..])
+    } else {
+        (false, bytes)
+    }
+}
+
+fn unified_diff(before: &str, after: &str, file_path: &str) -> String {
+    TextDiff::from_lines(before, after)
+        .unified_diff()
+        .header(&format!("a/{file_path}"), &format!("b/{file_path}"))
+        .to_string()
+}
+
+fn diff_stats(before: &str, after: &str) -> (usize, usize) {
+    let diff = TextDiff::from_lines(before, after);
+    let mut added = 0;
+    let mut removed = 0;
+    for change in diff.iter_all_changes() {
+        match change.tag() {
+            ChangeTag::Insert => added += 1,
+            ChangeTag::Delete => removed += 1,
+            ChangeTag::Equal => {}
+        }
+    }
+    (added, removed)
+}
+
+#[async_trait]
+impl Tool for EditTool {
+    fn name(&self) -> &str {
+        "edit"
+    }
+
+    fn description(&self) -> &str {
+        "Replaces an exact span of text in a file (fuzzy-matched against whitespace/indentation drift)."
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        serde_json::to_value(schemars::schema_for!(EditParams)).unwrap()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let params: EditParams =
+            serde_json::from_value(input).map_err(|e| ToolError::Invalid(e.to_string()))?;
+
+        if params.old_string == params.new_string {
+            return Err(ToolError::Invalid(
+                "No changes to apply: oldString and newString are identical.".into(),
+            ));
+        }
+
+        let path = paths::resolve(&ctx.cwd, &params.file_path);
+        let pattern = paths::relative_pattern(&ctx.cwd, &path);
+        let lock = self.lock_for(&path);
+        let _guard = lock.lock().await;
+
+        let metadata = tokio::fs::metadata(&path).await.ok();
+        if metadata.as_ref().is_some_and(|m| m.is_dir()) {
+            return Err(ToolError::Invalid(format!(
+                "Path is a directory, not a file: {}",
+                path.display()
+            )));
+        }
+
+        let (content_old, content_new, had_bom) = if params.old_string.is_empty() {
+            if metadata.is_some() {
+                return Err(ToolError::Invalid(
+                    "oldString cannot be empty when editing an existing file. Provide the exact text to replace, or use write for an intentional full-file replacement.".into(),
+                ));
+            }
+            (String::new(), params.new_string.clone(), false)
+        } else {
+            if metadata.is_none() {
+                return Err(ToolError::Invalid(format!(
+                    "File {} not found",
+                    path.display()
+                )));
+            }
+            let bytes = tokio::fs::read(&path)
+                .await
+                .map_err(|e| ToolError::Other(format!("{}: {e}", path.display())))?;
+            let (had_bom, stripped) = strip_bom(&bytes);
+            let text = String::from_utf8_lossy(stripped).into_owned();
+
+            let ending = detect_line_ending(&text);
+            let old = convert_line_ending(&normalize_line_endings(&params.old_string), ending);
+            let new_ = convert_line_ending(&normalize_line_endings(&params.new_string), ending);
+            let replaced =
+                replacers::replace(&text, &old, &new_, params.replace_all.unwrap_or(false))
+                    .map_err(|e| ToolError::Invalid(e.to_string()))?;
+            (text, replaced, had_bom)
+        };
+
+        let diff = unified_diff(&content_old, &content_new, &params.file_path);
+        ctx.ask
+            .ask(
+                "edit",
+                pattern.clone(),
+                "*",
+                serde_json::json!({"file_path": params.file_path, "diff": diff}),
+            )
+            .await?;
+
+        if let Some(parent) = path.parent() {
+            tokio::fs::create_dir_all(parent)
+                .await
+                .map_err(|e| ToolError::Other(format!("{}: {e}", parent.display())))?;
+        }
+        let mut out_bytes = Vec::with_capacity(content_new.len() + 3);
+        if had_bom {
+            out_bytes.extend_from_slice(&UTF8_BOM);
+        }
+        out_bytes.extend_from_slice(content_new.as_bytes());
+        tokio::fs::write(&path, &out_bytes)
+            .await
+            .map_err(|e| ToolError::Other(format!("{}: {e}", path.display())))?;
+
+        let (added, removed) = diff_stats(&content_old, &content_new);
+        let mut output = ToolOutput::new(pattern, "Edit applied successfully.".to_string());
+        output.metadata = serde_json::json!({"diff": diff, "added": added, "removed": removed});
+        Ok(output)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use harness_core::event::EventBus;
+    use harness_core::permission::{spawn_auto_approve, PermissionService};
+    use harness_core::tool::{MetadataSink, PermissionHandle};
+    use harness_core::types::SessionId;
+    use std::sync::Mutex;
+    use tokio_util::sync::CancellationToken;
+
+    fn ctx(cwd: PathBuf) -> ToolCtx {
+        let bus = EventBus::new();
+        let service = Arc::new(PermissionService::new(bus.clone()));
+        spawn_auto_approve(bus, service.clone());
+        let (metadata, _rx) = MetadataSink::channel();
+        ToolCtx {
+            session_id: SessionId::new(),
+            message_id: harness_core::types::MessageId::new(),
+            call_id: "call_1".into(),
+            data_dir: cwd.join("tool-output"),
+            cwd,
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                service,
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+        }
+    }
+
+    #[tokio::test]
+    async fn creates_new_file_when_old_string_is_empty() {
+        let dir = tempfile::tempdir().unwrap();
+        let result = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "newfile.txt", "old_string": "", "new_string": "new content"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(result.metadata["diff"]
+            .as_str()
+            .unwrap()
+            .contains("new content"));
+        assert_eq!(
+            std::fs::read_to_string(dir.path().join("newfile.txt")).unwrap(),
+            "new content"
+        );
+    }
+
+    #[tokio::test]
+    async fn rejects_empty_old_string_on_existing_file() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("existing.txt"), "original").unwrap();
+        let err = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "existing.txt", "old_string": "", "new_string": "x"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap_err();
+        assert!(
+            matches!(&err, ToolError::Invalid(msg) if msg.contains("oldString cannot be empty"))
+        );
+        assert_eq!(
+            std::fs::read_to_string(dir.path().join("existing.txt")).unwrap(),
+            "original"
+        );
+    }
+
+    #[tokio::test]
+    async fn creates_new_file_with_nested_directories() {
+        let dir = tempfile::tempdir().unwrap();
+        EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "nested/dir/file.txt", "old_string": "", "new_string": "nested file"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert_eq!(
+            std::fs::read_to_string(dir.path().join("nested/dir/file.txt")).unwrap(),
+            "nested file"
+        );
+    }
+
+    #[tokio::test]
+    async fn replaces_text_in_existing_file() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("existing.txt"), "old content here").unwrap();
+        let result = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "existing.txt", "old_string": "old content", "new_string": "new content"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(result.output.contains("Edit applied successfully"));
+        assert_eq!(
+            std::fs::read_to_string(dir.path().join("existing.txt")).unwrap(),
+            "new content here"
+        );
+    }
+
+    #[tokio::test]
+    async fn preserves_bom_and_only_edits_visible_content() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut original = UTF8_BOM.to_vec();
+        original.extend_from_slice(b"using System;\nclass Test {}\n");
+        std::fs::write(dir.path().join("existing.cs"), &original).unwrap();
+
+        let result = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "existing.cs", "old_string": "using System;", "new_string": "using Up;"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(result.metadata["diff"]
+            .as_str()
+            .unwrap()
+            .contains("-using System;"));
+        assert!(result.metadata["diff"]
+            .as_str()
+            .unwrap()
+            .contains("+using Up;"));
+
+        let content = std::fs::read(dir.path().join("existing.cs")).unwrap();
+        assert_eq!(&content[..3], &UTF8_BOM);
+        assert_eq!(&content[3..], b"using Up;\nclass Test {}\n");
+    }
+
+    #[tokio::test]
+    async fn errors_when_file_does_not_exist() {
+        let dir = tempfile::tempdir().unwrap();
+        let err = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "nonexistent.txt", "old_string": "old", "new_string": "new"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap_err();
+        assert!(matches!(&err, ToolError::Invalid(msg) if msg.contains("not found")));
+    }
+
+    #[tokio::test]
+    async fn errors_when_old_equals_new() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("file.txt"), "content").unwrap();
+        let err = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "file.txt", "old_string": "same", "new_string": "same"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap_err();
+        assert!(matches!(&err, ToolError::Invalid(msg) if msg.contains("identical")));
+    }
+
+    #[tokio::test]
+    async fn errors_when_path_is_a_directory() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::create_dir(dir.path().join("adir")).unwrap();
+        let err = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "adir", "old_string": "old", "new_string": "new"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap_err();
+        assert!(matches!(&err, ToolError::Invalid(msg) if msg.contains("directory")));
+    }
+
+    #[tokio::test]
+    async fn replaces_all_occurrences_with_replace_all_option() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("file.txt"), "foo bar foo baz foo").unwrap();
+        EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "file.txt", "old_string": "foo", "new_string": "qux", "replace_all": true}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert_eq!(
+            std::fs::read_to_string(dir.path().join("file.txt")).unwrap(),
+            "qux bar qux baz qux"
+        );
+    }
+
+    #[tokio::test]
+    async fn handles_crlf_line_endings() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("file.txt"), "line1\r\nold\r\nline3").unwrap();
+        EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "file.txt", "old_string": "old", "new_string": "new"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert_eq!(
+            std::fs::read_to_string(dir.path().join("file.txt")).unwrap(),
+            "line1\r\nnew\r\nline3"
+        );
+    }
+
+    #[tokio::test]
+    async fn tracks_file_diff_statistics() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("file.txt"), "line1\nline2\nline3").unwrap();
+        let result = EditTool::new()
+            .execute(
+                serde_json::json!({"file_path": "file.txt", "old_string": "line2", "new_string": "new line a\nnew line b"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(result.metadata["added"].as_u64().unwrap() > 0);
+    }
+
+    #[tokio::test]
+    async fn concurrent_edits_to_the_same_file_serialize_without_losing_either_change() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(
+            dir.path().join("file.txt"),
+            "top = 0\nmiddle = keep\nbottom = 0\n",
+        )
+        .unwrap();
+        let tool = Arc::new(EditTool::new());
+
+        let t1 = tool.clone();
+        let cwd1 = dir.path().to_path_buf();
+        let h1 = tokio::spawn(async move {
+            t1.execute(
+                serde_json::json!({"file_path": "file.txt", "old_string": "top = 0", "new_string": "top = 1"}),
+                ctx(cwd1),
+            )
+            .await
+        });
+
+        let t2 = tool.clone();
+        let cwd2 = dir.path().to_path_buf();
+        let h2 = tokio::spawn(async move {
+            t2.execute(
+                serde_json::json!({"file_path": "file.txt", "old_string": "bottom = 0", "new_string": "bottom = 2"}),
+                ctx(cwd2),
+            )
+            .await
+        });
+
+        h1.await.unwrap().unwrap();
+        h2.await.unwrap().unwrap();
+
+        assert_eq!(
+            std::fs::read_to_string(dir.path().join("file.txt")).unwrap(),
+            "top = 1\nmiddle = keep\nbottom = 2\n"
+        );
+    }
+}

crates/harness-tools/src/edit/replacers.rs

@@ -0,0 +1,696 @@
+//! Verbatim port of opencode's replacer chain (`packages/opencode/src/tool/edit.ts`).
+//! Each replacer returns candidate matches in the same order the TS generator yields them;
+//! `replace()` walks replacers in a fixed sequence and takes the first candidate that
+//! resolves unambiguously (or throws — the disproportionate-match guard is a hard stop,
+//! not a "try the next candidate").
+
+#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
+pub enum EditError {
+    #[error("No changes to apply: oldString and newString are identical.")]
+    Identical,
+    #[error("oldString cannot be empty when editing an existing file. Provide the exact text to replace, or use write for an intentional full-file replacement.")]
+    EmptyOldString,
+    #[error("Could not find oldString in the file. It must match exactly, including whitespace, indentation, and line endings.")]
+    NotFound,
+    #[error("Found multiple matches for oldString. Provide more surrounding context to make the match unique.")]
+    MultipleMatches,
+    #[error("Refusing replacement because the matched span is much larger than oldString. Re-read the file and provide the full exact oldString for the intended replacement.")]
+    DisproportionateMatch,
+}
+
+const SINGLE_CANDIDATE_SIMILARITY_THRESHOLD: f64 = 0.65;
+const MULTIPLE_CANDIDATES_SIMILARITY_THRESHOLD: f64 = 0.65;
+
+fn levenshtein(a: &str, b: &str) -> usize {
+    let a: Vec<char> = a.chars().collect();
+    let b: Vec<char> = b.chars().collect();
+    if a.is_empty() || b.is_empty() {
+        return a.len().max(b.len());
+    }
+    let mut matrix = vec![vec![0usize; b.len() + 1]; a.len() + 1];
+    for (i, row) in matrix.iter_mut().enumerate() {
+        row[0] = i;
+    }
+    for (j, cell) in matrix[0].iter_mut().enumerate() {
+        *cell = j;
+    }
+    for i in 1..=a.len() {
+        for j in 1..=b.len() {
+            let cost = if a[i - 1] == b[j - 1] { 0 } else { 1 };
+            matrix[i][j] = (matrix[i - 1][j] + 1)
+                .min(matrix[i][j - 1] + 1)
+                .min(matrix[i - 1][j - 1] + cost);
+        }
+    }
+    matrix[a.len()][b.len()]
+}
+
+fn simple(_content: &str, find: &str) -> Vec<String> {
+    vec![find.to_string()]
+}
+
+fn line_trimmed(content: &str, find: &str) -> Vec<String> {
+    let original_lines: Vec<&str> = content.split('\n').collect();
+    let mut search_lines: Vec<&str> = find.split('\n').collect();
+    if search_lines.last() == Some(&"") {
+        search_lines.pop();
+    }
+    let mut out = Vec::new();
+    if search_lines.is_empty() || original_lines.len() < search_lines.len() {
+        return out;
+    }
+
+    for i in 0..=(original_lines.len() - search_lines.len()) {
+        let matches =
+            (0..search_lines.len()).all(|j| original_lines[i + j].trim() == search_lines[j].trim());
+        if matches {
+            let mut match_start = 0usize;
+            for line in &original_lines[..i] {
+                match_start += line.len() + 1;
+            }
+            let mut match_end = match_start;
+            for k in 0..search_lines.len() {
+                match_end += original_lines[i + k].len();
+                if k < search_lines.len() - 1 {
+                    match_end += 1;
+                }
+            }
+            out.push(content[match_start..match_end].to_string());
+        }
+    }
+    out
+}
+
+fn extract_lines(original_lines: &[&str], content: &str, start: usize, end: usize) -> String {
+    let mut match_start = 0usize;
+    for line in &original_lines[..start] {
+        match_start += line.len() + 1;
+    }
+    let mut match_end = match_start;
+    for (k, line) in original_lines.iter().enumerate().take(end + 1).skip(start) {
+        match_end += line.len();
+        if k < end {
+            match_end += 1;
+        }
+    }
+    content[match_start..match_end].to_string()
+}
+
+fn block_anchor(content: &str, find: &str) -> Vec<String> {
+    let original_lines: Vec<&str> = content.split('\n').collect();
+    let mut search_lines: Vec<&str> = find.split('\n').collect();
+    if search_lines.len() < 3 {
+        return vec![];
+    }
+    if search_lines.last() == Some(&"") {
+        search_lines.pop();
+    }
+
+    let first_line_search = search_lines[0].trim();
+    let last_line_search = search_lines[search_lines.len() - 1].trim();
+    let search_block_size = search_lines.len();
+    let max_line_delta = ((search_block_size as f64 * 0.25).floor() as i64).max(1);
+
+    let mut candidates: Vec<(usize, usize)> = Vec::new();
+    for i in 0..original_lines.len() {
+        if original_lines[i].trim() != first_line_search {
+            continue;
+        }
+        let mut j = i + 2;
+        while j < original_lines.len() {
+            if original_lines[j].trim() == last_line_search {
+                let actual_block_size = (j - i + 1) as i64;
+                if (actual_block_size - search_block_size as i64).abs() <= max_line_delta {
+                    candidates.push((i, j));
+                }
+                break;
+            }
+            j += 1;
+        }
+    }
+    if candidates.is_empty() {
+        return vec![];
+    }
+
+    let middle_similarity = |start: usize, end: usize, early_exit: bool| -> f64 {
+        let actual_block_size = end - start + 1;
+        let lines_to_check = (search_block_size as i64 - 2).min(actual_block_size as i64 - 2);
+        if lines_to_check <= 0 {
+            return 1.0;
+        }
+        let lines_to_check = lines_to_check as usize;
+        let upper = (search_block_size - 1).min(actual_block_size - 1);
+        let mut similarity = 0.0f64;
+        for j in 1..upper {
+            let original_line = original_lines[start + j].trim();
+            let search_line = search_lines[j].trim();
+            let max_len = original_line.len().max(search_line.len());
+            if max_len == 0 {
+                continue;
+            }
+            let distance = levenshtein(original_line, search_line);
+            similarity += (1.0 - distance as f64 / max_len as f64) / lines_to_check as f64;
+            if early_exit && similarity >= SINGLE_CANDIDATE_SIMILARITY_THRESHOLD {
+                break;
+            }
+        }
+        similarity
+    };
+
+    if candidates.len() == 1 {
+        let (start, end) = candidates[0];
+        let similarity = middle_similarity(start, end, true);
+        if similarity >= SINGLE_CANDIDATE_SIMILARITY_THRESHOLD {
+            return vec![extract_lines(&original_lines, content, start, end)];
+        }
+        return vec![];
+    }
+
+    let mut best: Option<(usize, usize)> = None;
+    let mut max_similarity = -1.0f64;
+    for &(start, end) in &candidates {
+        let actual_block_size = end - start + 1;
+        let lines_to_check = (search_block_size as i64 - 2).min(actual_block_size as i64 - 2);
+        let similarity = if lines_to_check <= 0 {
+            1.0
+        } else {
+            let upper = (search_block_size - 1).min(actual_block_size - 1);
+            let mut sum = 0.0;
+            for j in 1..upper {
+                let original_line = original_lines[start + j].trim();
+                let search_line = search_lines[j].trim();
+                let max_len = original_line.len().max(search_line.len());
+                if max_len == 0 {
+                    continue;
+                }
+                sum += 1.0 - levenshtein(original_line, search_line) as f64 / max_len as f64;
+            }
+            sum / lines_to_check as f64
+        };
+        if similarity > max_similarity {
+            max_similarity = similarity;
+            best = Some((start, end));
+        }
+    }
+    if max_similarity >= MULTIPLE_CANDIDATES_SIMILARITY_THRESHOLD {
+        if let Some((start, end)) = best {
+            return vec![extract_lines(&original_lines, content, start, end)];
+        }
+    }
+    vec![]
+}
+
+fn normalize_whitespace(text: &str) -> String {
+    let mut out = String::new();
+    let mut in_ws = false;
+    for c in text.chars() {
+        if c.is_whitespace() {
+            if !in_ws {
+                out.push(' ');
+                in_ws = true;
+            }
+        } else {
+            out.push(c);
+            in_ws = false;
+        }
+    }
+    out.trim().to_string()
+}
+
+/// Finds `words` in `line` separated by runs of whitespace, returning the exact
+/// (non-normalized) substring — a hand-rolled stand-in for the TS version's regex built
+/// from escaped words joined by `\s+`.
+fn find_flexible_whitespace_match(line: &str, words: &[&str]) -> Option<String> {
+    if words.is_empty() {
+        return None;
+    }
+    let chars: Vec<char> = line.chars().collect();
+    let n = chars.len();
+    'outer: for start in 0..n {
+        let mut pos = start;
+        for (wi, word) in words.iter().enumerate() {
+            let wchars: Vec<char> = word.chars().collect();
+            if pos + wchars.len() > n || chars[pos..pos + wchars.len()] != wchars[..] {
+                continue 'outer;
+            }
+            pos += wchars.len();
+            if wi < words.len() - 1 {
+                let ws_start = pos;
+                while pos < n && chars[pos].is_whitespace() {
+                    pos += 1;
+                }
+                if pos == ws_start {
+                    continue 'outer;
+                }
+            }
+        }
+        return Some(chars[start..pos].iter().collect());
+    }
+    None
+}
+
+fn whitespace_normalized(content: &str, find: &str) -> Vec<String> {
+    let mut out = Vec::new();
+    let normalized_find = normalize_whitespace(find);
+
+    let lines: Vec<&str> = content.split('\n').collect();
+    for line in &lines {
+        if normalize_whitespace(line) == normalized_find {
+            out.push(line.to_string());
+        } else if normalize_whitespace(line).contains(&normalized_find) {
+            let words: Vec<&str> = find.split_whitespace().collect();
+            if let Some(m) = find_flexible_whitespace_match(line, &words) {
+                out.push(m);
+            }
+        }
+    }
+
+    let find_lines: Vec<&str> = find.split('\n').collect();
+    if find_lines.len() > 1 && lines.len() >= find_lines.len() {
+        for i in 0..=(lines.len() - find_lines.len()) {
+            let block = lines[i..i + find_lines.len()].join("\n");
+            if normalize_whitespace(&block) == normalized_find {
+                out.push(block);
+            }
+        }
+    }
+    out
+}
+
+fn remove_indentation(text: &str) -> String {
+    let lines: Vec<&str> = text.split('\n').collect();
+    let non_empty: Vec<&&str> = lines.iter().filter(|l| !l.trim().is_empty()).collect();
+    if non_empty.is_empty() {
+        return text.to_string();
+    }
+    let min_indent = non_empty
+        .iter()
+        .map(|l| l.len() - l.trim_start().len())
+        .min()
+        .unwrap_or(0);
+    lines
+        .iter()
+        .map(|l| {
+            if l.trim().is_empty() {
+                l.to_string()
+            } else {
+                l[min_indent.min(l.len())..].to_string()
+            }
+        })
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+fn indentation_flexible(content: &str, find: &str) -> Vec<String> {
+    let normalized_find = remove_indentation(find);
+    let content_lines: Vec<&str> = content.split('\n').collect();
+    let find_lines: Vec<&str> = find.split('\n').collect();
+    let mut out = Vec::new();
+    if content_lines.len() < find_lines.len() {
+        return out;
+    }
+    for i in 0..=(content_lines.len() - find_lines.len()) {
+        let block = content_lines[i..i + find_lines.len()].join("\n");
+        if remove_indentation(&block) == normalized_find {
+            out.push(block);
+        }
+    }
+    out
+}
+
+fn unescape_string(s: &str) -> String {
+    let mut out = String::new();
+    let mut chars = s.chars().peekable();
+    while let Some(c) = chars.next() {
+        if c == '\\' {
+            match chars.peek().copied() {
+                Some('n') => {
+                    out.push('\n');
+                    chars.next();
+                }
+                Some('t') => {
+                    out.push('\t');
+                    chars.next();
+                }
+                Some('r') => {
+                    out.push('\r');
+                    chars.next();
+                }
+                Some('\'') => {
+                    out.push('\'');
+                    chars.next();
+                }
+                Some('"') => {
+                    out.push('"');
+                    chars.next();
+                }
+                Some('`') => {
+                    out.push('`');
+                    chars.next();
+                }
+                Some('\\') => {
+                    out.push('\\');
+                    chars.next();
+                }
+                Some('\n') => {
+                    out.push('\n');
+                    chars.next();
+                }
+                Some('$') => {
+                    out.push('$');
+                    chars.next();
+                }
+                _ => out.push(c),
+            }
+        } else {
+            out.push(c);
+        }
+    }
+    out
+}
+
+fn escape_normalized(content: &str, find: &str) -> Vec<String> {
+    let mut out = Vec::new();
+    let unescaped_find = unescape_string(find);
+    if content.contains(&unescaped_find) {
+        out.push(unescaped_find.clone());
+    }
+    let lines: Vec<&str> = content.split('\n').collect();
+    let find_lines: Vec<&str> = unescaped_find.split('\n').collect();
+    if lines.len() >= find_lines.len() {
+        for i in 0..=(lines.len() - find_lines.len()) {
+            let block = lines[i..i + find_lines.len()].join("\n");
+            if unescape_string(&block) == unescaped_find {
+                out.push(block);
+            }
+        }
+    }
+    out
+}
+
+fn multi_occurrence(content: &str, find: &str) -> Vec<String> {
+    let mut out = Vec::new();
+    if find.is_empty() {
+        return out;
+    }
+    let mut start = 0usize;
+    while let Some(idx) = content[start..].find(find) {
+        out.push(find.to_string());
+        start += idx + find.len();
+    }
+    out
+}
+
+fn trimmed_boundary(content: &str, find: &str) -> Vec<String> {
+    let trimmed_find = find.trim();
+    let mut out = Vec::new();
+    if trimmed_find == find {
+        return out;
+    }
+    if content.contains(trimmed_find) {
+        out.push(trimmed_find.to_string());
+    }
+    let lines: Vec<&str> = content.split('\n').collect();
+    let find_lines: Vec<&str> = find.split('\n').collect();
+    if lines.len() >= find_lines.len() {
+        for i in 0..=(lines.len() - find_lines.len()) {
+            let block = lines[i..i + find_lines.len()].join("\n");
+            if block.trim() == trimmed_find {
+                out.push(block);
+            }
+        }
+    }
+    out
+}
+
+fn context_aware(content: &str, find: &str) -> Vec<String> {
+    let mut find_lines: Vec<&str> = find.split('\n').collect();
+    if find_lines.len() < 3 {
+        return vec![];
+    }
+    if find_lines.last() == Some(&"") {
+        find_lines.pop();
+    }
+    let content_lines: Vec<&str> = content.split('\n').collect();
+    let first_line = find_lines[0].trim();
+    let last_line = find_lines[find_lines.len() - 1].trim();
+    let mut out = Vec::new();
+
+    for i in 0..content_lines.len() {
+        if content_lines[i].trim() != first_line {
+            continue;
+        }
+        let mut j = i + 2;
+        while j < content_lines.len() {
+            if content_lines[j].trim() == last_line {
+                let block_lines = &content_lines[i..=j];
+                if block_lines.len() == find_lines.len() {
+                    let mut matching = 0usize;
+                    let mut total_non_empty = 0usize;
+                    for k in 1..block_lines.len() - 1 {
+                        let block_line = block_lines[k].trim();
+                        let find_line = find_lines[k].trim();
+                        if !block_line.is_empty() || !find_line.is_empty() {
+                            total_non_empty += 1;
+                            if block_line == find_line {
+                                matching += 1;
+                            }
+                        }
+                    }
+                    if total_non_empty == 0 || (matching as f64 / total_non_empty as f64) >= 0.5 {
+                        out.push(block_lines.join("\n"));
+                    }
+                }
+                break;
+            }
+            j += 1;
+        }
+    }
+    out
+}
+
+type ReplacerFn = fn(&str, &str) -> Vec<String>;
+
+const REPLACER_CHAIN: &[ReplacerFn] = &[
+    simple,
+    line_trimmed,
+    block_anchor,
+    whitespace_normalized,
+    indentation_flexible,
+    escape_normalized,
+    trimmed_boundary,
+    context_aware,
+    multi_occurrence,
+];
+
+fn is_disproportionate_match(search: &str, old_string: &str) -> bool {
+    let old_lines = old_string.split('\n').count() as i64;
+    let search_lines = search.split('\n').count() as i64;
+    if search_lines >= (old_lines + 3).max(old_lines * 2) {
+        return true;
+    }
+    if old_lines == 1 {
+        return false;
+    }
+    let old_trimmed_len = old_string.trim().len() as i64;
+    search.trim().len() as i64 > (old_trimmed_len + 500).max(old_trimmed_len * 4)
+}
+
+/// Applies the replacer chain in order, returning the new content on the first
+/// unambiguous match. Mirrors opencode's `replace()` exactly, including the
+/// disproportionate-match guard short-circuiting immediately (not skipping ahead).
+pub fn replace(
+    content: &str,
+    old_string: &str,
+    new_string: &str,
+    replace_all: bool,
+) -> Result<String, EditError> {
+    if old_string == new_string {
+        return Err(EditError::Identical);
+    }
+    if old_string.is_empty() {
+        return Err(EditError::EmptyOldString);
+    }
+
+    let mut not_found = true;
+    for replacer in REPLACER_CHAIN {
+        for search in replacer(content, old_string) {
+            let index = match content.find(&search) {
+                Some(i) => i,
+                None => continue,
+            };
+            not_found = false;
+            if is_disproportionate_match(&search, old_string) {
+                return Err(EditError::DisproportionateMatch);
+            }
+            if replace_all {
+                return Ok(content.replace(&search, new_string));
+            }
+            let last_index = content.rfind(&search).unwrap();
+            if index != last_index {
+                continue;
+            }
+            return Ok(format!(
+                "{}{}{}",
+                &content[..index],
+                new_string,
+                &content[index + search.len()..]
+            ));
+        }
+    }
+
+    if not_found {
+        Err(EditError::NotFound)
+    } else {
+        Err(EditError::MultipleMatches)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn replaces_exact_text_in_existing_content() {
+        let result = replace("old content here", "old content", "new content", false).unwrap();
+        assert_eq!(result, "new content here");
+    }
+
+    #[test]
+    fn rejects_identical_old_and_new() {
+        assert_eq!(
+            replace("content", "same", "same", false),
+            Err(EditError::Identical)
+        );
+    }
+
+    #[test]
+    fn rejects_empty_old_string() {
+        assert_eq!(
+            replace("content", "", "x", false),
+            Err(EditError::EmptyOldString)
+        );
+    }
+
+    #[test]
+    fn reports_not_found() {
+        let err = replace("actual content", "not in file", "replacement", false).unwrap_err();
+        assert_eq!(err, EditError::NotFound);
+    }
+
+    #[test]
+    fn rejects_loose_block_anchor_matches() {
+        let original = [
+            "function configure() {",
+            "  keepImportantState()",
+            "  removeAllUserData()",
+            "  archiveBackups()",
+            "  auditLog()",
+            "}",
+        ]
+        .join("\n");
+        let old = ["function configure() {", "  const enabled = true", "}"].join("\n");
+        let new = ["function configure() {", "  const enabled = false", "}"].join("\n");
+        assert_eq!(
+            replace(&original, &old, &new, false),
+            Err(EditError::NotFound)
+        );
+    }
+
+    #[test]
+    fn rejects_block_anchor_matches_with_unrelated_middle_content() {
+        let original = ["function configure() {", "  removeAllUserData()", "}"].join("\n");
+        let old = ["function configure() {", "  const enabled = true", "}"].join("\n");
+        let new = ["function configure() {", "  const enabled = false", "}"].join("\n");
+        assert_eq!(
+            replace(&original, &old, &new, false),
+            Err(EditError::NotFound)
+        );
+    }
+
+    #[test]
+    fn replaces_all_occurrences_with_replace_all() {
+        let result = replace("foo bar foo baz foo", "foo", "qux", true).unwrap();
+        assert_eq!(result, "qux bar qux baz qux");
+    }
+
+    #[test]
+    fn rejects_ambiguous_single_occurrence_request() {
+        let err = replace("foo bar foo", "foo", "qux", false).unwrap_err();
+        assert_eq!(err, EditError::MultipleMatches);
+    }
+
+    #[test]
+    fn handles_multiline_replacements() {
+        let result = replace(
+            "line1\nline2\nline3",
+            "line2",
+            "new line 2\nextra line",
+            false,
+        )
+        .unwrap();
+        assert_eq!(result, "line1\nnew line 2\nextra line\nline3");
+    }
+
+    #[test]
+    fn line_trimmed_replacer_matches_despite_surrounding_whitespace_changes() {
+        let content = "  line one  \n  line two  \n";
+        let matches = line_trimmed(content, "line one\nline two");
+        assert_eq!(matches, vec!["  line one  \n  line two  ".to_string()]);
+    }
+
+    #[test]
+    fn whitespace_normalized_matches_reformatted_line() {
+        let content = "const x   =    1;\n";
+        let matches = whitespace_normalized(content, "const x = 1;");
+        assert!(matches.iter().any(|m| m == "const x   =    1;"));
+    }
+
+    #[test]
+    fn indentation_flexible_matches_reindented_block() {
+        let content = "    if true {\n        foo()\n    }\n";
+        let find = "if true {\n    foo()\n}";
+        let matches = indentation_flexible(content, find);
+        assert_eq!(matches.len(), 1);
+    }
+
+    #[test]
+    fn escape_normalized_matches_escaped_newline_in_find() {
+        let content = "line one\nline two\n";
+        let matches = escape_normalized(content, "line one\\nline two");
+        assert!(matches.contains(&"line one\nline two".to_string()));
+    }
+
+    #[test]
+    fn trimmed_boundary_matches_padded_find() {
+        let content = "exact text here\n";
+        let matches = trimmed_boundary(content, "  exact text here  ");
+        assert!(matches.contains(&"exact text here".to_string()));
+    }
+
+    #[test]
+    fn disproportionate_match_flags_far_larger_line_count() {
+        let old = "one line";
+        let search = "a\nb\nc\nd"; // 4 lines >= max(1+3, 1*2) = 4
+        assert!(is_disproportionate_match(search, old));
+    }
+
+    #[test]
+    fn disproportionate_match_flags_far_larger_char_span_same_line_count() {
+        let old = "short";
+        let search = "x".repeat(600); // > max(5+500, 5*4) = 505, still 1 line like `old`
+        assert!(!is_disproportionate_match(&search, old)); // old_lines == 1 short-circuits to false
+
+        let old_multiline = "short\nline";
+        let search_multiline = format!("{}\nline", "x".repeat(600));
+        assert!(is_disproportionate_match(&search_multiline, old_multiline));
+    }
+
+    #[test]
+    fn disproportionate_match_allows_reasonably_sized_matches() {
+        assert!(!is_disproportionate_match(
+            "old content",
+            "old content here"
+        ));
+    }
+}

crates/harness-tools/src/glob.rs

@@ -0,0 +1,181 @@
+use std::path::PathBuf;
+use std::time::SystemTime;
+
+use async_trait::async_trait;
+use globset::GlobBuilder;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use ignore::WalkBuilder;
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+use crate::paths;
+
+const MAX_RESULTS: usize = 100;
+
+#[derive(Debug, Deserialize, JsonSchema)]
+struct GlobParams {
+    pattern: String,
+    path: Option<String>,
+}
+
+/// Gitignore-aware file finder — read-only, no permission ask.
+pub struct GlobTool;
+
+#[async_trait]
+impl Tool for GlobTool {
+    fn name(&self) -> &str {
+        "glob"
+    }
+
+    fn description(&self) -> &str {
+        "Finds files by glob pattern (gitignore-aware), most recently modified first."
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        serde_json::to_value(schemars::schema_for!(GlobParams)).unwrap()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let params: GlobParams =
+            serde_json::from_value(input).map_err(|e| ToolError::Invalid(e.to_string()))?;
+        let root = params
+            .path
+            .as_ref()
+            .map(|p| paths::resolve(&ctx.cwd, p))
+            .unwrap_or_else(|| ctx.cwd.clone());
+        let pattern = params.pattern.clone();
+
+        let matcher = GlobBuilder::new(&pattern)
+            .literal_separator(true)
+            .build()
+            .map_err(|e| ToolError::Invalid(e.to_string()))?
+            .compile_matcher();
+
+        let mut matches: Vec<(PathBuf, SystemTime)> = Vec::new();
+        for entry in WalkBuilder::new(&root)
+            .hidden(false)
+            .require_git(false)
+            .build()
+        {
+            let entry = match entry {
+                Ok(e) => e,
+                Err(_) => continue,
+            };
+            if !entry.file_type().map(|t| t.is_file()).unwrap_or(false) {
+                continue;
+            }
+            let rel = entry.path().strip_prefix(&root).unwrap_or(entry.path());
+            if !matcher.is_match(rel) {
+                continue;
+            }
+            let mtime = entry
+                .metadata()
+                .ok()
+                .and_then(|m| m.modified().ok())
+                .unwrap_or(SystemTime::UNIX_EPOCH);
+            matches.push((entry.path().to_path_buf(), mtime));
+        }
+
+        matches.sort_by_key(|(_, mtime)| std::cmp::Reverse(*mtime));
+        let truncated = matches.len() > MAX_RESULTS;
+        matches.truncate(MAX_RESULTS);
+
+        let mut output = matches
+            .iter()
+            .map(|(p, _)| p.display().to_string())
+            .collect::<Vec<_>>()
+            .join("\n");
+        if truncated {
+            output.push_str(
+                "\n\n(Results are truncated: showing first 100 results. Consider using a more specific path or pattern.)",
+            );
+        }
+        if output.is_empty() {
+            output = "(no matches)".to_string();
+        }
+        Ok(ToolOutput::new(pattern, output))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use harness_core::event::EventBus;
+    use harness_core::permission::PermissionService;
+    use harness_core::tool::{MetadataSink, PermissionHandle};
+    use harness_core::types::SessionId;
+    use std::sync::{Arc, Mutex};
+    use tokio_util::sync::CancellationToken;
+
+    fn ctx(cwd: std::path::PathBuf) -> ToolCtx {
+        let bus = EventBus::new();
+        let service = Arc::new(PermissionService::new(bus));
+        let (metadata, _rx) = MetadataSink::channel();
+        ToolCtx {
+            session_id: SessionId::new(),
+            message_id: harness_core::types::MessageId::new(),
+            call_id: "call_1".into(),
+            data_dir: cwd.join("tool-output"),
+            cwd,
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                service,
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+        }
+    }
+
+    #[tokio::test]
+    async fn finds_files_matching_pattern() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("a.rs"), "").unwrap();
+        std::fs::write(dir.path().join("b.txt"), "").unwrap();
+        let output = GlobTool
+            .execute(
+                serde_json::json!({"pattern": "*.rs"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output.output.ends_with("a.rs"));
+        assert!(!output.output.contains("b.txt"));
+    }
+
+    #[tokio::test]
+    async fn respects_gitignore() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join(".gitignore"), "ignored.rs\n").unwrap();
+        std::fs::write(dir.path().join("ignored.rs"), "").unwrap();
+        std::fs::write(dir.path().join("kept.rs"), "").unwrap();
+        let output = GlobTool
+            .execute(
+                serde_json::json!({"pattern": "*.rs"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output.output.contains("kept.rs"));
+        assert!(!output.output.contains("ignored.rs"));
+    }
+
+    #[tokio::test]
+    async fn no_matches_reports_clearly() {
+        let dir = tempfile::tempdir().unwrap();
+        let output = GlobTool
+            .execute(
+                serde_json::json!({"pattern": "*.nonexistent"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert_eq!(output.output, "(no matches)");
+    }
+}

crates/harness-tools/src/grep.rs

@@ -0,0 +1,209 @@
+use std::path::PathBuf;
+use std::time::SystemTime;
+
+use async_trait::async_trait;
+use globset::GlobBuilder;
+use grep_regex::RegexMatcher;
+use grep_searcher::sinks::UTF8;
+use grep_searcher::SearcherBuilder;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use ignore::WalkBuilder;
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+use crate::paths;
+
+const MAX_RESULTS: usize = 100;
+
+#[derive(Debug, Deserialize, JsonSchema)]
+struct GrepParams {
+    pattern: String,
+    path: Option<String>,
+    include: Option<String>,
+}
+
+struct Hit {
+    file: PathBuf,
+    line_number: u64,
+    line: String,
+    mtime: SystemTime,
+}
+
+/// Content search via ripgrep's engine as a library — read-only, no permission ask.
+pub struct GrepTool;
+
+#[async_trait]
+impl Tool for GrepTool {
+    fn name(&self) -> &str {
+        "grep"
+    }
+
+    fn description(&self) -> &str {
+        "Searches file contents by regex (gitignore-aware), most recently modified first."
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        serde_json::to_value(schemars::schema_for!(GrepParams)).unwrap()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let params: GrepParams =
+            serde_json::from_value(input).map_err(|e| ToolError::Invalid(e.to_string()))?;
+        let root = params
+            .path
+            .as_ref()
+            .map(|p| paths::resolve(&ctx.cwd, p))
+            .unwrap_or_else(|| ctx.cwd.clone());
+
+        let matcher =
+            RegexMatcher::new(&params.pattern).map_err(|e| ToolError::Invalid(e.to_string()))?;
+        let include_matcher = params
+            .include
+            .as_ref()
+            .map(|inc| GlobBuilder::new(inc).literal_separator(true).build())
+            .transpose()
+            .map_err(|e| ToolError::Invalid(e.to_string()))?
+            .map(|g| g.compile_matcher());
+
+        let mut hits: Vec<Hit> = Vec::new();
+        for entry in WalkBuilder::new(&root)
+            .hidden(false)
+            .require_git(false)
+            .build()
+        {
+            let entry = match entry {
+                Ok(e) => e,
+                Err(_) => continue,
+            };
+            if !entry.file_type().map(|t| t.is_file()).unwrap_or(false) {
+                continue;
+            }
+            let rel = entry.path().strip_prefix(&root).unwrap_or(entry.path());
+            if let Some(glob) = &include_matcher {
+                if !glob.is_match(rel) {
+                    continue;
+                }
+            }
+            let mtime = entry
+                .metadata()
+                .ok()
+                .and_then(|m| m.modified().ok())
+                .unwrap_or(SystemTime::UNIX_EPOCH);
+
+            let path = entry.path().to_path_buf();
+            let mut searcher = SearcherBuilder::new().line_number(true).build();
+            let _ = searcher.search_path(
+                &matcher,
+                &path,
+                UTF8(|line_number, line| {
+                    hits.push(Hit {
+                        file: path.clone(),
+                        line_number,
+                        line: line.trim_end().to_string(),
+                        mtime,
+                    });
+                    Ok(true)
+                }),
+            );
+        }
+
+        hits.sort_by_key(|h| std::cmp::Reverse(h.mtime));
+        let truncated = hits.len() > MAX_RESULTS;
+        hits.truncate(MAX_RESULTS);
+
+        let mut output = hits
+            .iter()
+            .map(|h| format!("{}:{}:{}", h.file.display(), h.line_number, h.line))
+            .collect::<Vec<_>>()
+            .join("\n");
+        if truncated {
+            output.push_str("\n\n(Results are truncated: showing first 100 matches. Consider a more specific pattern.)");
+        }
+        if output.is_empty() {
+            output = "(no matches)".to_string();
+        }
+        Ok(ToolOutput::new(params.pattern, output))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use harness_core::event::EventBus;
+    use harness_core::permission::PermissionService;
+    use harness_core::tool::{MetadataSink, PermissionHandle};
+    use harness_core::types::SessionId;
+    use std::sync::{Arc, Mutex};
+    use tokio_util::sync::CancellationToken;
+
+    fn ctx(cwd: std::path::PathBuf) -> ToolCtx {
+        let bus = EventBus::new();
+        let service = Arc::new(PermissionService::new(bus));
+        let (metadata, _rx) = MetadataSink::channel();
+        ToolCtx {
+            session_id: SessionId::new(),
+            message_id: harness_core::types::MessageId::new(),
+            call_id: "call_1".into(),
+            data_dir: cwd.join("tool-output"),
+            cwd,
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                service,
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+        }
+    }
+
+    #[tokio::test]
+    async fn finds_matching_lines() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("a.rs"), "fn main() {}\nfn helper() {}\n").unwrap();
+        let output = GrepTool
+            .execute(
+                serde_json::json!({"pattern": "fn main"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output.output.contains("a.rs:1:fn main() {}"));
+        assert!(!output.output.contains("helper"));
+    }
+
+    #[tokio::test]
+    async fn filters_by_include_glob() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("a.rs"), "needle\n").unwrap();
+        std::fs::write(dir.path().join("b.txt"), "needle\n").unwrap();
+        let output = GrepTool
+            .execute(
+                serde_json::json!({"pattern": "needle", "include": "*.rs"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output.output.contains("a.rs"));
+        assert!(!output.output.contains("b.txt"));
+    }
+
+    #[tokio::test]
+    async fn no_matches_reports_clearly() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("a.rs"), "content\n").unwrap();
+        let output = GrepTool
+            .execute(
+                serde_json::json!({"pattern": "nonexistent"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert_eq!(output.output, "(no matches)");
+    }
+}

crates/harness-tools/src/lib.rs

@@ -1 +1,28 @@
-// Built-in tools (bash, read, edit, write, glob, grep, todo, webfetch, task) land here in M1/M4.
+mod bash;
+mod edit;
+mod glob;
+mod grep;
+mod paths;
+mod read;
+mod write;
+
+pub use bash::BashTool;
+pub use edit::EditTool;
+pub use glob::GlobTool;
+pub use grep::GrepTool;
+pub use read::ReadTool;
+pub use write::WriteTool;
+
+use std::sync::Arc;
+
+use harness_core::tool::ToolRegistry;
+
+/// Registers all M1 built-ins: read, write, edit, bash, glob, grep.
+pub fn register_builtins(registry: &mut ToolRegistry) {
+    registry.register(Arc::new(ReadTool));
+    registry.register(Arc::new(WriteTool));
+    registry.register(Arc::new(EditTool::new()));
+    registry.register(Arc::new(BashTool));
+    registry.register(Arc::new(GlobTool));
+    registry.register(Arc::new(GrepTool));
+}

crates/harness-tools/src/paths.rs

@@ -0,0 +1,52 @@
+use std::path::{Path, PathBuf};
+
+/// Resolves a (possibly relative) tool-supplied path against `cwd`.
+pub fn resolve(cwd: &Path, file_path: &str) -> PathBuf {
+    let path = Path::new(file_path);
+    if path.is_absolute() {
+        path.to_path_buf()
+    } else {
+        cwd.join(path)
+    }
+}
+
+/// Slash-separated path relative to `cwd`, used as the permission pattern — falls back to
+/// the absolute path (still slash-separated) when `path` isn't under `cwd`.
+pub fn relative_pattern(cwd: &Path, path: &Path) -> String {
+    let rel = path.strip_prefix(cwd).unwrap_or(path);
+    rel.to_string_lossy().replace('\\', "/")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn resolve_joins_relative_paths_to_cwd() {
+        let cwd = Path::new("/home/user/project");
+        assert_eq!(
+            resolve(cwd, "src/main.rs"),
+            PathBuf::from("/home/user/project/src/main.rs")
+        );
+    }
+
+    #[test]
+    fn resolve_leaves_absolute_paths_untouched() {
+        let cwd = Path::new("/home/user/project");
+        assert_eq!(resolve(cwd, "/etc/hosts"), PathBuf::from("/etc/hosts"));
+    }
+
+    #[test]
+    fn relative_pattern_strips_cwd_prefix() {
+        let cwd = Path::new("/home/user/project");
+        let path = Path::new("/home/user/project/src/main.rs");
+        assert_eq!(relative_pattern(cwd, path), "src/main.rs");
+    }
+
+    #[test]
+    fn relative_pattern_falls_back_to_absolute_outside_cwd() {
+        let cwd = Path::new("/home/user/project");
+        let path = Path::new("/etc/hosts");
+        assert_eq!(relative_pattern(cwd, path), "/etc/hosts");
+    }
+}

crates/harness-tools/src/read.rs

@@ -0,0 +1,202 @@
+use async_trait::async_trait;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+use crate::paths;
+
+const DEFAULT_LIMIT: usize = 2000;
+const LINE_CHAR_CLAMP: usize = 2000;
+const IMAGE_EXTENSIONS: &[&str] = &["png", "jpg", "jpeg", "gif", "webp", "bmp", "ico", "svg"];
+
+#[derive(Debug, Deserialize, JsonSchema)]
+struct ReadParams {
+    file_path: String,
+    offset: Option<usize>,
+    limit: Option<usize>,
+}
+
+pub struct ReadTool;
+
+fn is_binary(bytes: &[u8]) -> bool {
+    bytes.iter().take(8000).any(|&b| b == 0)
+}
+
+fn clamp_line(line: &str) -> String {
+    if line.chars().count() <= LINE_CHAR_CLAMP {
+        line.to_string()
+    } else {
+        let clamped: String = line.chars().take(LINE_CHAR_CLAMP).collect();
+        format!("{clamped}... [line truncated]")
+    }
+}
+
+#[async_trait]
+impl Tool for ReadTool {
+    fn name(&self) -> &str {
+        "read"
+    }
+
+    fn description(&self) -> &str {
+        "Reads a file from the local filesystem, returning `cat -n`-style numbered lines."
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        serde_json::to_value(schemars::schema_for!(ReadParams)).unwrap()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let params: ReadParams =
+            serde_json::from_value(input).map_err(|e| ToolError::Invalid(e.to_string()))?;
+        let path = paths::resolve(&ctx.cwd, &params.file_path);
+        let pattern = paths::relative_pattern(&ctx.cwd, &path);
+        ctx.ask
+            .ask(
+                "read",
+                pattern.clone(),
+                pattern,
+                serde_json::json!({"file_path": params.file_path}),
+            )
+            .await?;
+
+        if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
+            if IMAGE_EXTENSIONS.contains(&ext.to_lowercase().as_str()) {
+                return Ok(ToolOutput::new(
+                    params.file_path.clone(),
+                    format!(
+                        "{}: image file — binary content not shown",
+                        params.file_path
+                    ),
+                ));
+            }
+        }
+
+        let bytes = tokio::fs::read(&path)
+            .await
+            .map_err(|e| ToolError::Other(format!("{}: {e}", path.display())))?;
+        if is_binary(&bytes) {
+            return Ok(ToolOutput::new(
+                params.file_path.clone(),
+                format!("{}: binary file — content not shown", params.file_path),
+            ));
+        }
+
+        let text = String::from_utf8_lossy(&bytes);
+        let offset = params.offset.unwrap_or(0);
+        let limit = params.limit.unwrap_or(DEFAULT_LIMIT);
+        let numbered: Vec<String> = text
+            .lines()
+            .enumerate()
+            .skip(offset)
+            .take(limit)
+            .map(|(i, line)| format!("{:>6}\t{}", i + 1, clamp_line(line)))
+            .collect();
+
+        if numbered.is_empty() {
+            return Ok(ToolOutput::new(
+                params.file_path.clone(),
+                "(empty file or offset past end)".to_string(),
+            ));
+        }
+        Ok(ToolOutput::new(
+            params.file_path.clone(),
+            numbered.join("\n"),
+        ))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use harness_core::event::EventBus;
+    use harness_core::permission::{spawn_auto_approve, PermissionService};
+    use harness_core::tool::{MetadataSink, PermissionHandle};
+    use harness_core::types::SessionId;
+    use std::sync::{Arc, Mutex};
+    use tokio_util::sync::CancellationToken;
+
+    fn ctx(cwd: std::path::PathBuf) -> ToolCtx {
+        let bus = EventBus::new();
+        let service = Arc::new(PermissionService::new(bus.clone()));
+        spawn_auto_approve(bus, service.clone());
+        let (metadata, _rx) = MetadataSink::channel();
+        ToolCtx {
+            session_id: SessionId::new(),
+            message_id: harness_core::types::MessageId::new(),
+            call_id: "call_1".into(),
+            data_dir: cwd.join("tool-output"),
+            cwd,
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                service,
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+        }
+    }
+
+    #[tokio::test]
+    async fn reads_and_numbers_lines() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("a.txt"), "line one\nline two\nline three").unwrap();
+        let output = ReadTool
+            .execute(
+                serde_json::json!({"file_path": "a.txt"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert_eq!(
+            output.output,
+            "     1\tline one\n     2\tline two\n     3\tline three"
+        );
+    }
+
+    #[tokio::test]
+    async fn respects_offset_and_limit() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("a.txt"), "a\nb\nc\nd\ne").unwrap();
+        let output = ReadTool
+            .execute(
+                serde_json::json!({"file_path": "a.txt", "offset": 1, "limit": 2}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert_eq!(output.output, "     2\tb\n     3\tc");
+    }
+
+    #[tokio::test]
+    async fn detects_binary_content() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("bin.dat"), [0u8, 1, 2, 3]).unwrap();
+        let output = ReadTool
+            .execute(
+                serde_json::json!({"file_path": "bin.dat"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        assert!(output.output.contains("binary file"));
+    }
+
+    #[tokio::test]
+    async fn missing_file_is_a_tool_error() {
+        let dir = tempfile::tempdir().unwrap();
+        let err = ReadTool
+            .execute(
+                serde_json::json!({"file_path": "missing.txt"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap_err();
+        assert!(matches!(err, ToolError::Other(_)));
+    }
+}

crates/harness-tools/src/write.rs

@@ -0,0 +1,150 @@
+use async_trait::async_trait;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use schemars::JsonSchema;
+use serde::Deserialize;
+use similar::TextDiff;
+
+use crate::paths;
+
+#[derive(Debug, Deserialize, JsonSchema)]
+struct WriteParams {
+    file_path: String,
+    content: String,
+}
+
+pub struct WriteTool;
+
+fn unified_diff(before: &str, after: &str, file_path: &str) -> String {
+    TextDiff::from_lines(before, after)
+        .unified_diff()
+        .header(&format!("a/{file_path}"), &format!("b/{file_path}"))
+        .to_string()
+}
+
+#[async_trait]
+impl Tool for WriteTool {
+    fn name(&self) -> &str {
+        "write"
+    }
+
+    fn description(&self) -> &str {
+        "Writes content to a file, creating it (and parent directories) if needed."
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        serde_json::to_value(schemars::schema_for!(WriteParams)).unwrap()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let params: WriteParams =
+            serde_json::from_value(input).map_err(|e| ToolError::Invalid(e.to_string()))?;
+        let path = paths::resolve(&ctx.cwd, &params.file_path);
+        let pattern = paths::relative_pattern(&ctx.cwd, &path);
+
+        let before = tokio::fs::read_to_string(&path).await.unwrap_or_default();
+        let diff = unified_diff(&before, &params.content, &params.file_path);
+
+        // Same permission key as opencode's edit tool — writing is an edit.
+        ctx.ask
+            .ask(
+                "edit",
+                pattern,
+                "*",
+                serde_json::json!({"file_path": params.file_path, "diff": diff}),
+            )
+            .await?;
+
+        if let Some(parent) = path.parent() {
+            tokio::fs::create_dir_all(parent)
+                .await
+                .map_err(|e| ToolError::Other(format!("{}: {e}", parent.display())))?;
+        }
+        tokio::fs::write(&path, &params.content)
+            .await
+            .map_err(|e| ToolError::Other(format!("{}: {e}", path.display())))?;
+
+        let added = diff
+            .lines()
+            .filter(|l| l.starts_with('+') && !l.starts_with("+++"))
+            .count();
+        let removed = diff
+            .lines()
+            .filter(|l| l.starts_with('-') && !l.starts_with("---"))
+            .count();
+        let mut output = ToolOutput::new(
+            params.file_path.clone(),
+            format!("wrote {} bytes", params.content.len()),
+        );
+        output.metadata = serde_json::json!({"diff": diff, "added": added, "removed": removed});
+        Ok(output)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use harness_core::event::EventBus;
+    use harness_core::permission::{spawn_auto_approve, PermissionService};
+    use harness_core::tool::{MetadataSink, PermissionHandle};
+    use harness_core::types::SessionId;
+    use std::sync::{Arc, Mutex};
+    use tokio_util::sync::CancellationToken;
+
+    fn ctx(cwd: std::path::PathBuf) -> ToolCtx {
+        let bus = EventBus::new();
+        let service = Arc::new(PermissionService::new(bus.clone()));
+        spawn_auto_approve(bus, service.clone());
+        let (metadata, _rx) = MetadataSink::channel();
+        ToolCtx {
+            session_id: SessionId::new(),
+            message_id: harness_core::types::MessageId::new(),
+            call_id: "call_1".into(),
+            data_dir: cwd.join("tool-output"),
+            cwd,
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                service,
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+        }
+    }
+
+    #[tokio::test]
+    async fn creates_file_and_parent_dirs() {
+        let dir = tempfile::tempdir().unwrap();
+        WriteTool
+            .execute(
+                serde_json::json!({"file_path": "nested/dir/a.txt", "content": "hello"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        let content = std::fs::read_to_string(dir.path().join("nested/dir/a.txt")).unwrap();
+        assert_eq!(content, "hello");
+    }
+
+    #[tokio::test]
+    async fn overwrites_existing_file_and_reports_diff() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("a.txt"), "old\n").unwrap();
+        let output = WriteTool
+            .execute(
+                serde_json::json!({"file_path": "a.txt", "content": "new\n"}),
+                ctx(dir.path().to_path_buf()),
+            )
+            .await
+            .unwrap();
+        let content = std::fs::read_to_string(dir.path().join("a.txt")).unwrap();
+        assert_eq!(content, "new\n");
+        assert_eq!(output.metadata["added"], 1);
+        assert_eq!(output.metadata["removed"], 1);
+    }
+}

crates/harness-tui/Cargo.toml

@@ -10,6 +10,23 @@ path = "src/main.rs"
 
 [dependencies]
 harness-app = { workspace = true }
+harness-core = { workspace = true }
+tokio = { workspace = true }
+ratatui = { workspace = true }
+crossterm = { workspace = true }
+tui-textarea = { workspace = true }
+pulldown-cmark = { workspace = true }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+tracing-appender = { workspace = true }
+dirs = { workspace = true }
+anyhow = { workspace = true }
+serde_json = { workspace = true }
+futures = { workspace = true }
+tokio-util = { workspace = true }
+
+[dev-dependencies]
+insta = "1"
 
 [lints]
 workspace = true

crates/harness-tui/src/app.rs

@@ -0,0 +1,106 @@
+use std::io;
+use std::path::PathBuf;
+
+use crossterm::event::EventStream;
+use futures::StreamExt;
+use harness_app::EngineHandle;
+use harness_core::event::AppEvent;
+use ratatui::backend::CrosstermBackend;
+use ratatui::Terminal;
+use tokio::sync::broadcast;
+use tokio::time::{interval, Duration};
+
+use crate::input::{apply_action, handle_event};
+use crate::render::render;
+use crate::state::AppState;
+use crate::terminal::TerminalGuard;
+
+const DEFAULT_MODEL: &str = "anthropic/claude-sonnet-4-5";
+const DEFAULT_AGENT: &str = "orchestrator";
+const RENDER_TICK_MS: u64 = 33;
+
+pub struct App {
+    engine: EngineHandle,
+    state: AppState,
+    _terminal_guard: TerminalGuard,
+    terminal: Terminal<CrosstermBackend<io::Stdout>>,
+    events: EventStream,
+    bus_rx: broadcast::Receiver<AppEvent>,
+    render_interval: tokio::time::Interval,
+}
+
+impl App {
+    pub async fn new(cwd: PathBuf) -> anyhow::Result<Self> {
+        let engine = EngineHandle::init(cwd)?;
+        let bus_rx = engine.bus().subscribe();
+
+        let config = engine.config();
+        let model_ref = config
+            .model
+            .clone()
+            .unwrap_or_else(|| DEFAULT_MODEL.to_string());
+        let agent = DEFAULT_AGENT.to_string();
+
+        let mut state = AppState::new(model_ref, agent);
+
+        // Create a default session so the user can start typing immediately.
+        match engine.new_session(&state.agent, &state.model_ref).await {
+            Ok(session_id) => {
+                state.session_id = Some(session_id);
+            }
+            Err(e) => {
+                tracing::warn!(error = %e, "failed to create default session");
+            }
+        }
+
+        let terminal_guard = TerminalGuard::enter()?;
+        let backend = CrosstermBackend::new(io::stdout());
+        let terminal = Terminal::new(backend)?;
+        let events = EventStream::new();
+        let render_interval = interval(Duration::from_millis(RENDER_TICK_MS));
+
+        Ok(Self {
+            engine,
+            state,
+            _terminal_guard: terminal_guard,
+            terminal,
+            events,
+            bus_rx,
+            render_interval,
+        })
+    }
+
+    pub async fn run(&mut self) -> anyhow::Result<()> {
+        self.state.dirty = true;
+        while !self.state.quit {
+            tokio::select! {
+                biased;
+
+                maybe_event = self.events.next() => {
+                    match maybe_event {
+                        Some(Ok(event)) => {
+                            let action = handle_event(event, &mut self.state, &self.engine);
+                            apply_action(action, &mut self.state, &self.engine).await;
+                        }
+                        Some(Err(e)) => {
+                            tracing::error!(error = %e, "input error");
+                        }
+                        None => break,
+                    }
+                }
+
+                Ok(event) = self.bus_rx.recv() => {
+                    self.state.apply_event(event);
+                }
+
+                _ = self.render_interval.tick() => {
+                    if self.state.dirty {
+                        self.terminal.draw(|frame| render(frame, &mut self.state))?;
+                        self.state.dirty = false;
+                    }
+                }
+            }
+        }
+        Ok(())
+    }
+}

crates/harness-tui/src/input.rs

@@ -0,0 +1,264 @@
+use crossterm::event::{Event, KeyCode, KeyEvent, KeyModifiers};
+use harness_app::EngineHandle;
+use harness_core::permission::PermissionReply;
+
+use crate::modal;
+use crate::state::{AppState, ModalState};
+
+/// Possible high-level actions produced by key input.
+#[derive(Debug)]
+pub enum InputAction {
+    None,
+    Submit { text: String },
+    Abort,
+    Quit,
+    LoadSessions,
+    NewSession,
+    SetModel(String),
+    SetAgent(String),
+    CloseModal,
+    ScrollUp(u16),
+    ScrollDown(u16),
+    PermissionReply(PermissionReply),
+    SessionPickerUp,
+    SessionPickerDown,
+    SessionPickerSelect,
+}
+
+/// Translate a crossterm event into an action and/or mutate `state` directly.
+pub fn handle_event(event: Event, state: &mut AppState, _engine: &EngineHandle) -> InputAction {
+    match event {
+        Event::Key(key) => {
+            let action = handle_key(key, state);
+            // Any keypress can change the input buffer or cursor; force a redraw so typed
+            // characters appear immediately rather than only when some other event sets dirty.
+            state.dirty = true;
+            action
+        }
+        Event::Resize(_, _) => {
+            state.dirty = true;
+            InputAction::None
+        }
+        Event::Mouse(_) | Event::FocusGained | Event::FocusLost | Event::Paste(_) => {
+            InputAction::None
+        }
+    }
+}
+
+fn handle_key(key: KeyEvent, state: &mut AppState) -> InputAction {
+    match &state.modal {
+        ModalState::Permission { .. } => handle_permission_key(key, state),
+        ModalState::SessionPicker { .. } => handle_session_picker_key(key, state),
+        ModalState::None => handle_normal_key(key, state),
+    }
+}
+
+fn handle_permission_key(key: KeyEvent, _state: &mut AppState) -> InputAction {
+    match key.code {
+        KeyCode::Char('y') | KeyCode::Char('Y') => {
+            InputAction::PermissionReply(PermissionReply::Once)
+        }
+        KeyCode::Char('a') | KeyCode::Char('A') => {
+            InputAction::PermissionReply(PermissionReply::Always)
+        }
+        KeyCode::Char('n') | KeyCode::Char('N') => {
+            InputAction::PermissionReply(PermissionReply::Reject)
+        }
+        // Esc rejects the pending request rather than merely hiding the modal — closing
+        // without a reply would leave the tool call blocked on its oneshot forever.
+        KeyCode::Esc => InputAction::PermissionReply(PermissionReply::Reject),
+        _ => InputAction::None,
+    }
+}
+
+fn handle_session_picker_key(key: KeyEvent, _state: &mut AppState) -> InputAction {
+    match key.code {
+        KeyCode::Up => InputAction::SessionPickerUp,
+        KeyCode::Down => InputAction::SessionPickerDown,
+        KeyCode::Enter => InputAction::SessionPickerSelect,
+        KeyCode::Esc => InputAction::CloseModal,
+        _ => InputAction::None,
+    }
+}
+
+fn handle_normal_key(key: KeyEvent, state: &mut AppState) -> InputAction {
+    let ctrl = key.modifiers.contains(KeyModifiers::CONTROL);
+
+    // Reset the double-Ctrl+C guard on any key that isn't another Ctrl+C.
+    if !(matches!(key.code, KeyCode::Char('c') | KeyCode::Char('C')) && ctrl) {
+        state.ctrl_c_pressed = false;
+    }
+
+    match key.code {
+        KeyCode::Enter => {
+            let text = state.input.lines().join("\n");
+            state.input = tui_textarea::TextArea::default();
+            state
+                .input
+                .set_cursor_line_style(ratatui::style::Style::default());
+            if let Some(action) = parse_slash_command(&text) {
+                action
+            } else {
+                InputAction::Submit { text }
+            }
+        }
+        KeyCode::Char('c') if ctrl => {
+            if state.ctrl_c_pressed {
+                InputAction::Quit
+            } else {
+                state.ctrl_c_pressed = true;
+                state.dirty = true;
+                InputAction::None
+            }
+        }
+        KeyCode::Esc => {
+            if state.running {
+                InputAction::Abort
+            } else {
+                InputAction::None
+            }
+        }
+        KeyCode::Char('s') if ctrl => InputAction::LoadSessions,
+        KeyCode::Up => {
+            let (row, _) = state.input.cursor();
+            if row == 0 {
+                InputAction::ScrollUp(3)
+            } else {
+                state.input.input(key);
+                InputAction::None
+            }
+        }
+        KeyCode::Down => {
+            let (row, _) = state.input.cursor();
+            let last_line = state.input.lines().len().saturating_sub(1);
+            if row == last_line {
+                InputAction::ScrollDown(3)
+            } else {
+                state.input.input(key);
+                InputAction::None
+            }
+        }
+        _ => {
+            state.input.input(key);
+            InputAction::None
+        }
+    }
+}
+
+fn parse_slash_command(text: &str) -> Option<InputAction> {
+    let trimmed = text.trim();
+    if !trimmed.starts_with('/') {
+        return None;
+    }
+    let mut parts = trimmed.split_whitespace();
+    let cmd = parts.next()?;
+    let rest: String = parts.collect::<Vec<_>>().join(" ");
+    match cmd {
+        "/new" => Some(InputAction::NewSession),
+        "/model" if !rest.is_empty() => Some(InputAction::SetModel(rest)),
+        "/agent" if !rest.is_empty() => Some(InputAction::SetAgent(rest)),
+        "/sessions" => Some(InputAction::LoadSessions),
+        "/quit" => Some(InputAction::Quit),
+        _ => None,
+    }
+}
+
+/// Apply an action that needs async engine calls. Non-async decisions are applied inline.
+pub async fn apply_action(action: InputAction, state: &mut AppState, engine: &EngineHandle) {
+    match action {
+        InputAction::Submit { text } => {
+            if let Some(session_id) = state.session_id.clone() {
+                if let Err(e) = engine.prompt(session_id, text, &state.model_ref).await {
+                    tracing::error!(error = %e, "prompt failed");
+                }
+            }
+        }
+        InputAction::Abort => {
+            if let Some(session_id) = state.session_id.clone() {
+                engine.abort(&session_id);
+            }
+        }
+        InputAction::LoadSessions => match engine.list_sessions().await {
+            Ok(sessions) => modal::open_session_picker(state, sessions),
+            Err(e) => tracing::error!(error = %e, "failed to list sessions"),
+        },
+        InputAction::NewSession => match engine.new_session(&state.agent, &state.model_ref).await {
+            Ok(id) => {
+                state.session_id = Some(id);
+                state.messages.clear();
+                state.scroll_offset = 0;
+                state.session_cost = 0.0;
+                state.session_tokens = 0;
+                state.dirty = true;
+            }
+            Err(e) => tracing::error!(error = %e, "failed to create session"),
+        },
+        InputAction::SetModel(model_ref) => {
+            state.model_ref = model_ref;
+            state.dirty = true;
+        }
+        InputAction::SetAgent(agent) => {
+            state.agent = agent;
+            state.dirty = true;
+        }
+        InputAction::Quit => state.quit = true,
+        InputAction::CloseModal => state.close_modal(),
+        InputAction::ScrollUp(n) => state.scroll_up(n),
+        InputAction::ScrollDown(n) => state.scroll_down(n),
+        InputAction::PermissionReply(reply) => {
+            modal::resolve_permission(state, engine, reply);
+        }
+        InputAction::SessionPickerUp => {
+            if let ModalState::SessionPicker {
+                selected,
+                sessions: _,
+            } = &mut state.modal
+            {
+                *selected = selected.saturating_sub(1);
+            }
+            state.dirty = true;
+        }
+        InputAction::SessionPickerDown => {
+            if let ModalState::SessionPicker { selected, sessions } = &mut state.modal {
+                if *selected + 1 < sessions.len() {
+                    *selected += 1;
+                }
+            }
+            state.dirty = true;
+        }
+        InputAction::SessionPickerSelect => {
+            if let ModalState::SessionPicker { sessions, selected } = &state.modal {
+                if let Some(session) = sessions.get(*selected).cloned() {
+                    if let Err(e) = modal::select_session(state, engine, session).await {
+                        tracing::error!(error = %e, "failed to load session");
+                    }
+                }
+            }
+        }
+        InputAction::None => {}
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::state::AppState;
+    use crossterm::event::KeyEvent;
+
+    #[tokio::test]
+    async fn typing_marks_frame_dirty_and_updates_input() {
+        let engine = EngineHandle::init_in_memory(std::env::temp_dir()).unwrap();
+        let mut state = AppState::new("anthropic/claude".into(), "orchestrator".into());
+        state.dirty = false;
+
+        let action = handle_event(
+            Event::Key(KeyEvent::from(KeyCode::Char('x'))),
+            &mut state,
+            &engine,
+        );
+
+        assert!(matches!(action, InputAction::None));
+        assert!(state.dirty, "a keystroke must request a redraw");
+        assert_eq!(state.input.lines().join("\n"), "x");
+    }
+}

crates/harness-tui/src/main.rs

@@ -1,3 +1,156 @@
-fn main() {
+mod app;
-    println!("harness {}", env!("CARGO_PKG_VERSION"));
+mod input;
+mod markdown;
+mod modal;
+mod render;
+mod state;
+mod terminal;
+
+use std::path::PathBuf;
+
+use harness_core::event::RunOutcome;
+
+const DEFAULT_MODEL: &str = "anthropic/claude-sonnet-4-5";
+
+fn print_usage() {
+    eprintln!("usage: harness [run -p \"<prompt>\" [-m provider/model]] | [tui]");
+}
+
+fn parse_run_args(args: &[String]) -> Option<(String, Option<String>)> {
+    let mut prompt = None;
+    let mut model = None;
+    let mut i = 0;
+    while i < args.len() {
+        match args[i].as_str() {
+            "-p" | "--prompt" => {
+                prompt = args.get(i + 1).cloned();
+                i += 2;
+            }
+            "-m" | "--model" => {
+                model = args.get(i + 1).cloned();
+                i += 2;
+            }
+            _ => i += 1,
+        }
+    }
+    prompt.map(|p| (p, model))
+}
+
+async fn run_headless(args: &[String]) -> i32 {
+    let Some((prompt, model_arg)) = parse_run_args(args) else {
+        print_usage();
+        return 2;
+    };
+
+    let cwd = match std::env::current_dir() {
+        Ok(cwd) => cwd,
+        Err(e) => {
+            eprintln!("error: could not read cwd: {e}");
+            return 1;
+        }
+    };
+
+    let app = match harness_app::App::init(cwd) {
+        Ok(app) => app,
+        Err(e) => {
+            eprintln!("error: {e}");
+            return 1;
+        }
+    };
+
+    let model_ref = model_arg
+        .or_else(|| app.config().model.clone())
+        .unwrap_or_else(|| DEFAULT_MODEL.to_string());
+
+    match app.run_prompt(prompt, &model_ref).await {
+        Ok((RunOutcome::Stopped, session_id)) => {
+            let text = app.final_text(&session_id).await.unwrap_or_default();
+            println!("{text}");
+            0
+        }
+        Ok((RunOutcome::Aborted, _)) => {
+            eprintln!("run aborted");
+            1
+        }
+        Ok((RunOutcome::Errored { message }, _)) => {
+            eprintln!("error: {message}");
+            1
+        }
+        Err(e) => {
+            eprintln!("error: {e}");
+            1
+        }
+    }
+}
+
+fn log_dir() -> PathBuf {
+    dirs::data_dir()
+        .unwrap_or_else(std::env::temp_dir)
+        .join("ai-harness")
+        .join("log")
+}
+
+fn setup_tracing() -> anyhow::Result<tracing_appender::non_blocking::WorkerGuard> {
+    let log_dir = log_dir();
+    std::fs::create_dir_all(&log_dir)?;
+    let file_appender = tracing_appender::rolling::daily(log_dir, "harness-tui.log");
+    let (non_blocking, guard) = tracing_appender::non_blocking(file_appender);
+    tracing_subscriber::fmt()
+        .with_writer(non_blocking)
+        .with_env_filter(tracing_subscriber::EnvFilter::from_default_env())
+        .init();
+    Ok(guard)
+}
+
+async fn run_tui() -> i32 {
+    let cwd = match std::env::current_dir() {
+        Ok(cwd) => cwd,
+        Err(e) => {
+            eprintln!("error: could not read cwd: {e}");
+            return 1;
+        }
+    };
+
+    let _guard = match setup_tracing() {
+        Ok(guard) => guard,
+        Err(e) => {
+            eprintln!("error: could not initialize logging: {e}");
+            return 1;
+        }
+    };
+
+    let mut app = match crate::app::App::new(cwd).await {
+        Ok(app) => app,
+        Err(e) => {
+            tracing::error!(error = %e, "failed to start TUI");
+            eprintln!("error: {e}");
+            return 1;
+        }
+    };
+
+    if let Err(e) = app.run().await {
+        tracing::error!(error = %e, "TUI error");
+        eprintln!("error: {e}");
+        return 1;
+    }
+    0
+}
+
+#[tokio::main(flavor = "current_thread")]
+async fn main() {
+    let args: Vec<String> = std::env::args().skip(1).collect();
+    let exit_code = match args.first().map(String::as_str) {
+        Some("run") => run_headless(&args[1..]).await,
+        Some("tui") | None => run_tui().await,
+        Some("help") | Some("--help") | Some("-h") => {
+            print_usage();
+            0
+        }
+        Some(cmd) => {
+            eprintln!("unknown command: {cmd}");
+            print_usage();
+            2
+        }
+    };
+    std::process::exit(exit_code);
 }

crates/harness-tui/src/markdown.rs

@@ -0,0 +1,255 @@
+use pulldown_cmark::{CodeBlockKind, Event, HeadingLevel, Parser, Tag, TagEnd};
+use ratatui::style::{Color, Modifier, Style};
+use ratatui::text::{Line, Span};
+
+/// Render a markdown string into wrapped ratatui lines.
+pub fn render_markdown(text: &str, width: u16) -> Vec<Line<'static>> {
+    let mut lines: Vec<Line<'static>> = Vec::new();
+    let mut current_spans: Vec<Span<'static>> = Vec::new();
+    let mut current_text = String::new();
+
+    let mut in_bold = false;
+    let mut in_italic = false;
+    let mut in_code_block = false;
+    let mut code_block_language = String::new();
+    let mut list_stack: Vec<u64> = Vec::new();
+
+    let flush = |current: &mut String, spans: &mut Vec<Span<'static>>, bold, italic| {
+        if current.is_empty() {
+            return;
+        }
+        let style = base_style(bold, italic);
+        let span = Span::styled(std::mem::take(current), style);
+        spans.push(span);
+    };
+
+    for event in Parser::new(text) {
+        match event {
+            Event::Start(tag) => match tag {
+                Tag::Paragraph => {
+                    if !lines.is_empty() && !current_text.is_empty() {
+                        flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                        wrap_spans(&mut lines, &mut current_spans, width);
+                    }
+                }
+                Tag::Heading { level, .. } => {
+                    let level_num = heading_level_from(level);
+                    current_text.push_str(&"#".repeat(level_num as usize));
+                    current_text.push(' ');
+                }
+                Tag::Strong => {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    in_bold = true;
+                }
+                Tag::Emphasis => {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    in_italic = true;
+                }
+                Tag::List(start) => {
+                    list_stack.push(start.unwrap_or(1));
+                }
+                Tag::Item => {
+                    let prefix = if let Some(n) = list_stack.last_mut() {
+                        let p = format!("{n}. ");
+                        *n += 1;
+                        p
+                    } else {
+                        "• ".to_string()
+                    };
+                    current_text.push_str(&prefix);
+                }
+                Tag::CodeBlock(lang) => {
+                    in_code_block = true;
+                    code_block_language = match lang {
+                        CodeBlockKind::Fenced(name) => name.to_string(),
+                        CodeBlockKind::Indented => String::new(),
+                    };
+                    if !current_text.is_empty() || !current_spans.is_empty() {
+                        flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                        wrap_spans(&mut lines, &mut current_spans, width);
+                    }
+                    let border = if code_block_language.is_empty() {
+                        "┌────".to_string()
+                    } else {
+                        format!("┌──── {code_block_language}")
+                    };
+                    lines.push(Line::from(Span::styled(border, code_style())));
+                }
+                _ => {}
+            },
+            Event::End(tag_end) => match tag_end {
+                TagEnd::Heading(_) => {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    let mut spans = std::mem::take(&mut current_spans);
+                    for span in &mut spans {
+                        span.style = span.style.add_modifier(Modifier::BOLD);
+                    }
+                    wrap_spans(&mut lines, &mut spans, width);
+                }
+                TagEnd::Paragraph => {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    wrap_spans(&mut lines, &mut current_spans, width);
+                }
+                TagEnd::Strong => {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    in_bold = false;
+                }
+                TagEnd::Emphasis => {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    in_italic = false;
+                }
+                TagEnd::List(_) => {
+                    list_stack.pop();
+                }
+                TagEnd::CodeBlock => {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    if !current_spans.is_empty() {
+                        for line in wrap_spans_to_lines(&mut current_spans, width) {
+                            lines.push(prefix_code_block_line(line));
+                        }
+                    }
+                    lines.push(Line::from(Span::styled("└────", code_style())));
+                    in_code_block = false;
+                    code_block_language.clear();
+                }
+                _ => {}
+            },
+            Event::Text(t) => {
+                if in_code_block {
+                    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                    current_spans.push(Span::styled(t.to_string(), code_style()));
+                } else {
+                    current_text.push_str(&t);
+                }
+            }
+            Event::Code(c) => {
+                flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                current_spans.push(Span::styled(c.to_string(), code_style()));
+            }
+            Event::Html(h) | Event::InlineHtml(h) => {
+                current_text.push_str(&h);
+            }
+            Event::SoftBreak | Event::HardBreak => {
+                flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                wrap_spans(&mut lines, &mut current_spans, width);
+            }
+            Event::Rule => {
+                flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+                wrap_spans(&mut lines, &mut current_spans, width);
+                lines.push(Line::from("─".repeat(width as usize)));
+            }
+            _ => {}
+        }
+    }
+
+    flush(&mut current_text, &mut current_spans, in_bold, in_italic);
+    if in_code_block {
+        for line in wrap_spans_to_lines(&mut current_spans, width) {
+            lines.push(prefix_code_block_line(line));
+        }
+        lines.push(Line::from(Span::styled("└────", code_style())));
+    } else {
+        wrap_spans(&mut lines, &mut current_spans, width);
+    }
+
+    if lines.is_empty() {
+        lines.push(Line::default());
+    }
+    lines
+}
+
+fn heading_level_from(level: HeadingLevel) -> u8 {
+    match level {
+        HeadingLevel::H1 => 1,
+        HeadingLevel::H2 => 2,
+        HeadingLevel::H3 => 3,
+        HeadingLevel::H4 => 4,
+        HeadingLevel::H5 => 5,
+        HeadingLevel::H6 => 6,
+    }
+}
+
+fn base_style(bold: bool, italic: bool) -> Style {
+    let mut style = Style::new();
+    if bold {
+        style = style.add_modifier(Modifier::BOLD);
+    }
+    if italic {
+        style = style.add_modifier(Modifier::ITALIC);
+    }
+    style
+}
+
+fn code_style() -> Style {
+    Style::new().fg(Color::Yellow)
+}
+
+fn prefix_code_block_line(line: Line<'static>) -> Line<'static> {
+    let mut spans = vec![Span::styled("│ ", code_style())];
+    spans.extend(line.spans);
+    Line::from(spans)
+}
+
+/// Flush accumulated spans into `lines`, wrapping to `width`.
+fn wrap_spans(lines: &mut Vec<Line<'static>>, spans: &mut Vec<Span<'static>>, width: u16) {
+    if spans.is_empty() {
+        return;
+    }
+    for line in wrap_spans_to_lines(spans, width) {
+        lines.push(line);
+    }
+    spans.clear();
+}
+
+fn wrap_spans_to_lines(spans: &mut Vec<Span<'static>>, width: u16) -> Vec<Line<'static>> {
+    let width = width.max(1) as usize;
+    let mut out: Vec<Line<'static>> = Vec::new();
+    let mut current_line: Vec<Span<'static>> = Vec::new();
+    let mut current_width = 0usize;
+
+    for span in spans.drain(..) {
+        for word in span.content.split(' ') {
+            let word_width = word.chars().count();
+            let sep = if current_width == 0 { 0 } else { 1 };
+            if current_width + sep + word_width > width && current_width > 0 {
+                out.push(Line::from(std::mem::take(&mut current_line)));
+                current_width = 0;
+            }
+            if current_width > 0 {
+                current_line.push(Span::styled(" ", span.style));
+                current_width += 1;
+            }
+            current_line.push(Span::styled(word.to_string(), span.style));
+            current_width += word_width;
+        }
+    }
+
+    if !current_line.is_empty() {
+        out.push(Line::from(current_line));
+    }
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn renders_heading_and_bold() {
+        let lines = render_markdown("# Hello\n\n**bold** text", 40);
+        assert!(!lines.is_empty());
+        let first: String = lines[0]
+            .spans
+            .iter()
+            .map(|s| s.content.to_string())
+            .collect();
+        assert!(first.contains("# Hello"));
+    }
+
+    #[test]
+    fn wraps_long_paragraphs() {
+        let text = "a ".repeat(50);
+        let lines = render_markdown(&text, 20);
+        assert!(lines.len() > 1);
+    }
+}

crates/harness-tui/src/modal.rs

@@ -0,0 +1,38 @@
+//! Modal helpers for permission dialogs and the session picker.
+//!
+//! Rendering lives in `render.rs`; key handling lives in `input.rs`. This module
+//! provides shared utilities for modal state transitions.
+
+use harness_app::EngineHandle;
+use harness_core::permission::PermissionReply;
+use harness_core::types::Session;
+
+use crate::state::{AppState, ModalState};
+
+/// Resolve the current permission request with `reply`.
+pub fn resolve_permission(state: &mut AppState, engine: &EngineHandle, reply: PermissionReply) {
+    state.handle_permission_reply(reply, engine);
+}
+
+/// Open the session picker with the supplied sessions.
+pub fn open_session_picker(state: &mut AppState, sessions: Vec<Session>) {
+    state.open_session_picker(sessions);
+}
+
+/// Load a session's messages and parts into state.
+pub async fn select_session(
+    state: &mut AppState,
+    engine: &EngineHandle,
+    session: Session,
+) -> Result<(), harness_app::AppError> {
+    let session_id = session.id.clone();
+    let messages = engine.session_messages(session_id).await?;
+    let mut all_parts = Vec::new();
+    for message in &messages {
+        let parts = engine.message_parts(message.id.clone()).await?;
+        all_parts.extend(parts);
+    }
+    state.set_session(session, messages, all_parts);
+    state.modal = ModalState::None;
+    Ok(())
+}

crates/harness-tui/src/render.rs

@@ -0,0 +1,511 @@
+use ratatui::layout::{Constraint, Direction, Layout, Rect};
+use ratatui::style::{Color, Modifier, Style};
+use ratatui::text::{Line, Span};
+use ratatui::widgets::{Block, Borders, Clear, Paragraph, Wrap};
+use ratatui::Frame;
+
+use crate::markdown::render_markdown;
+use crate::state::{AppState, ModalState, PartView, ToolStateView};
+use harness_core::types::Role;
+
+/// Render the full UI into the terminal frame.
+pub fn render(frame: &mut Frame, state: &mut AppState) {
+    let area = frame.area();
+    let chunks = Layout::default()
+        .direction(Direction::Vertical)
+        .constraints([
+            Constraint::Length(1),
+            Constraint::Min(3),
+            Constraint::Length(3),
+            Constraint::Length(1),
+        ])
+        .split(area);
+
+    render_header(frame, state, chunks[0]);
+    render_chat(frame, state, chunks[1]);
+    render_input(frame, state, chunks[2]);
+    render_status(frame, state, chunks[3]);
+
+    match &state.modal {
+        ModalState::Permission { requests } => {
+            render_permission_modal(frame, &requests[0], area);
+        }
+        ModalState::SessionPicker { sessions, selected } => {
+            render_session_picker(frame, sessions, *selected, area);
+        }
+        ModalState::None => {}
+    }
+}
+
+fn render_header(frame: &mut Frame, state: &AppState, area: Rect) {
+    let title = if state.session_title.is_empty() {
+        "new session"
+    } else {
+        state.session_title.as_str()
+    };
+    let text = format!("{} · {} · {}", title, state.agent, state.model_ref);
+    let paragraph = Paragraph::new(text)
+        .style(Style::new().add_modifier(Modifier::BOLD))
+        .alignment(ratatui::layout::Alignment::Center);
+    frame.render_widget(paragraph, area);
+}
+
+fn render_chat(frame: &mut Frame, state: &mut AppState, area: Rect) {
+    let effective_width = area.width.saturating_sub(4).max(1);
+    // Cached lines are wrapped to a specific width, so a resize must drop them.
+    if state.render_width != effective_width {
+        state.invalidate_caches();
+        state.render_width = effective_width;
+    }
+
+    let mut lines: Vec<Line<'static>> = Vec::new();
+    for message in &mut state.messages {
+        let role_style = match message.role {
+            Role::User => Style::new().fg(Color::Cyan),
+            Role::Assistant => Style::new(),
+        };
+        let prefix = match message.role {
+            Role::User => "> ",
+            Role::Assistant => "",
+        };
+        let mut first = true;
+        for part in &mut message.parts {
+            let part_lines = render_part(part, effective_width, role_style, prefix, first);
+            lines.extend(part_lines);
+            first = false;
+        }
+        lines.push(Line::default());
+    }
+
+    let paragraph = Paragraph::new(lines)
+        .block(Block::default().borders(Borders::ALL).title(" chat "))
+        .scroll((state.scroll_offset, 0));
+    frame.render_widget(paragraph, area);
+}
+
+fn render_part(
+    part: &mut PartView,
+    effective_width: u16,
+    role_style: Style,
+    prefix: &'static str,
+    first: bool,
+) -> Vec<Line<'static>> {
+    match part {
+        PartView::Text {
+            text, cached_lines, ..
+        } => {
+            // Cache the prefix-free markdown render; the `> ` prefix is cheap to re-apply
+            // to the clone each frame and would otherwise poison a shared cache.
+            let base = cached_lines.get_or_insert_with(|| render_markdown(text, effective_width));
+            let mut rendered = base.clone();
+            if first && !prefix.is_empty() {
+                if let Some(first_line) = rendered.first_mut() {
+                    let mut spans = vec![Span::styled(prefix, role_style)];
+                    spans.extend(first_line.spans.clone());
+                    *first_line = Line::from(spans);
+                }
+            }
+            rendered
+        }
+        PartView::Reasoning { text, .. } => {
+            let style = Style::new()
+                .fg(Color::DarkGray)
+                .add_modifier(Modifier::ITALIC);
+            render_markdown(text, effective_width)
+                .into_iter()
+                .map(|line| {
+                    let mut spans = vec![Span::styled("🧠 ", style)];
+                    spans.extend(line.spans);
+                    Line::from(spans)
+                })
+                .collect()
+        }
+        PartView::Tool {
+            name,
+            state,
+            cached_lines,
+            ..
+        } => {
+            if let Some(cached) = cached_lines {
+                return cached.clone();
+            }
+            let status = state.status_label();
+            let icon = match state {
+                ToolStateView::Pending { .. } => "⏳",
+                ToolStateView::Running { .. } => "⚙",
+                ToolStateView::Completed { .. } => "✓",
+                ToolStateView::Error { .. } => "✗",
+            };
+            let title = format!("{icon} {name} — {status}");
+            let mut lines = vec![Line::from(title)];
+            match state {
+                ToolStateView::Pending { partial_input } => {
+                    lines.extend(render_markdown(partial_input, effective_width));
+                }
+                ToolStateView::Completed { output, .. } => {
+                    lines.extend(render_markdown(output, effective_width));
+                }
+                ToolStateView::Error { error } => {
+                    lines.extend(render_markdown(error, effective_width));
+                }
+                ToolStateView::Running { .. } => {}
+            }
+            *cached_lines = Some(lines.clone());
+            lines
+        }
+        PartView::StepFinish { .. } => vec![Line::from("─── step ───")],
+    }
+}
+
+fn render_input(frame: &mut Frame, state: &mut AppState, area: Rect) {
+    let block = Block::default().borders(Borders::ALL).title(" input ");
+    state.input.set_block(block);
+    frame.render_widget(&state.input, area);
+}
+
+fn render_status(frame: &mut Frame, state: &AppState, area: Rect) {
+    let spinner = if state.running { "⠋ " } else { "" };
+    let status = if state.running { "running" } else { "idle" };
+    let hints = if state.ctrl_c_pressed {
+        "Press Ctrl+C again to quit"
+    } else {
+        "Ctrl+S: sessions | Esc: abort | Ctrl+C: quit"
+    };
+    let usage = format!(
+        "{} · ${:.4}",
+        format_tokens(state.session_tokens),
+        state.session_cost
+    );
+    let text = format!("{spinner}{status} · {usage} · {hints}");
+    let paragraph = Paragraph::new(text).style(Style::new().fg(Color::Gray));
+    frame.render_widget(paragraph, area);
+}
+
+/// Compact token count: `1234` → `1.2k`, `2_000_000` → `2.0M`.
+fn format_tokens(n: u64) -> String {
+    match n {
+        0..=999 => format!("{n} tok"),
+        1_000..=999_999 => format!("{:.1}k tok", n as f64 / 1_000.0),
+        _ => format!("{:.1}M tok", n as f64 / 1_000_000.0),
+    }
+}
+
+fn render_permission_modal(
+    frame: &mut Frame,
+    request: &harness_core::event::PermissionRequest,
+    area: Rect,
+) {
+    let popup = centered_rect(60, 60, area);
+    frame.render_widget(Clear, popup);
+
+    let text = vec![
+        Line::from("Permission required").style(Style::new().add_modifier(Modifier::BOLD)),
+        Line::default(),
+        Line::from(format!("permission: {}", request.permission)),
+        Line::from(format!("pattern: {}", request.pattern)),
+        Line::default(),
+        Line::from("y: allow once  a: allow always  n: reject"),
+    ];
+
+    let block = Block::default().borders(Borders::ALL).title(" permission ");
+    let paragraph = Paragraph::new(text).block(block).wrap(Wrap { trim: true });
+    frame.render_widget(paragraph, popup);
+}
+
+fn render_session_picker(
+    frame: &mut Frame,
+    sessions: &[harness_core::types::Session],
+    selected: usize,
+    area: Rect,
+) {
+    let popup = centered_rect(60, 60, area);
+    frame.render_widget(Clear, popup);
+
+    let mut text: Vec<Line<'static>> =
+        vec![Line::from("Select session").style(Style::new().add_modifier(Modifier::BOLD))];
+    for (i, session) in sessions.iter().enumerate() {
+        let marker = if i == selected { "> " } else { "  " };
+        let line = format!(
+            "{}{} · {} · {}/{}",
+            marker, session.id, session.agent, session.model.provider_id, session.model.model_id
+        );
+        let style = if i == selected {
+            Style::new().bg(Color::Blue).fg(Color::White)
+        } else {
+            Style::new()
+        };
+        text.push(Line::from(Span::styled(line, style)));
+    }
+
+    let block = Block::default().borders(Borders::ALL).title(" sessions ");
+    let paragraph = Paragraph::new(text).block(block).wrap(Wrap { trim: false });
+    frame.render_widget(paragraph, popup);
+}
+
+fn centered_rect(percent_x: u16, percent_y: u16, r: Rect) -> Rect {
+    let popup_layout = Layout::default()
+        .direction(Direction::Vertical)
+        .constraints([
+            Constraint::Percentage((100 - percent_y) / 2),
+            Constraint::Percentage(percent_y),
+            Constraint::Percentage((100 - percent_y) / 2),
+        ])
+        .split(r);
+
+    Layout::default()
+        .direction(Direction::Horizontal)
+        .constraints([
+            Constraint::Percentage((100 - percent_x) / 2),
+            Constraint::Percentage(percent_x),
+            Constraint::Percentage((100 - percent_x) / 2),
+        ])
+        .split(popup_layout[1])[1]
+}
+
+#[cfg(test)]
+mod tests {
+    use ratatui::backend::TestBackend;
+    use ratatui::Terminal;
+
+    use super::*;
+    use crate::state::{AppState, MessageView, PartView, ToolStateView};
+    use harness_core::event::PermissionRequest;
+    use harness_core::types::{MessageId, ModelRef, PartId, Role, Session, SessionId};
+
+    fn buffer_to_string(backend: &TestBackend) -> String {
+        let buffer = backend.buffer();
+        let area = buffer.area;
+        let mut result = String::new();
+        for y in 0..area.height {
+            for x in 0..area.width {
+                result.push_str(buffer[(x, y)].symbol());
+            }
+            while result.ends_with(' ') {
+                result.pop();
+            }
+            result.push('\n');
+        }
+        result
+    }
+
+    #[test]
+    fn snapshot_empty_session() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new(
+            "anthropic/claude-sonnet-4-5".to_string(),
+            "orchestrator".to_string(),
+        );
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        insta::assert_snapshot!(buffer_to_string(terminal.backend()));
+    }
+
+    #[test]
+    fn snapshot_user_and_assistant_messages() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new(
+            "anthropic/claude-sonnet-4-5".to_string(),
+            "orchestrator".to_string(),
+        );
+        state.session_id = Some(SessionId("ses_test_001".to_string()));
+        state.messages.push(MessageView {
+            id: MessageId("msg_test_001".to_string()),
+            role: Role::User,
+            parts: vec![PartView::Text {
+                id: PartId("prt_test_001".to_string()),
+                text: "Hello, can you read foo.txt?".to_string(),
+                cached_lines: None,
+            }],
+        });
+        state.messages.push(MessageView {
+            id: MessageId("msg_test_002".to_string()),
+            role: Role::Assistant,
+            parts: vec![PartView::Text {
+                id: PartId("prt_test_002".to_string()),
+                text: "I'll read that file for you.\n\n```rust\nlet x = 42;\n```".to_string(),
+                cached_lines: None,
+            }],
+        });
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        insta::assert_snapshot!(buffer_to_string(terminal.backend()));
+    }
+
+    #[test]
+    fn snapshot_tool_card_completed() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new(
+            "anthropic/claude-sonnet-4-5".to_string(),
+            "orchestrator".to_string(),
+        );
+        state.session_id = Some(SessionId("ses_test_001".to_string()));
+        state.messages.push(MessageView {
+            id: MessageId("msg_test_003".to_string()),
+            role: Role::Assistant,
+            parts: vec![PartView::Tool {
+                id: PartId("prt_test_003".to_string()),
+                name: "read".to_string(),
+                state: ToolStateView::Completed {
+                    title: "read foo.txt".to_string(),
+                    output: "file contents here".to_string(),
+                },
+                cached_lines: None,
+            }],
+        });
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        insta::assert_snapshot!(buffer_to_string(terminal.backend()));
+    }
+
+    #[test]
+    fn snapshot_tool_card_running() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new(
+            "anthropic/claude-sonnet-4-5".to_string(),
+            "orchestrator".to_string(),
+        );
+        state.session_id = Some(SessionId("ses_test_001".to_string()));
+        state.messages.push(MessageView {
+            id: MessageId("msg_test_004".to_string()),
+            role: Role::Assistant,
+            parts: vec![PartView::Tool {
+                id: PartId("prt_test_004".to_string()),
+                name: "read".to_string(),
+                state: ToolStateView::Running {
+                    title: Some("read foo.txt".to_string()),
+                },
+                cached_lines: None,
+            }],
+        });
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        insta::assert_snapshot!(buffer_to_string(terminal.backend()));
+    }
+
+    #[test]
+    fn snapshot_tool_card_error() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new(
+            "anthropic/claude-sonnet-4-5".to_string(),
+            "orchestrator".to_string(),
+        );
+        state.session_id = Some(SessionId("ses_test_001".to_string()));
+        state.messages.push(MessageView {
+            id: MessageId("msg_test_005".to_string()),
+            role: Role::Assistant,
+            parts: vec![PartView::Tool {
+                id: PartId("prt_test_005".to_string()),
+                name: "read".to_string(),
+                state: ToolStateView::Error {
+                    error: "file not found".to_string(),
+                },
+                cached_lines: None,
+            }],
+        });
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        insta::assert_snapshot!(buffer_to_string(terminal.backend()));
+    }
+
+    #[test]
+    fn snapshot_permission_modal() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new(
+            "anthropic/claude-sonnet-4-5".to_string(),
+            "orchestrator".to_string(),
+        );
+        let request = PermissionRequest {
+            id: "test".to_string(),
+            session_id: SessionId("ses_test_001".to_string()),
+            permission: "bash".to_string(),
+            pattern: "rm -rf /".to_string(),
+            always_pattern: "rm *".to_string(),
+            metadata: serde_json::Value::Null,
+        };
+        state.modal = ModalState::Permission {
+            requests: vec![request],
+        };
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        insta::assert_snapshot!(buffer_to_string(terminal.backend()));
+    }
+
+    #[test]
+    fn snapshot_session_picker_modal() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new(
+            "anthropic/claude-sonnet-4-5".to_string(),
+            "orchestrator".to_string(),
+        );
+        let sessions = vec![
+            Session {
+                id: SessionId("ses_test_001".to_string()),
+                parent_id: None,
+                depth: 0,
+                title: "Session One".to_string(),
+                agent: "orchestrator".to_string(),
+                model: ModelRef::new("anthropic", "claude-sonnet-4-5"),
+                usage: harness_core::types::TokenUsage::default(),
+                cost: 0.0,
+                extra_rules: Vec::new(),
+                created_at: 1,
+                updated_at: 1,
+            },
+            Session {
+                id: SessionId("ses_test_002".to_string()),
+                parent_id: None,
+                depth: 0,
+                title: "Session Two".to_string(),
+                agent: "coder".to_string(),
+                model: ModelRef::new("openai", "gpt-4o"),
+                usage: harness_core::types::TokenUsage::default(),
+                cost: 0.0,
+                extra_rules: Vec::new(),
+                created_at: 2,
+                updated_at: 2,
+            },
+            Session {
+                id: SessionId("ses_test_003".to_string()),
+                parent_id: None,
+                depth: 0,
+                title: "Session Three".to_string(),
+                agent: "reviewer".to_string(),
+                model: ModelRef::new("google", "gemini-2.5"),
+                usage: harness_core::types::TokenUsage::default(),
+                cost: 0.0,
+                extra_rules: Vec::new(),
+                created_at: 3,
+                updated_at: 3,
+            },
+        ];
+        state.modal = ModalState::SessionPicker {
+            sessions,
+            selected: 1,
+        };
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        insta::assert_snapshot!(buffer_to_string(terminal.backend()));
+    }
+
+    #[test]
+    fn render_empty_state_produces_frame() {
+        let backend = TestBackend::new(80, 24);
+        let mut terminal = Terminal::new(backend).unwrap();
+        let mut state = AppState::new("anthropic/claude".to_string(), "orchestrator".to_string());
+        terminal.draw(|frame| render(frame, &mut state)).unwrap();
+        let buffer = terminal.backend().buffer().clone();
+        assert_eq!(buffer.area.width, 80);
+        assert_eq!(buffer.area.height, 24);
+        // Header should contain the agent/model line.
+        let header_row: String = buffer
+            .content
+            .chunks(80)
+            .next()
+            .unwrap()
+            .iter()
+            .map(|c| c.symbol())
+            .collect();
+        assert!(header_row.contains("orchestrator"));
+        assert!(header_row.contains("anthropic/claude"));
+    }
+}

crates/harness-tui/src/snapshots/harness__render__tests__snapshot_empty_session.snap

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 300
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit

crates/harness-tui/src/snapshots/harness__render__tests__snapshot_permission_modal.snap

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 430
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│               ┌ permission ──────────────────────────────────┐               │
+│               │Permission required                           │               │
+│               │                                              │               │
+│               │permission: bash                              │               │
+│               │pattern: rm -rf /                             │               │
+│               │                                              │               │
+│               │y: allow once  a: allow always  n: reject     │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               └──────────────────────────────────────────────┘               │
+└──────────────────────────────────────────────────────────────────────────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit

crates/harness-tui/src/snapshots/harness__render__tests__snapshot_session_picker_modal.snap

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 487
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│               ┌ sessions ────────────────────────────────────┐               │
+│               │Select session                                │               │
+│               │  ses_test_001 · orchestrator ·               │               │
+│               │anthropic/claude-sonnet-4-5                   │               │
+│               │> ses_test_002 · coder · openai/gpt-4o        │               │
+│               │  ses_test_003 · reviewer · google/gemini-2.5 │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               │                                              │               │
+│               └──────────────────────────────────────────────┘               │
+└──────────────────────────────────────────────────────────────────────────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit

crates/harness-tui/src/snapshots/harness__render__tests__snapshot_tool_card_completed.snap

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 357
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│✓ read — read foo.txt                                                         │
+│file contents here                                                            │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit

crates/harness-tui/src/snapshots/harness__render__tests__snapshot_tool_card_error.snap

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 407
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│✗ read — error: file not found                                                │
+│file not found                                                                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit

crates/harness-tui/src/snapshots/harness__render__tests__snapshot_tool_card_running.snap

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 382
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│⚙ read — read foo.txt                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit

crates/harness-tui/src/snapshots/harness__render__tests__snapshot_user_and_assistant_messages.snap

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 331
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│> Hello, can you read foo.txt?                                                │
+│                                                                              │
+│I'll read that file for you.                                                  │
+│┌──── rust                                                                    │
+││ let x = 42;                                                                 │
+│└────                                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit

crates/harness-tui/src/state.rs

@@ -0,0 +1,406 @@
+use harness_app::EngineHandle;
+use harness_core::event::{AppEvent, PermissionRequest, RunOutcome};
+use harness_core::permission::PermissionReply;
+use harness_core::types::{Message, Part, PartBody, PartId, Role, Session, SessionId, ToolState};
+use ratatui::text::Line;
+
+/// Cached view of a message in the chat viewport.
+#[derive(Debug)]
+pub struct MessageView {
+    pub id: harness_core::types::MessageId,
+    pub role: Role,
+    pub parts: Vec<PartView>,
+}
+
+/// Cached view of a part. The optional `cached_lines` field stores a pre-rendered markdown
+/// representation; it is invalidated whenever the underlying text changes.
+#[derive(Debug)]
+pub enum PartView {
+    Text {
+        id: PartId,
+        text: String,
+        cached_lines: Option<Vec<Line<'static>>>,
+    },
+    Reasoning {
+        id: PartId,
+        text: String,
+    },
+    Tool {
+        id: PartId,
+        name: String,
+        state: ToolStateView,
+        cached_lines: Option<Vec<Line<'static>>>,
+    },
+    StepFinish {
+        id: PartId,
+    },
+}
+
+/// Simplified tool state for rendering.
+#[derive(Debug)]
+pub enum ToolStateView {
+    Pending { partial_input: String },
+    Running { title: Option<String> },
+    Completed { title: String, output: String },
+    Error { error: String },
+}
+
+impl ToolStateView {
+    fn from_tool_state(state: &ToolState) -> Self {
+        match state {
+            ToolState::Pending { partial_input } => Self::Pending {
+                partial_input: partial_input.clone(),
+            },
+            ToolState::Running { title, .. } => Self::Running {
+                title: title.clone(),
+            },
+            ToolState::Completed { title, output, .. } => Self::Completed {
+                title: title.clone(),
+                output: output.clone(),
+            },
+            ToolState::Error { error, .. } => Self::Error {
+                error: error.clone(),
+            },
+        }
+    }
+
+    pub fn status_label(&self) -> String {
+        match self {
+            Self::Pending { .. } => "pending".to_string(),
+            Self::Running { title } => title.clone().unwrap_or_else(|| "running".to_string()),
+            Self::Completed { title, .. } => title.clone(),
+            Self::Error { error } => format!("error: {error}"),
+        }
+    }
+}
+
+/// Active modal overlay state.
+#[derive(Debug, Default)]
+pub enum ModalState {
+    #[default]
+    None,
+    Permission {
+        requests: Vec<PermissionRequest>,
+    },
+    SessionPicker {
+        sessions: Vec<Session>,
+        selected: usize,
+    },
+}
+
+/// All mutable UI state lives here.
+#[derive(Debug)]
+pub struct AppState {
+    pub session_id: Option<SessionId>,
+    pub session_title: String,
+    pub messages: Vec<MessageView>,
+    pub modal: ModalState,
+    pub input: tui_textarea::TextArea<'static>,
+    pub scroll_offset: u16,
+    pub running: bool,
+    pub model_ref: String,
+    pub agent: String,
+    pub quit: bool,
+    pub dirty: bool,
+    pub ctrl_c_pressed: bool,
+    /// Width the currently cached part lines were wrapped to; a change invalidates them.
+    pub render_width: u16,
+    /// Accumulated session cost in USD, from `SessionCreated`/`SessionUpdated` events.
+    pub session_cost: f64,
+    /// Accumulated session tokens (input + output), for the status bar.
+    pub session_tokens: u64,
+}
+
+impl AppState {
+    pub fn new(model_ref: String, agent: String) -> Self {
+        let mut input = tui_textarea::TextArea::default();
+        input.set_cursor_line_style(ratatui::style::Style::default());
+        Self {
+            session_id: None,
+            session_title: String::new(),
+            messages: Vec::new(),
+            modal: ModalState::None,
+            input,
+            scroll_offset: 0,
+            running: false,
+            model_ref,
+            agent,
+            quit: false,
+            dirty: true,
+            ctrl_c_pressed: false,
+            render_width: 0,
+            session_cost: 0.0,
+            session_tokens: 0,
+        }
+    }
+
+    /// Drop every part's cached render (e.g. after a resize changes the wrap width).
+    pub fn invalidate_caches(&mut self) {
+        for message in &mut self.messages {
+            for part in &mut message.parts {
+                match part {
+                    PartView::Text { cached_lines, .. } | PartView::Tool { cached_lines, .. } => {
+                        *cached_lines = None
+                    }
+                    PartView::Reasoning { .. } | PartView::StepFinish { .. } => {}
+                }
+            }
+        }
+    }
+
+    pub fn set_session(&mut self, session: Session, messages: Vec<Message>, parts: Vec<Part>) {
+        self.session_id = Some(session.id.clone());
+        self.session_title = session.title;
+        self.agent = session.agent;
+        self.model_ref = format!("{}/{}", session.model.provider_id, session.model.model_id);
+        self.session_cost = session.cost;
+        self.session_tokens = session.usage.input + session.usage.output;
+        self.messages.clear();
+
+        let mut messages = messages;
+        messages.sort_by_key(|m| m.created_at);
+        for message in messages {
+            let message_parts: Vec<Part> = parts
+                .iter()
+                .filter(|p| p.message_id == message.id)
+                .cloned()
+                .collect();
+            self.messages.push(message_view(message, message_parts));
+        }
+        self.scroll_offset = 0;
+        self.dirty = true;
+    }
+
+    pub fn apply_event(&mut self, event: AppEvent) {
+        match event {
+            AppEvent::SessionCreated { session } | AppEvent::SessionUpdated { session } => {
+                if self.session_id.as_ref() == Some(&session.id) {
+                    self.session_title = session.title;
+                    self.agent = session.agent;
+                    self.model_ref =
+                        format!("{}/{}", session.model.provider_id, session.model.model_id);
+                    self.session_cost = session.cost;
+                    self.session_tokens = session.usage.input + session.usage.output;
+                    self.dirty = true;
+                }
+            }
+            AppEvent::MessageCreated { message } => {
+                if self.session_id.as_ref() == Some(&message.session_id) {
+                    self.messages.push(message_view(message, Vec::new()));
+                    self.dirty = true;
+                }
+            }
+            AppEvent::MessageUpdated { message } => {
+                if let Some(view) = self.messages.iter_mut().find(|m| m.id == message.id) {
+                    view.role = message.role;
+                    self.dirty = true;
+                }
+            }
+            AppEvent::PartUpdated { part } => {
+                if self.session_id.as_ref() == Some(&part.session_id) {
+                    self.update_or_insert_part(part);
+                    self.dirty = true;
+                }
+            }
+            AppEvent::PartDelta {
+                part_id,
+                message_id,
+                delta,
+            } => {
+                self.apply_delta(part_id, message_id, delta);
+                self.dirty = true;
+            }
+            AppEvent::RunStarted { session_id } => {
+                if self.session_id.as_ref() == Some(&session_id) {
+                    self.running = true;
+                    self.dirty = true;
+                }
+            }
+            AppEvent::RunFinished {
+                session_id,
+                outcome,
+            } => {
+                if self.session_id.as_ref() == Some(&session_id) {
+                    self.running = false;
+                    if matches!(outcome, RunOutcome::Errored { .. }) {
+                        self.messages.push(MessageView {
+                            id: harness_core::types::MessageId::new(),
+                            role: Role::Assistant,
+                            parts: vec![PartView::Text {
+                                id: PartId::new(),
+                                text: match outcome {
+                                    RunOutcome::Errored { message } => message,
+                                    _ => String::new(),
+                                },
+                                cached_lines: None,
+                            }],
+                        });
+                    }
+                    self.dirty = true;
+                }
+            }
+            AppEvent::PermissionAsked { request } => {
+                if let ModalState::Permission { requests } = &mut self.modal {
+                    requests.push(request);
+                } else {
+                    self.modal = ModalState::Permission {
+                        requests: vec![request],
+                    };
+                }
+                self.dirty = true;
+            }
+            AppEvent::PermissionResolved { id } => {
+                if let ModalState::Permission { requests } = &mut self.modal {
+                    requests.retain(|r| r.id != id);
+                    if requests.is_empty() {
+                        self.modal = ModalState::None;
+                    }
+                }
+                self.dirty = true;
+            }
+            AppEvent::JobUpdated { .. }
+            | AppEvent::AuthPrompt { .. }
+            | AppEvent::ServerNotice { .. } => {
+                self.dirty = true;
+            }
+        }
+    }
+
+    pub fn scroll_up(&mut self, n: u16) {
+        self.scroll_offset = self.scroll_offset.saturating_sub(n);
+        self.dirty = true;
+    }
+
+    pub fn scroll_down(&mut self, n: u16) {
+        self.scroll_offset = self.scroll_offset.saturating_add(n);
+        self.dirty = true;
+    }
+
+    pub fn handle_permission_reply(&mut self, reply: PermissionReply, engine: &EngineHandle) {
+        let request_id = if let ModalState::Permission { requests } = &self.modal {
+            requests.first().map(|r| r.id.clone())
+        } else {
+            None
+        };
+
+        if let Some(id) = request_id {
+            engine.permission_reply(&id, reply);
+            if let ModalState::Permission { requests } = &mut self.modal {
+                requests.retain(|r| r.id != id);
+                if requests.is_empty() {
+                    self.modal = ModalState::None;
+                }
+            }
+            self.dirty = true;
+        }
+    }
+
+    pub fn open_session_picker(&mut self, sessions: Vec<Session>) {
+        self.modal = ModalState::SessionPicker {
+            sessions,
+            selected: 0,
+        };
+        self.dirty = true;
+    }
+
+    pub fn close_modal(&mut self) {
+        self.modal = ModalState::None;
+        self.dirty = true;
+    }
+
+    fn update_or_insert_part(&mut self, part: Part) {
+        let target_id = part.id.clone();
+        let Some(message_view) = self.messages.iter_mut().find(|m| m.id == part.message_id) else {
+            return;
+        };
+
+        let new_view = part_view(part);
+        if let Some(existing) = message_view
+            .parts
+            .iter_mut()
+            .find(|p| view_part_id(p) == target_id)
+        {
+            *existing = new_view;
+        } else {
+            message_view.parts.push(new_view);
+        }
+    }
+
+    fn apply_delta(
+        &mut self,
+        target_id: PartId,
+        message_id: harness_core::types::MessageId,
+        delta: String,
+    ) {
+        let Some(message_view) = self.messages.iter_mut().find(|m| m.id == message_id) else {
+            return;
+        };
+
+        for part in &mut message_view.parts {
+            if view_part_id(part) == target_id {
+                match part {
+                    PartView::Text {
+                        text, cached_lines, ..
+                    } => {
+                        text.push_str(&delta);
+                        *cached_lines = None;
+                    }
+                    PartView::Reasoning { text, .. } => {
+                        text.push_str(&delta);
+                    }
+                    PartView::Tool { cached_lines, .. } => {
+                        *cached_lines = None;
+                    }
+                    PartView::StepFinish { .. } => {}
+                }
+                return;
+            }
+        }
+    }
+}
+
+fn message_view(message: Message, mut parts: Vec<Part>) -> MessageView {
+    parts.sort_by_key(|p| p.idx);
+    MessageView {
+        id: message.id,
+        role: message.role,
+        parts: parts.into_iter().map(part_view).collect(),
+    }
+}
+
+fn part_view(part: Part) -> PartView {
+    match part.body {
+        PartBody::Text { text, .. } => PartView::Text {
+            id: part.id,
+            text,
+            cached_lines: None,
+        },
+        PartBody::Reasoning { text, .. } => PartView::Reasoning { id: part.id, text },
+        PartBody::Tool { name, state, .. } => PartView::Tool {
+            id: part.id,
+            name,
+            state: ToolStateView::from_tool_state(&state),
+            cached_lines: None,
+        },
+        PartBody::StepStart | PartBody::StepFinish { .. } => PartView::StepFinish { id: part.id },
+        PartBody::Subtask { description, .. } => PartView::Text {
+            id: part.id,
+            text: description,
+            cached_lines: None,
+        },
+        PartBody::Compaction { summary, .. } => PartView::Text {
+            id: part.id,
+            text: format!("_Compaction summary: {summary}_"),
+            cached_lines: None,
+        },
+    }
+}
+
+fn view_part_id(part: &PartView) -> PartId {
+    match part {
+        PartView::Text { id, .. }
+        | PartView::Reasoning { id, .. }
+        | PartView::Tool { id, .. }
+        | PartView::StepFinish { id } => id.clone(),
+    }
+}

crates/harness-tui/src/terminal.rs

@@ -0,0 +1,41 @@
+use std::io;
+use std::panic;
+
+use crossterm::cursor::Show;
+use crossterm::execute;
+use crossterm::terminal::{
+    disable_raw_mode, enable_raw_mode, EnterAlternateScreen, LeaveAlternateScreen,
+};
+
+/// Restores the terminal when dropped or on panic.
+pub struct TerminalGuard;
+
+impl TerminalGuard {
+    /// Enables raw mode, enters the alternate screen, and installs a panic hook
+    /// that restores the terminal before printing panic info.
+    pub fn enter() -> io::Result<Self> {
+        enable_raw_mode()?;
+        let mut stdout = io::stdout();
+        execute!(stdout, EnterAlternateScreen, Show)?;
+
+        let original_hook = panic::take_hook();
+        panic::set_hook(Box::new(move |info| {
+            let _ = restore();
+            original_hook(info);
+        }));
+
+        Ok(Self)
+    }
+}
+
+impl Drop for TerminalGuard {
+    fn drop(&mut self) {
+        let _ = restore();
+    }
+}
+
+fn restore() -> io::Result<()> {
+    disable_raw_mode()?;
+    let mut stdout = io::stdout();
+    execute!(stdout, LeaveAlternateScreen, Show)
+}