M6: crash-resume hardening

Hardens resume-after-crash: an unfinished assistant message left by a killed process is marked aborted on load rather than replayed, covered in harness-app and a small harness-tui::modal fix.

Cargo.lock

@@ -35,6 +35,15 @@ version = "0.2.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "683d7910e743518b0e34f1186f92494becacb047c7b6bf616c96772180fef923"
 
+[[package]]
+name = "android_system_properties"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "819e7219dbd41043ac279b19830f2efc897156490d7fd6ea916720117ee66311"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "anyhow"
 version = "1.0.103"
@@ -92,6 +101,18 @@ version = "1.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1505bd5d3d116872e7271a6d4e16d81d0c8570876c8de68093a09ac269d8aac0"
 
+[[package]]
+name = "autocfg"
+version = "1.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2032f911046de80f0a198e0901378627c33f59ea0ac00e363d481118bd70a53"
+
+[[package]]
+name = "base64"
+version = "0.21.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9d297deb1925b89f2ccc13d7635fa0714f12c87adce1c75356b39ca9b7178567"
+
 [[package]]
 name = "base64"
 version = "0.22.1"
@@ -175,6 +196,20 @@ dependencies = [
  "rand_core 0.10.1",
 ]
 
+[[package]]
+name = "chrono"
+version = "0.4.45"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1aa79e62e7697b8e29b513a68abacf485adcd1fe8284a4316c5ae868e6633327"
+dependencies = [
+ "iana-time-zone",
+ "js-sys",
+ "num-traits",
+ "serde",
+ "wasm-bindgen",
+ "windows-link",
+]
+
 [[package]]
 name = "compact_str"
 version = "0.8.2"
@@ -217,6 +252,12 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
+[[package]]
+name = "core-foundation-sys"
+version = "0.8.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
+
 [[package]]
 name = "cpufeatures"
 version = "0.3.0"
@@ -669,16 +710,20 @@ dependencies = [
 name = "harness-app"
 version = "0.1.0"
 dependencies = [
+ "async-trait",
  "dirs",
+ "futures",
  "harness-core",
  "harness-lsp",
  "harness-mcp",
  "harness-providers",
  "harness-tools",
+ "serde_json",
  "tempfile",
  "thiserror 2.0.18",
  "tokio",
  "tokio-util",
+ "tracing",
 ]
 
 [[package]]
@@ -694,6 +739,7 @@ dependencies = [
  "schemars",
  "serde",
  "serde_json",
+ "serde_yaml_ng",
  "tempfile",
  "thiserror 2.0.18",
  "tokio",
@@ -706,14 +752,28 @@ dependencies = [
 name = "harness-lsp"
 version = "0.1.0"
 dependencies = [
+ "async-trait",
  "harness-core",
+ "serde_json",
+ "tempfile",
+ "tokio",
+ "tracing",
 ]
 
 [[package]]
 name = "harness-mcp"
 version = "0.1.0"
 dependencies = [
+ "async-trait",
  "harness-core",
+ "rmcp",
+ "serde",
+ "serde_json",
+ "tempfile",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-util",
+ "tracing",
 ]
 
 [[package]]
@@ -723,12 +783,14 @@ dependencies = [
  "async-stream",
  "async-trait",
  "bytes",
+ "dirs",
  "eventsource-stream",
  "futures",
  "harness-core",
  "reqwest",
  "serde",
  "serde_json",
+ "tempfile",
  "thiserror 2.0.18",
  "tokio",
  "tokio-util",
@@ -800,6 +862,12 @@ dependencies = [
  "foldhash",
 ]
 
+[[package]]
+name = "hashbrown"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+
 [[package]]
 name = "hashlink"
 version = "0.9.1"
@@ -896,7 +964,7 @@ version = "0.1.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
 dependencies = [
- "base64",
+ "base64 0.22.1",
  "bytes",
  "futures-channel",
  "futures-util",
@@ -913,6 +981,30 @@ dependencies = [
  "tracing",
 ]
 
+[[package]]
+name = "iana-time-zone"
+version = "0.1.65"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e31bc9ad994ba00e440a8aa5c9ef0ec67d5cb5e5cb0cc7f8b744a35b389cc470"
+dependencies = [
+ "android_system_properties",
+ "core-foundation-sys",
+ "iana-time-zone-haiku",
+ "js-sys",
+ "log",
+ "wasm-bindgen",
+ "windows-core",
+]
+
+[[package]]
+name = "iana-time-zone-haiku"
+version = "0.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f31827a206f56af32e590ba56d5d2d085f558508192593743f16b2306495269f"
+dependencies = [
+ "cc",
+]
+
 [[package]]
 name = "icu_collections"
 version = "2.2.0"
@@ -1038,6 +1130,16 @@ dependencies = [
  "winapi-util",
 ]
 
+[[package]]
+name = "indexmap"
+version = "2.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d466e9454f08e4a911e14806c24e16fba1b4c121d1ea474396f396069cf949d9"
+dependencies = [
+ "equivalent",
+ "hashbrown 0.17.1",
+]
+
 [[package]]
 name = "indoc"
 version = "2.0.7"
@@ -1270,6 +1372,15 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
 
+[[package]]
+name = "num-traits"
+version = "0.2.19"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "071dfc062690e90b734c0b2273ce72ad0ffa95f0c74596bc250dcfd960262841"
+dependencies = [
+ "autocfg",
+]
+
 [[package]]
 name = "once_cell"
 version = "1.21.4"
@@ -1577,7 +1688,7 @@ version = "0.12.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
 dependencies = [
- "base64",
+ "base64 0.22.1",
  "bytes",
  "futures-core",
  "futures-util",
@@ -1626,6 +1737,38 @@ dependencies = [
  "windows-sys 0.52.0",
 ]
 
+[[package]]
+name = "rmcp"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "33a0110d28bd076f39e14bfd5b0340216dd18effeb5d02b43215944cc3e5c751"
+dependencies = [
+ "base64 0.21.7",
+ "chrono",
+ "futures",
+ "paste",
+ "pin-project-lite",
+ "rmcp-macros",
+ "schemars",
+ "serde",
+ "serde_json",
+ "thiserror 2.0.18",
+ "tokio",
+ "tokio-util",
+ "tracing",
+]
+
+[[package]]
+name = "rmcp-macros"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6e2b2fd7497540489fa2db285edd43b7ed14c49157157438664278da6e42a7a"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
 [[package]]
 name = "rusqlite"
 version = "0.32.1"
@@ -1824,6 +1967,19 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "serde_yaml_ng"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7b4db627b98b36d4203a7b458cf3573730f2bb591b28871d916dfa9efabfd41f"
+dependencies = [
+ "indexmap",
+ "itoa",
+ "ryu",
+ "serde",
+ "unsafe-libyaml",
+]
+
 [[package]]
 name = "sharded-slab"
 version = "0.1.7"
@@ -2354,6 +2510,12 @@ version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1fc81956842c57dac11422a97c3b8195a1ff727f06e85c84ed2e8aa277c9a0fd"
 
+[[package]]
+name = "unsafe-libyaml"
+version = "0.2.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "673aac59facbab8a9007c7f6108d11f63b603f7cabff99fabf650fea5c32b861"
+
 [[package]]
 name = "untrusted"
 version = "0.9.0"
@@ -2558,12 +2720,65 @@ version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "712e227841d057c1ee1cd2fb22fa7e5a5461ae8e48fa2ca79ec42cfc1931183f"
 
+[[package]]
+name = "windows-core"
+version = "0.62.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
+dependencies = [
+ "windows-implement",
+ "windows-interface",
+ "windows-link",
+ "windows-result",
+ "windows-strings",
+]
+
+[[package]]
+name = "windows-implement"
+version = "0.60.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "windows-interface"
+version = "0.59.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
 [[package]]
 name = "windows-link"
 version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f0805222e57f7521d6a62e36fa9163bc891acd422f971defe97d64e70d0a4fe5"
 
+[[package]]
+name = "windows-result"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7781fa89eaf60850ac3d2da7af8e5242a5ea78d1a11c49bf2910bb5a73853eb5"
+dependencies = [
+ "windows-link",
+]
+
+[[package]]
+name = "windows-strings"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
+dependencies = [
+ "windows-link",
+]
+
 [[package]]
 name = "windows-sys"
 version = "0.48.0"

crates/harness-app/Cargo.toml

@@ -12,11 +12,17 @@ harness-mcp = { workspace = true }
 harness-lsp = { workspace = true }
 tokio = { workspace = true }
 tokio-util = { workspace = true }
+futures = { workspace = true }
+serde_json = { workspace = true }
+async-trait = { workspace = true }
 dirs = { workspace = true }
 thiserror = { workspace = true }
+tracing = { workspace = true }
 
 [dev-dependencies]
 tempfile = { workspace = true }
+futures = { workspace = true }
+serde_json = { workspace = true }
 
 [lints]
 workspace = true

crates/harness-core/Cargo.toml

@@ -11,6 +11,7 @@ futures = { workspace = true }
 async-trait = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
+serde_yaml_ng = { workspace = true }
 schemars = { workspace = true }
 rusqlite = { workspace = true }
 globset = { workspace = true }

crates/harness-core/assets/agents/designer.md

@@ -0,0 +1,12 @@
+---
+description: UI and interaction design specialist for user-facing surfaces
+mode: subagent
+temperature: 0.4
+tools: { task: false }
+---
+You are Designer, a UI and interaction design specialist. You handle user-facing surfaces:
+layout, component structure, styling, and interaction details.
+
+- Understand the existing design language before proposing changes; stay consistent with it.
+- When implementing, make focused edits and describe the visual/interaction effect.
+- Call out accessibility and responsive concerns relevant to the change.

crates/harness-core/assets/agents/explorer.md

@@ -0,0 +1,16 @@
+---
+description: Read-only reconnaissance specialist for mapping code and finding relevant files
+mode: subagent
+temperature: 0.1
+tools: { write: false, edit: false, task: false }
+permission:
+  - { permission: "edit", pattern: "*", action: deny }
+  - { permission: "write", pattern: "*", action: deny }
+---
+You are Explorer, a read-only reconnaissance specialist. You locate the code, files, and
+facts the orchestrator needs and report back concisely.
+
+- Use read, glob, grep, and bash (read-only commands) to investigate.
+- Never modify files. Return a focused summary with concrete `path:line` references, not
+  file dumps.
+- State what you found and, briefly, what you could not find.

crates/harness-core/assets/agents/fixer.md

@@ -0,0 +1,13 @@
+---
+description: Implementation specialist that makes focused code changes and verifies them
+mode: subagent
+temperature: 0.2
+tools: { task: false }
+---
+You are Fixer, an implementation specialist. You take a well-scoped change, implement it,
+and verify it compiles/tests.
+
+- Make the smallest change that satisfies the request; match the surrounding code's style.
+- Use read/grep to understand context before editing; use bash to build and run tests.
+- Report exactly what you changed (files and the essence of the diff) and the result of any
+  verification you ran.

crates/harness-core/assets/agents/librarian.md

@@ -0,0 +1,15 @@
+---
+description: Documentation and knowledge lookup specialist
+mode: subagent
+temperature: 0.1
+tools: { write: false, edit: false, task: false }
+permission:
+  - { permission: "edit", pattern: "*", action: deny }
+  - { permission: "write", pattern: "*", action: deny }
+---
+You are Librarian. You find and summarize documentation, comments, READMEs, config, and
+other in-repo knowledge on request.
+
+- Search docs and source for the relevant material with read, glob, and grep.
+- Quote the authoritative source with its `path:line`; do not invent details.
+- Return a concise, well-organized summary with pointers back to the sources.

crates/harness-core/assets/agents/oracle.md

@@ -0,0 +1,15 @@
+---
+description: Deep-reasoning analyst for architecture, debugging, and design trade-offs
+mode: subagent
+temperature: 0.3
+tools: { write: false, edit: false, task: false }
+permission:
+  - { permission: "edit", pattern: "*", action: deny }
+  - { permission: "write", pattern: "*", action: deny }
+---
+You are Oracle, a deep-reasoning analyst. You are consulted for hard questions:
+root-causing bugs, weighing architectural trade-offs, and reviewing designs.
+
+- Read whatever code and context you need, but do not modify anything.
+- Reason carefully and explicitly; state assumptions and the evidence behind conclusions.
+- Return a decisive recommendation with the reasoning that supports it.

crates/harness-core/assets/agents/orchestrator.md

@@ -0,0 +1,19 @@
+---
+description: Primary coordinator that plans work and delegates to specialists
+mode: primary
+temperature: 0.2
+---
+You are the orchestrator. You plan the work, delegate focused pieces to specialist
+subagents via the `task` tool, and synthesize their results into a final answer.
+
+Guidelines:
+- Break the request into concrete, independently-verifiable pieces.
+- Prefer delegating reconnaissance and analysis to subagents so your own context stays
+  focused; do the integration and final write-up yourself.
+- Launch background tasks for long-running independent work, then continue planning.
+  Do not poll running jobs — wait for completion and reconcile terminal jobs before your
+  final response.
+- Reuse a completed specialist session (by its job alias) when following up on the same
+  thread of work.
+
+{{SUBAGENTS}}

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

@@ -0,0 +1,496 @@
+//! Agent definitions and registry.
+//!
+//! All agent *behavior* lives in markdown + config — the engine only understands `mode`, tool
+//! filters, permissions, model, and depth. Definitions are layered (bundled → global → project
+//! → config patch); later layers win by name. See `docs/04-multiagent.md`.
+
+use std::collections::HashMap;
+use std::path::Path;
+
+use serde::Deserialize;
+
+use crate::config::AgentPatch;
+use crate::permission::{Rule, Ruleset};
+use crate::types::ModelRef;
+
+/// Marker in a primary agent's prompt replaced at load time with the routing list of enabled
+/// subagents (name + description). Keeps routing text in sync with the enabled agent set.
+const SUBAGENTS_MARKER: &str = "{{SUBAGENTS}}";
+
+const BUNDLED: &[(&str, &str)] = &[
+    (
+        "orchestrator",
+        include_str!("../../assets/agents/orchestrator.md"),
+    ),
+    ("explorer", include_str!("../../assets/agents/explorer.md")),
+    ("oracle", include_str!("../../assets/agents/oracle.md")),
+    (
+        "librarian",
+        include_str!("../../assets/agents/librarian.md"),
+    ),
+    ("fixer", include_str!("../../assets/agents/fixer.md")),
+    ("designer", include_str!("../../assets/agents/designer.md")),
+];
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum AgentMode {
+    Primary,
+    #[default]
+    Subagent,
+    All,
+}
+
+impl AgentMode {
+    fn parse(s: &str) -> Option<Self> {
+        match s.trim().to_ascii_lowercase().as_str() {
+            "primary" => Some(Self::Primary),
+            "subagent" => Some(Self::Subagent),
+            "all" => Some(Self::All),
+            _ => None,
+        }
+    }
+
+    /// Whether this agent can be invoked as a subagent via the `task` tool.
+    pub fn is_subagent(self) -> bool {
+        matches!(self, Self::Subagent | Self::All)
+    }
+
+    /// Whether this agent can drive a top-level (primary) session.
+    pub fn is_primary(self) -> bool {
+        matches!(self, Self::Primary | Self::All)
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum AgentSource {
+    Bundled,
+    Global,
+    Project,
+    Config,
+}
+
+#[derive(Debug, Clone)]
+pub struct AgentDef {
+    pub name: String,
+    pub description: String,
+    pub mode: AgentMode,
+    /// `None` = follow the session model.
+    pub model: Option<ModelRef>,
+    pub temperature: Option<f32>,
+    pub prompt: String,
+    pub permissions: Ruleset,
+    /// Tool enable/disable overrides (wildcard keys allowed); absent = inherit default.
+    pub tools: HashMap<String, bool>,
+    pub max_steps: Option<u32>,
+    pub source: AgentSource,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum AgentError {
+    #[error("agent {0}: {1}")]
+    Frontmatter(String, String),
+}
+
+/// YAML frontmatter shape (opencode-compatible).
+#[derive(Debug, Default, Deserialize)]
+#[serde(default)]
+struct Frontmatter {
+    description: Option<String>,
+    mode: Option<String>,
+    model: Option<String>,
+    temperature: Option<f32>,
+    tools: Option<HashMap<String, bool>>,
+    permission: Option<Vec<Rule>>,
+    max_steps: Option<u32>,
+    disable: Option<bool>,
+}
+
+fn parse_model_ref(s: &str) -> Option<ModelRef> {
+    s.split_once('/')
+        .map(|(p, m)| ModelRef::new(p.trim(), m.trim()))
+}
+
+/// Splits a markdown agent file into (frontmatter, body). A file without a leading `---`
+/// fence is treated as an all-body prompt with empty frontmatter.
+fn split_frontmatter(content: &str) -> (&str, &str) {
+    let rest = match content
+        .strip_prefix("---\n")
+        .or_else(|| content.strip_prefix("---\r\n"))
+    {
+        Some(r) => r,
+        None => return ("", content),
+    };
+    // Find the closing fence line.
+    for delim in ["\n---\n", "\n---\r\n"] {
+        if let Some(idx) = rest.find(delim) {
+            let body_start = idx + delim.len();
+            return (&rest[..idx], &rest[body_start..]);
+        }
+    }
+    // Trailing fence with no body / no trailing newline.
+    if let Some(fm) = rest.strip_suffix("\n---").or(Some(rest)) {
+        if rest.ends_with("\n---") {
+            return (fm, "");
+        }
+    }
+    ("", content)
+}
+
+/// Parses one markdown agent definition. Returns `Ok(None)` when the file marks itself
+/// `disable: true`.
+fn parse_agent(
+    name: &str,
+    source: AgentSource,
+    content: &str,
+) -> Result<Option<AgentDef>, AgentError> {
+    let (fm_raw, body) = split_frontmatter(content);
+    let fm: Frontmatter = if fm_raw.trim().is_empty() {
+        Frontmatter::default()
+    } else {
+        serde_yaml_ng::from_str(fm_raw)
+            .map_err(|e| AgentError::Frontmatter(name.to_string(), e.to_string()))?
+    };
+
+    if fm.disable == Some(true) {
+        return Ok(None);
+    }
+
+    Ok(Some(AgentDef {
+        name: name.to_string(),
+        description: fm.description.unwrap_or_default(),
+        mode: fm
+            .mode
+            .as_deref()
+            .and_then(AgentMode::parse)
+            .unwrap_or_default(),
+        model: fm.model.as_deref().and_then(parse_model_ref),
+        temperature: fm.temperature,
+        prompt: body.trim_end().to_string(),
+        permissions: fm.permission.unwrap_or_default(),
+        tools: fm.tools.unwrap_or_default(),
+        max_steps: fm.max_steps,
+        source,
+    }))
+}
+
+/// Applies a config `AgentPatch` onto an existing definition (only set fields override).
+fn apply_patch(def: &mut AgentDef, patch: &AgentPatch) {
+    if let Some(mode) = patch.mode.as_deref().and_then(AgentMode::parse) {
+        def.mode = mode;
+    }
+    if let Some(model) = patch.model.as_deref().and_then(parse_model_ref) {
+        def.model = Some(model);
+    }
+    if let Some(temp) = patch.temperature {
+        def.temperature = Some(temp);
+    }
+    if let Some(prompt) = &patch.prompt {
+        def.prompt = prompt.clone();
+    }
+    if let Some(tools) = &patch.tools {
+        def.tools.extend(tools.clone());
+    }
+    if let Some(permission) = &patch.permission {
+        def.permissions = permission.clone();
+    }
+}
+
+#[derive(Debug, Clone, Default)]
+pub struct AgentRegistry {
+    agents: HashMap<String, AgentDef>,
+}
+
+impl AgentRegistry {
+    pub fn get(&self, name: &str) -> Option<&AgentDef> {
+        self.agents.get(name)
+    }
+
+    pub fn all(&self) -> Vec<&AgentDef> {
+        self.agents.values().collect()
+    }
+
+    pub fn len(&self) -> usize {
+        self.agents.len()
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.agents.is_empty()
+    }
+
+    /// Just the bundled agents — the default when no overrides are configured (tests, headless).
+    pub fn bundled() -> Self {
+        let mut reg = Self::default();
+        reg.load_markdown_layer(
+            BUNDLED.iter().map(|(n, c)| (n.to_string(), *c)),
+            AgentSource::Bundled,
+        );
+        reg.generate_routing();
+        reg
+    }
+
+    /// Full layered load: bundled → global dir → project dir → config patches.
+    pub fn load(
+        config_agents: &HashMap<String, AgentPatch>,
+        global_dir: Option<&Path>,
+        project_dir: Option<&Path>,
+    ) -> Self {
+        let mut reg = Self::default();
+        reg.load_markdown_layer(
+            BUNDLED.iter().map(|(n, c)| (n.to_string(), *c)),
+            AgentSource::Bundled,
+        );
+        if let Some(dir) = global_dir {
+            reg.load_dir(dir, AgentSource::Global);
+        }
+        if let Some(dir) = project_dir {
+            reg.load_dir(dir, AgentSource::Project);
+        }
+        reg.apply_config(config_agents);
+        reg.generate_routing();
+        reg
+    }
+
+    fn load_markdown_layer<I, S>(&mut self, files: I, source: AgentSource)
+    where
+        I: IntoIterator<Item = (String, S)>,
+        S: AsRef<str>,
+    {
+        for (name, content) in files {
+            match parse_agent(&name, source, content.as_ref()) {
+                Ok(Some(def)) => {
+                    self.agents.insert(name, def);
+                }
+                Ok(None) => {
+                    self.agents.remove(&name); // disable: true removes an earlier layer
+                }
+                Err(e) => tracing::warn!(error = %e, "skipping malformed agent"),
+            }
+        }
+    }
+
+    fn load_dir(&mut self, dir: &Path, source: AgentSource) {
+        let Ok(entries) = std::fs::read_dir(dir) else {
+            return;
+        };
+        let mut files: Vec<(String, String)> = Vec::new();
+        for entry in entries.flatten() {
+            let path = entry.path();
+            if path.extension().and_then(|e| e.to_str()) != Some("md") {
+                continue;
+            }
+            let Some(name) = path.file_stem().and_then(|s| s.to_str()) else {
+                continue;
+            };
+            if let Ok(content) = std::fs::read_to_string(&path) {
+                files.push((name.to_string(), content));
+            }
+        }
+        files.sort();
+        self.load_markdown_layer(files, source);
+    }
+
+    fn apply_config(&mut self, config_agents: &HashMap<String, AgentPatch>) {
+        for (name, patch) in config_agents {
+            if patch.disable == Some(true) {
+                self.agents.remove(name);
+                continue;
+            }
+            if let Some(def) = self.agents.get_mut(name) {
+                apply_patch(def, patch);
+            } else if patch.model.is_some() || patch.prompt.is_some() {
+                // Unknown name with enough to stand on its own → custom agent.
+                let mut def = AgentDef {
+                    name: name.clone(),
+                    description: String::new(),
+                    mode: AgentMode::default(),
+                    model: None,
+                    temperature: None,
+                    prompt: String::new(),
+                    permissions: Vec::new(),
+                    tools: HashMap::new(),
+                    max_steps: None,
+                    source: AgentSource::Config,
+                };
+                apply_patch(&mut def, patch);
+                self.agents.insert(name.clone(), def);
+            }
+        }
+    }
+
+    /// Replaces `{{SUBAGENTS}}` in every primary agent's prompt with a generated routing list
+    /// of the enabled subagents (so disabling an agent removes it from routing text).
+    fn generate_routing(&mut self) {
+        let mut subagents: Vec<(String, String)> = self
+            .agents
+            .values()
+            .filter(|a| a.mode.is_subagent())
+            .map(|a| (a.name.clone(), a.description.clone()))
+            .collect();
+        subagents.sort();
+
+        let routing = if subagents.is_empty() {
+            "## Agents\n\nNo specialist subagents are available.".to_string()
+        } else {
+            let mut s = String::from("## Agents\n\nDelegate to these specialists via `task`:\n");
+            for (name, desc) in &subagents {
+                s.push_str(&format!("- **{name}** — {desc}\n"));
+            }
+            s
+        };
+
+        for agent in self.agents.values_mut() {
+            if agent.prompt.contains(SUBAGENTS_MARKER) {
+                agent.prompt = agent.prompt.replace(SUBAGENTS_MARKER, routing.trim_end());
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::permission::Action;
+
+    #[test]
+    fn bundled_loads_all_six_agents() {
+        let reg = AgentRegistry::bundled();
+        for name in [
+            "orchestrator",
+            "explorer",
+            "oracle",
+            "librarian",
+            "fixer",
+            "designer",
+        ] {
+            assert!(reg.get(name).is_some(), "missing {name}");
+        }
+        assert_eq!(reg.len(), 6);
+    }
+
+    #[test]
+    fn parses_mode_model_temperature_tools_and_permissions() {
+        let md = "---\n\
+            description: test agent\n\
+            mode: subagent\n\
+            model: anthropic/claude-haiku-4-5\n\
+            temperature: 0.1\n\
+            tools: { write: false, bash: true }\n\
+            permission:\n\
+            \x20 - { permission: \"edit\", pattern: \"*\", action: deny }\n\
+            ---\n\
+            You are a test agent.\n";
+        let def = parse_agent("tester", AgentSource::Bundled, md)
+            .unwrap()
+            .unwrap();
+        assert_eq!(def.description, "test agent");
+        assert_eq!(def.mode, AgentMode::Subagent);
+        assert_eq!(
+            def.model,
+            Some(ModelRef::new("anthropic", "claude-haiku-4-5"))
+        );
+        assert_eq!(def.temperature, Some(0.1));
+        assert_eq!(def.tools.get("write"), Some(&false));
+        assert_eq!(def.tools.get("bash"), Some(&true));
+        assert_eq!(def.permissions.len(), 1);
+        assert_eq!(def.permissions[0].action, Action::Deny);
+        assert_eq!(def.prompt, "You are a test agent.");
+    }
+
+    #[test]
+    fn body_without_frontmatter_is_all_prompt() {
+        let def = parse_agent("x", AgentSource::Global, "just a prompt")
+            .unwrap()
+            .unwrap();
+        assert_eq!(def.prompt, "just a prompt");
+        assert_eq!(def.mode, AgentMode::Subagent); // default
+    }
+
+    #[test]
+    fn disable_true_removes_agent() {
+        assert!(
+            parse_agent("x", AgentSource::Config, "---\ndisable: true\n---\nbody")
+                .unwrap()
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn explorer_denies_writes_and_disables_edit_tool() {
+        let reg = AgentRegistry::bundled();
+        let explorer = reg.get("explorer").unwrap();
+        assert_eq!(explorer.mode, AgentMode::Subagent);
+        assert_eq!(explorer.tools.get("edit"), Some(&false));
+        assert!(explorer
+            .permissions
+            .iter()
+            .any(|r| r.permission == "write" && r.action == Action::Deny));
+    }
+
+    #[test]
+    fn orchestrator_routing_lists_subagents_and_drops_marker() {
+        let reg = AgentRegistry::bundled();
+        let prompt = &reg.get("orchestrator").unwrap().prompt;
+        assert!(!prompt.contains("{{SUBAGENTS}}"));
+        assert!(prompt.contains("explorer"));
+        assert!(prompt.contains("fixer"));
+        // The orchestrator itself is primary and must not list itself.
+        assert!(!prompt.contains("- **orchestrator**"));
+    }
+
+    #[test]
+    fn config_patch_overrides_only_set_fields() {
+        let mut patches = HashMap::new();
+        patches.insert(
+            "explorer".to_string(),
+            AgentPatch {
+                temperature: Some(0.9),
+                ..Default::default()
+            },
+        );
+        let reg = AgentRegistry::load(&patches, None, None);
+        let explorer = reg.get("explorer").unwrap();
+        assert_eq!(explorer.temperature, Some(0.9)); // overridden
+        assert_eq!(explorer.mode, AgentMode::Subagent); // untouched
+        assert_eq!(explorer.tools.get("edit"), Some(&false)); // untouched
+    }
+
+    #[test]
+    fn config_disable_removes_and_unknown_with_model_creates() {
+        let mut patches = HashMap::new();
+        patches.insert(
+            "designer".to_string(),
+            AgentPatch {
+                disable: Some(true),
+                ..Default::default()
+            },
+        );
+        patches.insert(
+            "custom".to_string(),
+            AgentPatch {
+                model: Some("openai/gpt-5".into()),
+                prompt: Some("custom prompt".into()),
+                ..Default::default()
+            },
+        );
+        let reg = AgentRegistry::load(&patches, None, None);
+        assert!(reg.get("designer").is_none());
+        let custom = reg.get("custom").unwrap();
+        assert_eq!(custom.source, AgentSource::Config);
+        assert_eq!(custom.model, Some(ModelRef::new("openai", "gpt-5")));
+        assert_eq!(custom.prompt, "custom prompt");
+    }
+
+    #[test]
+    fn unknown_config_agent_without_model_or_prompt_is_ignored() {
+        let mut patches = HashMap::new();
+        patches.insert(
+            "ghost".to_string(),
+            AgentPatch {
+                temperature: Some(0.5),
+                ..Default::default()
+            },
+        );
+        let reg = AgentRegistry::load(&patches, None, None);
+        assert!(reg.get("ghost").is_none());
+    }
+}

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

@@ -29,6 +29,7 @@ const ENV_CONFIG_PATH: &str = "AI_HARNESS_CONFIG";
 const PROVIDER_ENV_KEYS: &[(&str, &str)] = &[
     ("anthropic", "ANTHROPIC_API_KEY"),
     ("openai", "OPENAI_API_KEY"),
+    ("opencode", "OPENCODE_API_KEY"),
 ];
 
 fn read_jsonc(path: &Path) -> Result<Value, ConfigError> {

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

@@ -0,0 +1,296 @@
+//! Markdown-defined commands and skills (M5). Both are YAML-frontmatter + body files loaded
+//! from a global dir (`~/.config/ai-harness/<kind>/`) and a project dir (`<project>/.harness/
+//! <kind>/`), project winning by name. See `docs/06-config.md` and `docs/09-integrations.md`.
+//!
+//! - Commands (`command/*.md`) are a pure input-layer concern: `/name args` expands the body
+//!   template (`$ARGUMENTS`, `$1..$9`) into the user message, optionally switching agent/model.
+//! - Skills (`skill/<name>/SKILL.md`) advertise `name + description` in the system prompt; the
+//!   model pulls a skill's body on demand via the built-in `skill` tool.
+
+use std::collections::HashMap;
+use std::path::Path;
+
+use serde::Deserialize;
+
+/// A slash command: a named prompt template with an optional agent/model override.
+#[derive(Debug, Clone, PartialEq)]
+pub struct CommandDef {
+    /// Invocation name (the file stem); used as `/name`.
+    pub name: String,
+    pub description: String,
+    /// Run the expanded prompt under this agent instead of the session's default.
+    pub agent: Option<String>,
+    /// Run under this `provider/model` instead of the session's default.
+    pub model: Option<String>,
+    /// The body, with `$ARGUMENTS` / `$1..$9` placeholders.
+    pub template: String,
+}
+
+impl CommandDef {
+    /// Substitutes `$ARGUMENTS` (the whole argument string) and `$1..$9` (whitespace-split
+    /// positionals; missing ones become empty) into the template.
+    pub fn expand(&self, arguments: &str) -> String {
+        let positionals: Vec<&str> = arguments.split_whitespace().collect();
+        let mut out = self.template.replace("$ARGUMENTS", arguments);
+        for i in 1..=9 {
+            let value = positionals.get(i - 1).copied().unwrap_or("");
+            out = out.replace(&format!("${i}"), value);
+        }
+        out
+    }
+}
+
+/// A skill: advertised by `name + description`, body loaded on demand by the `skill` tool.
+#[derive(Debug, Clone, PartialEq)]
+pub struct SkillDef {
+    /// Skill name (the containing directory name); used as the `skill` tool argument.
+    pub name: String,
+    pub description: String,
+    pub body: String,
+}
+
+/// Frontmatter fields shared by commands (all optional).
+#[derive(Debug, Default, Deserialize)]
+#[serde(default)]
+struct CommandFrontmatter {
+    description: Option<String>,
+    agent: Option<String>,
+    model: Option<String>,
+}
+
+#[derive(Debug, Default, Deserialize)]
+#[serde(default)]
+struct SkillFrontmatter {
+    description: Option<String>,
+}
+
+/// Splits `---\n…\n---\n` frontmatter from the body. A file with no leading fence is all body.
+fn split_frontmatter(content: &str) -> (&str, &str) {
+    let rest = match content
+        .strip_prefix("---\n")
+        .or_else(|| content.strip_prefix("---\r\n"))
+    {
+        Some(r) => r,
+        None => return ("", content),
+    };
+    for delim in ["\n---\n", "\n---\r\n"] {
+        if let Some(idx) = rest.find(delim) {
+            return (&rest[..idx], &rest[idx + delim.len()..]);
+        }
+    }
+    if rest.ends_with("\n---") {
+        return (rest.trim_end_matches("\n---"), "");
+    }
+    ("", content)
+}
+
+fn parse_command(name: &str, content: &str) -> CommandDef {
+    let (fm_raw, body) = split_frontmatter(content);
+    let fm: CommandFrontmatter = if fm_raw.trim().is_empty() {
+        CommandFrontmatter::default()
+    } else {
+        serde_yaml_ng::from_str(fm_raw).unwrap_or_default()
+    };
+    CommandDef {
+        name: name.to_string(),
+        description: fm.description.unwrap_or_default(),
+        agent: fm.agent,
+        model: fm.model,
+        template: body.trim().to_string(),
+    }
+}
+
+fn parse_skill(name: &str, content: &str) -> SkillDef {
+    let (fm_raw, body) = split_frontmatter(content);
+    let fm: SkillFrontmatter = if fm_raw.trim().is_empty() {
+        SkillFrontmatter::default()
+    } else {
+        serde_yaml_ng::from_str(fm_raw).unwrap_or_default()
+    };
+    SkillDef {
+        name: name.to_string(),
+        description: fm.description.unwrap_or_default(),
+        body: body.trim().to_string(),
+    }
+}
+
+/// Loads `command/*.md` from global then project dirs (project wins by name).
+pub fn load_commands(
+    global_dir: Option<&Path>,
+    project_dir: Option<&Path>,
+) -> HashMap<String, CommandDef> {
+    let mut commands = HashMap::new();
+    for dir in [global_dir, project_dir].into_iter().flatten() {
+        let Ok(entries) = std::fs::read_dir(dir) else {
+            continue;
+        };
+        for entry in entries.flatten() {
+            let path = entry.path();
+            if path.extension().and_then(|e| e.to_str()) != Some("md") {
+                continue;
+            }
+            let Some(name) = path.file_stem().and_then(|s| s.to_str()) else {
+                continue;
+            };
+            if let Ok(content) = std::fs::read_to_string(&path) {
+                commands.insert(name.to_string(), parse_command(name, &content));
+            }
+        }
+    }
+    commands
+}
+
+/// Loads `skill/<name>/SKILL.md` from global then project dirs (project wins by name),
+/// returned name-sorted for a stable system-prompt listing.
+pub fn load_skills(global_dir: Option<&Path>, project_dir: Option<&Path>) -> Vec<SkillDef> {
+    let mut by_name: HashMap<String, SkillDef> = HashMap::new();
+    for dir in [global_dir, project_dir].into_iter().flatten() {
+        let Ok(entries) = std::fs::read_dir(dir) else {
+            continue;
+        };
+        for entry in entries.flatten() {
+            let path = entry.path();
+            if !path.is_dir() {
+                continue;
+            }
+            let Some(name) = path.file_name().and_then(|s| s.to_str()) else {
+                continue;
+            };
+            let skill_file = path.join("SKILL.md");
+            if let Ok(content) = std::fs::read_to_string(&skill_file) {
+                by_name.insert(name.to_string(), parse_skill(name, &content));
+            }
+        }
+    }
+    let mut skills: Vec<SkillDef> = by_name.into_values().collect();
+    skills.sort_by(|a, b| a.name.cmp(&b.name));
+    skills
+}
+
+/// The system-prompt section advertising available skills (name + description). `None` when
+/// there are no skills, so no empty section is injected.
+pub fn skills_prompt(skills: &[SkillDef]) -> Option<String> {
+    if skills.is_empty() {
+        return None;
+    }
+    let mut section = String::from(
+        "## Skills\n\nThese skills are available. Load a skill's full instructions on demand \
+         by calling the `skill` tool with its name before doing the related work:\n",
+    );
+    for skill in skills {
+        section.push_str(&format!("- **{}** — {}\n", skill.name, skill.description));
+    }
+    Some(section.trim_end().to_string())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn expand_substitutes_arguments_and_positionals() {
+        let cmd = CommandDef {
+            name: "greet".into(),
+            description: String::new(),
+            agent: None,
+            model: None,
+            template: "Say $1 to $2. All: $ARGUMENTS".into(),
+        };
+        assert_eq!(cmd.expand("hi there"), "Say hi to there. All: hi there");
+        // Missing positionals collapse to empty.
+        assert_eq!(cmd.expand("solo"), "Say solo to . All: solo");
+    }
+
+    #[test]
+    fn parse_command_reads_frontmatter_and_body() {
+        let md = "---\ndescription: review a PR\nagent: oracle\nmodel: openai/gpt-5\n---\nReview $ARGUMENTS please.\n";
+        let cmd = parse_command("review", md);
+        assert_eq!(cmd.description, "review a PR");
+        assert_eq!(cmd.agent.as_deref(), Some("oracle"));
+        assert_eq!(cmd.model.as_deref(), Some("openai/gpt-5"));
+        assert_eq!(cmd.template, "Review $ARGUMENTS please.");
+    }
+
+    #[test]
+    fn parse_command_without_frontmatter_is_all_template() {
+        let cmd = parse_command("x", "just do $1");
+        assert_eq!(cmd.template, "just do $1");
+        assert!(cmd.agent.is_none());
+    }
+
+    #[test]
+    fn parse_skill_reads_description_and_body() {
+        let md = "---\ndescription: format code\n---\nRun the formatter.\n";
+        let skill = parse_skill("formatter", md);
+        assert_eq!(skill.name, "formatter");
+        assert_eq!(skill.description, "format code");
+        assert_eq!(skill.body, "Run the formatter.");
+    }
+
+    #[test]
+    fn skills_prompt_lists_each_and_is_none_when_empty() {
+        assert!(skills_prompt(&[]).is_none());
+        let skills = vec![
+            SkillDef {
+                name: "a".into(),
+                description: "does a".into(),
+                body: "…".into(),
+            },
+            SkillDef {
+                name: "b".into(),
+                description: "does b".into(),
+                body: "…".into(),
+            },
+        ];
+        let prompt = skills_prompt(&skills).unwrap();
+        assert!(prompt.contains("- **a** — does a"));
+        assert!(prompt.contains("- **b** — does b"));
+    }
+
+    #[test]
+    fn load_skills_reads_skill_dirs_and_project_wins() {
+        let dir = tempfile::tempdir().unwrap();
+        let global = dir.path().join("global");
+        let project = dir.path().join("project");
+        std::fs::create_dir_all(global.join("fmt")).unwrap();
+        std::fs::create_dir_all(project.join("fmt")).unwrap();
+        std::fs::create_dir_all(global.join("lint")).unwrap();
+        std::fs::write(
+            global.join("fmt/SKILL.md"),
+            "---\ndescription: global fmt\n---\nglobal body",
+        )
+        .unwrap();
+        std::fs::write(
+            project.join("fmt/SKILL.md"),
+            "---\ndescription: project fmt\n---\nproject body",
+        )
+        .unwrap();
+        std::fs::write(
+            global.join("lint/SKILL.md"),
+            "---\ndescription: lint\n---\nlint body",
+        )
+        .unwrap();
+
+        let skills = load_skills(Some(&global), Some(&project));
+        assert_eq!(skills.len(), 2);
+        // Sorted by name: fmt, lint.
+        assert_eq!(skills[0].name, "fmt");
+        assert_eq!(skills[0].description, "project fmt"); // project overrode global
+        assert_eq!(skills[0].body, "project body");
+        assert_eq!(skills[1].name, "lint");
+    }
+
+    #[test]
+    fn load_commands_project_overrides_global() {
+        let dir = tempfile::tempdir().unwrap();
+        let global = dir.path().join("g");
+        let project = dir.path().join("p");
+        std::fs::create_dir_all(&global).unwrap();
+        std::fs::create_dir_all(&project).unwrap();
+        std::fs::write(global.join("deploy.md"), "global deploy").unwrap();
+        std::fs::write(project.join("deploy.md"), "project deploy").unwrap();
+        let commands = load_commands(Some(&global), Some(&project));
+        assert_eq!(commands.len(), 1);
+        assert_eq!(commands["deploy"].template, "project deploy");
+    }
+}

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

@@ -1,7 +1,9 @@
 pub mod load;
+pub mod markdown;
 pub mod schema;
 
 pub use load::{load, ConfigError};
+pub use markdown::{load_commands, load_skills, skills_prompt, CommandDef, SkillDef};
 pub use schema::{
     AgentPatch, Config, LspServerConfig, McpServerConfig, OrchestrationConfig, ProviderConfig,
     TuiConfig,

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

@@ -0,0 +1,146 @@
+//! Auto-compaction (M6). When a session's context approaches the model's window the loop
+//! summarizes the conversation so far via a small model, writes a `Compaction` marker, and
+//! continues — subsequent requests replace the summarized history with the summary. See
+//! `docs/02-engine.md`.
+
+use async_trait::async_trait;
+
+use crate::event::AppEvent;
+use crate::llm::{ProviderError, WireMessage};
+use crate::types::{Message, MessageId, Part, PartBody, PartId, SessionId};
+
+use super::processor::StepContext;
+
+/// Produces a compact recap of a conversation. Implemented by the composition root over the
+/// configured `small_model`; absent in headless/test contexts (compaction then disabled).
+#[async_trait]
+pub trait Compactor: Send + Sync {
+    async fn summarize(&self, messages: &[WireMessage]) -> Result<String, ProviderError>;
+}
+
+/// System prompt handed to the small model to summarize the conversation for continuation.
+pub const SUMMARY_SYSTEM_PROMPT: &str = "You are compacting a long coding-assistant \
+    conversation so it can continue within a smaller context window. Write a dense, factual \
+    summary that preserves: the user's goal and constraints, decisions made and why, files \
+    and symbols touched, commands run and their results, and the exact next step in progress. \
+    Omit pleasantries. Output only the summary.";
+
+/// Frames a raw summary as the synthetic user turn the model sees after compaction.
+pub fn frame_summary(summary: &str) -> String {
+    format!(
+        "The earlier part of this conversation was summarized to save context. \
+         Continue from this summary:\n\n{summary}"
+    )
+}
+
+/// Persists a compaction: a new user message carrying the `Compaction` marker (used to cut
+/// history on the next load) plus a synthetic text part holding the framed summary (what the
+/// model actually reads). Returns the new message id.
+pub async fn write_compaction(
+    ctx: &StepContext,
+    session_id: &SessionId,
+    replaces_up_to: MessageId,
+    summary: String,
+    now: i64,
+) -> Result<MessageId, ProviderError> {
+    let message = Message::new_user(session_id.clone(), now);
+    let message_id = message.id.clone();
+    let store_err = |e: crate::store::StoreError| ProviderError::Decode(e.to_string());
+
+    ctx.store
+        .upsert_message(message.clone())
+        .await
+        .map_err(store_err)?;
+    ctx.bus.publish(AppEvent::MessageCreated {
+        message: message.clone(),
+    });
+
+    let marker = Part {
+        id: PartId::new(),
+        message_id: message_id.clone(),
+        session_id: session_id.clone(),
+        idx: 0,
+        body: PartBody::Compaction {
+            replaces_up_to,
+            summary: summary.clone(),
+        },
+    };
+    ctx.store.upsert_part(marker).await.map_err(store_err)?;
+
+    let text = Part {
+        id: PartId::new(),
+        message_id: message_id.clone(),
+        session_id: session_id.clone(),
+        idx: 1,
+        body: PartBody::Text {
+            text: frame_summary(&summary),
+            synthetic: true,
+        },
+    };
+    ctx.store
+        .upsert_part(text.clone())
+        .await
+        .map_err(store_err)?;
+    ctx.bus.publish(AppEvent::PartUpdated { part: text });
+
+    Ok(message_id)
+}
+
+/// Index of the last message carrying a `Compaction` part, given each message's parts. History
+/// before it is dropped when building the request; `None` means no compaction yet.
+pub fn last_compaction_index(parts_per_message: &[Vec<Part>]) -> Option<usize> {
+    parts_per_message.iter().rposition(|parts| {
+        parts
+            .iter()
+            .any(|p| matches!(p.body, PartBody::Compaction { .. }))
+    })
+}
+
+/// Whether accumulated `used` tokens have crossed the compaction threshold (90% of the
+/// model's context window). `context_limit == 0` (unknown) disables the trigger.
+pub fn over_threshold(used: u64, context_limit: u64) -> bool {
+    context_limit > 0 && used.saturating_mul(10) > context_limit.saturating_mul(9)
+}
+
+/// Effective context occupancy for a step's usage: prompt (incl. cache) plus generated output.
+pub fn tokens_used(usage: &crate::types::TokenUsage) -> u64 {
+    usage.input + usage.cache_read + usage.cache_write + usage.output
+}
+
+/// True if a message list has real history to compact before `cut_start` (avoids a useless
+/// compaction that would summarize nothing and could loop).
+pub fn has_history_to_compact(messages: &[Message], cut_start: usize) -> bool {
+    messages.len().saturating_sub(cut_start) > 1
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn over_threshold_respects_ninety_percent_and_unknown_limit() {
+        assert!(over_threshold(91, 100));
+        assert!(!over_threshold(90, 100)); // exactly 90% is not yet over
+        assert!(!over_threshold(50, 100));
+        assert!(!over_threshold(1_000_000, 0)); // unknown limit disables the trigger
+    }
+
+    #[test]
+    fn tokens_used_sums_prompt_cache_and_output() {
+        let usage = crate::types::TokenUsage {
+            input: 10,
+            output: 5,
+            reasoning: 0,
+            cache_read: 3,
+            cache_write: 2,
+        };
+        assert_eq!(tokens_used(&usage), 20);
+    }
+
+    #[test]
+    fn frame_summary_wraps_text() {
+        let framed = frame_summary("did X");
+        assert!(framed.contains("did X"));
+        assert!(framed.contains("summarized"));
+    }
+}

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

@@ -0,0 +1,579 @@
+//! Background job board — tracks subagent tasks spawned via the `task` tool so the
+//! orchestrator can see running work, reconcile terminal results, and reuse completed
+//! child sessions by alias. Simplified native port of oh-my-opencode-slim's
+//! `background-job-board.ts`. See `docs/04-multiagent.md`.
+//!
+//! The board is an in-memory `RwLock<HashMap>` mirrored to the `job` table so it survives a
+//! resume. All mutations persist through the passed-in `Store`.
+
+use std::collections::HashMap;
+use std::sync::RwLock;
+
+use serde::{Deserialize, Serialize};
+
+use crate::event::{AppEvent, EventBus, JobRecordEvent};
+use crate::store::{Store, StoreError};
+use crate::types::SessionId;
+
+/// A file a child session read, surfaced on the board so the orchestrator knows what a
+/// completed specialist already looked at.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ContextFile {
+    pub path: String,
+    pub lines: u32,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum JobState {
+    Running,
+    Completed,
+    Error,
+    Cancelled,
+}
+
+impl JobState {
+    /// Terminal jobs are candidates for reconciliation; a completed one is reusable.
+    pub fn is_terminal(self) -> bool {
+        !matches!(self, JobState::Running)
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct JobRecord {
+    pub task_id: String,
+    /// Human-friendly handle: first 3 chars of the agent name + a per-agent counter (`exp-1`).
+    pub alias: String,
+    pub parent_session: SessionId,
+    pub child_session: SessionId,
+    pub agent: String,
+    pub description: String,
+    pub objective: Option<String>,
+    pub state: JobState,
+    /// Whether the orchestrator has already seen this job's terminal result.
+    pub reconciled: bool,
+    pub result_summary: Option<String>,
+    pub context_files: Vec<ContextFile>,
+    pub launched_at: i64,
+    pub updated_at: i64,
+    pub last_used_at: i64,
+}
+
+/// Parameters for registering a newly launched task on the board.
+pub struct LaunchSpec {
+    pub task_id: String,
+    pub parent_session: SessionId,
+    pub child_session: SessionId,
+    pub agent: String,
+    pub description: String,
+    pub objective: Option<String>,
+}
+
+/// Result summaries are truncated to this many chars on the board.
+const SUMMARY_MAX: usize = 2000;
+/// Context files shown per job in the prompt injection.
+const CONTEXT_FILES_SHOWN: usize = 8;
+/// A read must cover at least this many lines to be worth reporting to the board.
+pub const MIN_REPORTED_LINES: u32 = 10;
+
+pub struct JobBoard {
+    store: Store,
+    bus: EventBus,
+    jobs: RwLock<HashMap<String, JobRecord>>,
+    max_reusable_per_agent: u32,
+}
+
+impl JobBoard {
+    /// Builds a board and loads any persisted jobs for `parent_session`'s tree from the store.
+    pub async fn load(
+        store: Store,
+        bus: EventBus,
+        parent_session: &SessionId,
+        max_reusable_per_agent: u32,
+    ) -> Result<Self, StoreError> {
+        let existing = store.jobs_for_parent(parent_session.clone()).await?;
+        let mut jobs = HashMap::new();
+        for job in existing {
+            jobs.insert(job.task_id.clone(), job);
+        }
+        Ok(Self {
+            store,
+            bus,
+            jobs: RwLock::new(jobs),
+            max_reusable_per_agent,
+        })
+    }
+
+    /// Assigns the next alias for `agent` under this board: `<3-char-prefix>-<n>`.
+    fn next_alias(&self, agent: &str) -> String {
+        let prefix: String = agent.chars().take(3).collect();
+        let prefix = if prefix.is_empty() {
+            "job".to_string()
+        } else {
+            prefix.to_ascii_lowercase()
+        };
+        let n = self
+            .jobs
+            .read()
+            .unwrap()
+            .values()
+            .filter(|j| j.agent == agent)
+            .count()
+            + 1;
+        format!("{prefix}-{n}")
+    }
+
+    /// Registers a freshly launched task and returns its assigned alias.
+    pub async fn register_launch(&self, spec: LaunchSpec, now: i64) -> Result<String, StoreError> {
+        let alias = self.next_alias(&spec.agent);
+        let record = JobRecord {
+            task_id: spec.task_id,
+            alias: alias.clone(),
+            parent_session: spec.parent_session,
+            child_session: spec.child_session,
+            agent: spec.agent,
+            description: spec.description,
+            objective: spec.objective,
+            state: JobState::Running,
+            reconciled: false,
+            result_summary: None,
+            context_files: Vec::new(),
+            launched_at: now,
+            updated_at: now,
+            last_used_at: now,
+        };
+        self.upsert(record).await?;
+        Ok(alias)
+    }
+
+    /// Marks a job terminal with an optional result summary (truncated).
+    pub async fn finish(
+        &self,
+        task_id: &str,
+        state: JobState,
+        result_summary: Option<String>,
+        now: i64,
+    ) -> Result<(), StoreError> {
+        let Some(mut record) = self.get(task_id) else {
+            return Ok(());
+        };
+        record.state = state;
+        record.result_summary = result_summary.map(|s| truncate_summary(&s));
+        record.updated_at = now;
+        record.last_used_at = now;
+        self.upsert(record).await?;
+        if state == JobState::Completed {
+            self.trim_reusable(now).await?;
+        }
+        Ok(())
+    }
+
+    /// Records a file a child session read (deduping by path, keeping the largest read).
+    pub async fn report_context_file(
+        &self,
+        task_id: &str,
+        path: String,
+        lines: u32,
+        now: i64,
+    ) -> Result<(), StoreError> {
+        if lines < MIN_REPORTED_LINES {
+            return Ok(());
+        }
+        let Some(mut record) = self.get(task_id) else {
+            return Ok(());
+        };
+        match record.context_files.iter_mut().find(|f| f.path == path) {
+            Some(existing) => existing.lines = existing.lines.max(lines),
+            None => record.context_files.push(ContextFile { path, lines }),
+        }
+        record.updated_at = now;
+        self.upsert(record).await
+    }
+
+    /// Resolves an alias or task id to a job for the given parent (reuse lookup). Only
+    /// completed (reusable) jobs match.
+    pub fn resolve_reusable(&self, parent: &SessionId, alias_or_id: &str) -> Option<JobRecord> {
+        self.jobs
+            .read()
+            .unwrap()
+            .values()
+            .find(|j| {
+                &j.parent_session == parent
+                    && j.state == JobState::Completed
+                    && (j.alias == alias_or_id || j.task_id == alias_or_id)
+            })
+            .cloned()
+    }
+
+    /// Bumps `last_used_at` when a completed session is reused, keeping it fresh in the LRU.
+    pub async fn touch(&self, task_id: &str, now: i64) -> Result<(), StoreError> {
+        let Some(mut record) = self.get(task_id) else {
+            return Ok(());
+        };
+        record.last_used_at = now;
+        self.upsert(record).await
+    }
+
+    /// Marks all terminal jobs `reconciled` — the orchestrator has now seen them.
+    pub async fn reconcile_terminal(&self, now: i64) -> Result<(), StoreError> {
+        let to_update: Vec<JobRecord> = {
+            let jobs = self.jobs.read().unwrap();
+            jobs.values()
+                .filter(|j| j.state.is_terminal() && !j.reconciled)
+                .cloned()
+                .collect()
+        };
+        for mut record in to_update {
+            record.reconciled = true;
+            record.updated_at = now;
+            self.upsert(record).await?;
+        }
+        Ok(())
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.jobs.read().unwrap().is_empty()
+    }
+
+    pub fn snapshot(&self) -> Vec<JobRecord> {
+        self.jobs.read().unwrap().values().cloned().collect()
+    }
+
+    /// Renders the board as a synthetic prompt block, or `None` when there is nothing to show.
+    /// Mirrors slim's `formatForPrompt`.
+    pub fn format_for_prompt(&self) -> Option<String> {
+        let jobs = self.jobs.read().unwrap();
+        if jobs.is_empty() {
+            return None;
+        }
+
+        let mut active: Vec<&JobRecord> = jobs
+            .values()
+            .filter(|j| !j.state.is_terminal() || !j.reconciled)
+            .collect();
+        let mut reusable: Vec<&JobRecord> = jobs
+            .values()
+            .filter(|j| j.state == JobState::Completed && j.reconciled)
+            .collect();
+        active.sort_by(|a, b| a.alias.cmp(&b.alias));
+        reusable.sort_by(|a, b| a.alias.cmp(&b.alias));
+
+        if active.is_empty() && reusable.is_empty() {
+            return None;
+        }
+
+        let mut out = String::from(
+            "### Background Job Board\n\
+             Do not poll running jobs; wait for completion. Reconcile terminal jobs before your \
+             final response.\nCompleted sessions are reusable by alias for the same specialist.\n",
+        );
+
+        if !active.is_empty() {
+            out.push_str("\n#### Active / Unreconciled\n");
+            for job in &active {
+                let state = state_label(job.state);
+                out.push_str(&format!(
+                    "- {} / {} / {} / {state}",
+                    job.alias, job.child_session, job.agent
+                ));
+                if let Some(obj) = &job.objective {
+                    out.push_str(&format!(" — Objective: {obj}"));
+                }
+                out.push('\n');
+                if let Some(summary) = &job.result_summary {
+                    out.push_str(&format!("    Result: {summary}\n"));
+                }
+            }
+        }
+
+        if !reusable.is_empty() {
+            out.push_str("\n#### Reusable Sessions\n");
+            for job in &reusable {
+                out.push_str(&format!(
+                    "- {} / {} / {} / completed\n",
+                    job.alias, job.child_session, job.agent
+                ));
+                if let Some(obj) = &job.objective {
+                    out.push_str(&format!("    Objective: {obj}\n"));
+                }
+                if !job.context_files.is_empty() {
+                    let files: Vec<&str> = job
+                        .context_files
+                        .iter()
+                        .take(CONTEXT_FILES_SHOWN)
+                        .map(|f| f.path.as_str())
+                        .collect();
+                    out.push_str(&format!("    Context read: {}\n", files.join(", ")));
+                }
+            }
+        }
+
+        Some(out)
+    }
+
+    fn get(&self, task_id: &str) -> Option<JobRecord> {
+        self.jobs.read().unwrap().get(task_id).cloned()
+    }
+
+    async fn upsert(&self, record: JobRecord) -> Result<(), StoreError> {
+        self.store.upsert_job(record.clone()).await?;
+        self.jobs
+            .write()
+            .unwrap()
+            .insert(record.task_id.clone(), record.clone());
+        self.bus.publish(AppEvent::JobUpdated {
+            job: JobRecordEvent(serde_json::to_value(&record).unwrap_or(serde_json::Value::Null)),
+        });
+        Ok(())
+    }
+
+    /// Keeps at most `max_reusable_per_agent` completed jobs per agent (LRU by `last_used_at`);
+    /// older completed jobs are dropped from the board and the store.
+    async fn trim_reusable(&self, _now: i64) -> Result<(), StoreError> {
+        let to_remove: Vec<String> = {
+            let jobs = self.jobs.read().unwrap();
+            let mut by_agent: HashMap<&str, Vec<&JobRecord>> = HashMap::new();
+            for job in jobs.values().filter(|j| j.state == JobState::Completed) {
+                by_agent.entry(job.agent.as_str()).or_default().push(job);
+            }
+            let mut remove = Vec::new();
+            for group in by_agent.values_mut() {
+                if group.len() as u32 <= self.max_reusable_per_agent {
+                    continue;
+                }
+                // Oldest last_used_at first; drop the excess from the front.
+                group.sort_by_key(|j| j.last_used_at);
+                let excess = group.len() - self.max_reusable_per_agent as usize;
+                for job in group.iter().take(excess) {
+                    remove.push(job.task_id.clone());
+                }
+            }
+            remove
+        };
+        for task_id in to_remove {
+            self.store.delete_job(task_id.clone()).await?;
+            self.jobs.write().unwrap().remove(&task_id);
+        }
+        Ok(())
+    }
+}
+
+fn state_label(state: JobState) -> &'static str {
+    match state {
+        JobState::Running => "running",
+        JobState::Completed => "completed",
+        JobState::Error => "error",
+        JobState::Cancelled => "cancelled",
+    }
+}
+
+fn truncate_summary(s: &str) -> String {
+    if s.len() <= SUMMARY_MAX {
+        return s.to_string();
+    }
+    let mut end = SUMMARY_MAX;
+    while !s.is_char_boundary(end) {
+        end -= 1;
+    }
+    format!("{}…", &s[..end])
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    async fn board(max_reusable: u32) -> (Store, JobBoard, SessionId) {
+        let store = Store::open_in_memory().unwrap();
+        let bus = EventBus::new();
+        let parent = SessionId::new();
+        let board = JobBoard::load(store.clone(), bus, &parent, max_reusable)
+            .await
+            .unwrap();
+        (store, board, parent)
+    }
+
+    fn spec(
+        task_id: &str,
+        parent: SessionId,
+        child: SessionId,
+        agent: &str,
+        objective: Option<&str>,
+    ) -> LaunchSpec {
+        LaunchSpec {
+            task_id: task_id.into(),
+            parent_session: parent,
+            child_session: child,
+            agent: agent.into(),
+            description: "d".into(),
+            objective: objective.map(Into::into),
+        }
+    }
+
+    #[tokio::test]
+    async fn alias_increments_per_agent() {
+        let (_store, board, parent) = board(2).await;
+        let a1 = board
+            .register_launch(
+                spec("t1", parent.clone(), SessionId::new(), "explorer", None),
+                1,
+            )
+            .await
+            .unwrap();
+        let a2 = board
+            .register_launch(
+                spec("t2", parent.clone(), SessionId::new(), "explorer", None),
+                2,
+            )
+            .await
+            .unwrap();
+        let f1 = board
+            .register_launch(
+                spec("t3", parent.clone(), SessionId::new(), "fixer", None),
+                3,
+            )
+            .await
+            .unwrap();
+        assert_eq!(a1, "exp-1");
+        assert_eq!(a2, "exp-2");
+        assert_eq!(f1, "fix-1");
+    }
+
+    #[tokio::test]
+    async fn finish_makes_job_reusable_and_resolvable_by_alias() {
+        let (_store, board, parent) = board(2).await;
+        let child = SessionId::new();
+        let alias = board
+            .register_launch(
+                spec(
+                    "t1",
+                    parent.clone(),
+                    child.clone(),
+                    "explorer",
+                    Some("map the auth flow"),
+                ),
+                1,
+            )
+            .await
+            .unwrap();
+        // Running jobs are not reusable.
+        assert!(board.resolve_reusable(&parent, &alias).is_none());
+        board
+            .finish("t1", JobState::Completed, Some("done".into()), 2)
+            .await
+            .unwrap();
+        let resolved = board.resolve_reusable(&parent, &alias).unwrap();
+        assert_eq!(resolved.child_session, child);
+        assert_eq!(resolved.result_summary.as_deref(), Some("done"));
+        // Also resolvable by task id.
+        assert!(board.resolve_reusable(&parent, "t1").is_some());
+    }
+
+    #[tokio::test]
+    async fn context_files_dedupe_and_respect_min_lines() {
+        let (_store, board, parent) = board(2).await;
+        board
+            .register_launch(spec("t1", parent, SessionId::new(), "explorer", None), 1)
+            .await
+            .unwrap();
+        // Below threshold — ignored.
+        board
+            .report_context_file("t1", "small.rs".into(), 3, 2)
+            .await
+            .unwrap();
+        board
+            .report_context_file("t1", "a.rs".into(), 20, 2)
+            .await
+            .unwrap();
+        // Same file again with a larger read keeps the max.
+        board
+            .report_context_file("t1", "a.rs".into(), 50, 3)
+            .await
+            .unwrap();
+        let job = board.snapshot().into_iter().next().unwrap();
+        assert_eq!(job.context_files.len(), 1);
+        assert_eq!(job.context_files[0].path, "a.rs");
+        assert_eq!(job.context_files[0].lines, 50);
+    }
+
+    #[tokio::test]
+    async fn trim_reusable_keeps_lru_within_limit() {
+        let (_store, board, parent) = board(2).await;
+        for (i, ts) in [(1, 10), (2, 20), (3, 30)] {
+            board
+                .register_launch(
+                    spec(
+                        &format!("t{i}"),
+                        parent.clone(),
+                        SessionId::new(),
+                        "explorer",
+                        None,
+                    ),
+                    ts,
+                )
+                .await
+                .unwrap();
+            board
+                .finish(&format!("t{i}"), JobState::Completed, None, ts)
+                .await
+                .unwrap();
+        }
+        // max_reusable = 2, so the oldest (t1, last_used 10) is dropped.
+        let ids: Vec<String> = board.snapshot().into_iter().map(|j| j.task_id).collect();
+        assert_eq!(ids.len(), 2);
+        assert!(!ids.contains(&"t1".to_string()));
+        assert!(ids.contains(&"t2".to_string()));
+        assert!(ids.contains(&"t3".to_string()));
+    }
+
+    #[tokio::test]
+    async fn reconcile_flips_terminal_jobs_and_moves_them_to_reusable_section() {
+        let (_store, board, parent) = board(2).await;
+        board
+            .register_launch(
+                spec("t1", parent, SessionId::new(), "explorer", Some("obj")),
+                1,
+            )
+            .await
+            .unwrap();
+        board
+            .finish("t1", JobState::Completed, Some("res".into()), 2)
+            .await
+            .unwrap();
+        // Before reconcile: appears under Active/Unreconciled.
+        let prompt = board.format_for_prompt().unwrap();
+        assert!(prompt.contains("Active / Unreconciled"));
+        board.reconcile_terminal(3).await.unwrap();
+        let prompt = board.format_for_prompt().unwrap();
+        assert!(prompt.contains("Reusable Sessions"));
+        assert!(prompt.contains("exp-1"));
+    }
+
+    #[tokio::test]
+    async fn board_reloads_persisted_jobs() {
+        let store = Store::open_in_memory().unwrap();
+        let bus = EventBus::new();
+        let parent = SessionId::new();
+        {
+            let board = JobBoard::load(store.clone(), bus.clone(), &parent, 2)
+                .await
+                .unwrap();
+            board
+                .register_launch(
+                    spec("t1", parent.clone(), SessionId::new(), "explorer", None),
+                    1,
+                )
+                .await
+                .unwrap();
+        }
+        // Fresh board over the same store sees the persisted job.
+        let board = JobBoard::load(store, bus, &parent, 2).await.unwrap();
+        assert!(!board.is_empty());
+        assert_eq!(board.snapshot().len(), 1);
+    }
+
+    #[tokio::test]
+    async fn empty_board_formats_to_none() {
+        let (_store, board, _parent) = board(2).await;
+        assert!(board.format_for_prompt().is_none());
+    }
+}

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

@@ -20,6 +20,44 @@ pub struct RunConfig {
     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>,
+    /// Whether to append the background job board to requests (primary/delegating agents).
+    pub inject_job_board: bool,
+    /// Optional user-provided reminder injected at the start of every turn (off by default).
+    pub reminder_turn_start: Option<String>,
+    /// Optional user-provided reminder injected on the turn after a file tool ran.
+    pub reminder_after_file_tool: Option<String>,
+    /// Pre-rendered "## Skills" system block advertising loadable skills. `None` = no skills.
+    pub skills_prompt: Option<String>,
+    /// The model's context window in tokens, for the auto-compaction trigger. `0` = unknown
+    /// (disables the threshold trigger; a hard `ContextOverflow` still compacts).
+    pub context_limit: u64,
+}
+
+/// 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
@@ -122,6 +160,54 @@ fn convert_message(message: &Message, parts: &[Part]) -> Vec<WireMessage> {
     }
 }
 
+/// Loads the session's messages, applies the compaction cut (drop everything before the last
+/// `Compaction` marker), and converts the surviving window to wire messages. Returns the full
+/// message list (for the "is there history to compact" check) alongside the wire window and
+/// the cut start index.
+async fn load_wire(
+    ctx: &StepContext,
+) -> Result<(Vec<Message>, Vec<WireMessage>, usize), ProviderError> {
+    let messages = ctx
+        .store
+        .messages(ctx.session_id.clone())
+        .await
+        .map_err(|e| ProviderError::Decode(e.to_string()))?;
+    let mut parts_per_message = Vec::with_capacity(messages.len());
+    for message in &messages {
+        let parts = ctx
+            .store
+            .parts(message.id.clone())
+            .await
+            .map_err(|e| ProviderError::Decode(e.to_string()))?;
+        parts_per_message.push(parts);
+    }
+    let cut_start = super::compact::last_compaction_index(&parts_per_message).unwrap_or(0);
+    let mut wire = Vec::new();
+    for (message, parts) in messages.iter().zip(&parts_per_message).skip(cut_start) {
+        wire.extend(convert_message(message, parts));
+    }
+    Ok((messages, wire, cut_start))
+}
+
+/// Summarizes the current context window and persists a compaction marker. Returns `true` if a
+/// compaction was actually written (there was history to summarize), `false` if there was
+/// nothing to compact. Requires `ctx.compactor` to be set.
+async fn do_compaction(ctx: &StepContext, now: i64) -> Result<bool, ProviderError> {
+    let Some(compactor) = ctx.compactor.clone() else {
+        return Ok(false);
+    };
+    let (messages, wire, cut_start) = load_wire(ctx).await?;
+    if !super::compact::has_history_to_compact(&messages, cut_start) {
+        return Ok(false);
+    }
+    let Some(last) = messages.last() else {
+        return Ok(false);
+    };
+    let summary = compactor.summarize(&wire).await?;
+    super::compact::write_compaction(ctx, &ctx.session_id, last.id.clone(), summary, now).await?;
+    Ok(true)
+}
+
 /// 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).
@@ -133,6 +219,8 @@ pub async fn run_session(
 ) -> RunOutcome {
     let mut doomloop = DoomLoopGuard::new();
     let mut steps = 0u32;
+    // Whether the previous step ran a file tool, gating the `after_file_tool` reminder.
+    let mut prev_used_file_tool = false;
 
     loop {
         match should_continue(&ctx.store, &ctx.session_id).await {
@@ -150,30 +238,50 @@ pub async fn run_session(
         steps += 1;
         ctx.now = now_fn();
 
-        let messages = match ctx.store.messages(ctx.session_id.clone()).await {
+        let mut wire_messages = match load_wire(&ctx).await {
-            Ok(m) => m,
+            Ok((_, wire, _)) => wire,
             Err(e) => {
                 return RunOutcome::Errored {
                     message: e.to_string(),
                 }
             }
         };
-        let mut wire_messages = Vec::new();
+
-        for message in &messages {
+        // Collect synthetic (non-persisted) blocks to append to the last user message this
-            let parts = match ctx.store.parts(message.id.clone()).await {
+        // turn: the optional turn-start reminder, the job board, and — if the previous step
-                Ok(p) => p,
+        // ran a file tool — the optional after-file-tool reminder. docs/04-multiagent.md.
-                Err(e) => {
+        let mut synthetic: Vec<String> = Vec::new();
-                    return RunOutcome::Errored {
+        if let Some(reminder) = &run_config.reminder_turn_start {
-                        message: e.to_string(),
+            synthetic.push(reminder.clone());
+        }
+        if run_config.inject_job_board {
+            if let Some(board) = &ctx.job_board {
+                if let Some(block) = board.format_for_prompt() {
+                    synthetic.push(block);
+                }
+            }
+        }
+        if prev_used_file_tool {
+            if let Some(reminder) = &run_config.reminder_after_file_tool {
+                synthetic.push(reminder.clone());
+            }
+        }
+        if !synthetic.is_empty() {
+            if let Some(last_user) = wire_messages
+                .iter_mut()
+                .rev()
+                .find(|m| m.role == WireRole::User)
+            {
+                for text in synthetic {
+                    last_user.content.push(WireContent::Text { text });
                 }
             }
-            };
-            wire_messages.extend(convert_message(message, &parts));
         }
 
         let system_blocks = system::assemble(
             system::env_header(&ctx.cwd),
             &run_config.agent_prompt,
+            run_config.skills_prompt.as_deref(),
             &run_config.instructions,
         );
         let tools: Vec<ToolSchema> = ctx
@@ -221,20 +329,71 @@ pub async fn run_session(
             &ctx,
             run_config.model.clone(),
             &run_config.agent_name,
+            run_config.cost,
             &mut doomloop,
         )
         .await;
 
         match step {
-            Ok(outcome) if outcome.aborted => return RunOutcome::Aborted,
+            Ok(outcome) if outcome.aborted => {
-            Ok(outcome) => match outcome.result {
+                accumulate_session_usage(&ctx, &outcome.usage, outcome.cost, now_fn()).await;
-                StepResult::Continue => continue,
+                return RunOutcome::Aborted;
+            }
+            Ok(outcome) => {
+                accumulate_session_usage(&ctx, &outcome.usage, outcome.cost, now_fn()).await;
+                prev_used_file_tool = outcome.used_file_tool;
+                // A completed step means the orchestrator has now seen any terminal jobs
+                // that were on the board this turn; mark them reconciled.
+                if run_config.inject_job_board {
+                    if let Some(board) = &ctx.job_board {
+                        if let Err(e) = board.reconcile_terminal(now_fn()).await {
+                            tracing::warn!(error = %e, "failed to reconcile job board");
+                        }
+                    }
+                }
+                // Proactive compaction: if the context has crossed ~90% of the window and a
+                // compactor is wired in, summarize before the next request. Best-effort — a
+                // compaction failure just means we continue with the full history.
+                let near_limit = ctx.compactor.is_some()
+                    && super::compact::over_threshold(
+                        super::compact::tokens_used(&outcome.usage),
+                        run_config.context_limit,
+                    );
+                match outcome.result {
+                    StepResult::Continue | StepResult::Compact => {
+                        if near_limit || matches!(outcome.result, StepResult::Compact) {
+                            if let Err(e) = do_compaction(&ctx, now_fn()).await {
+                                tracing::warn!(error = %e, "compaction failed; continuing with full history");
+                            }
+                        }
+                        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;
             }
+            // A hard context-window overflow is recoverable when a compactor is available:
+            // summarize and retry. If there was nothing left to compact, surface the error.
+            Err(step_err)
+                if matches!(step_err.source, ProviderError::ContextOverflow)
+                    && ctx.compactor.is_some() =>
+            {
+                match do_compaction(&ctx, now_fn()).await {
+                    Ok(true) => continue,
+                    Ok(false) => {
+                        return RunOutcome::Errored {
+                            message: step_err.source.to_string(),
+                        }
+                    }
+                    Err(e) => {
+                        return RunOutcome::Errored {
+                            message: e.to_string(),
+                        }
+                    }
+                }
+            }
             Err(step_err) => {
                 return RunOutcome::Errored {
                     message: step_err.source.to_string(),
@@ -342,11 +501,17 @@ mod tests {
             permissions,
             static_rules: Vec::new(),
             extra_rules: Arc::new(std::sync::Mutex::new(Vec::new())),
+            parent_rules: Vec::new(),
             session_id,
             cwd: cwd.clone(),
             data_dir: cwd.join("tool-output"),
             cancel: CancellationToken::new(),
             now: 1,
+            spawner: None,
+            job_board: None,
+            context_reporter: None,
+            diagnostics: None,
+            compactor: None,
         }
     }
 
@@ -422,11 +587,33 @@ mod tests {
             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()
+            }),
+            inject_job_board: false,
+            reminder_turn_start: None,
+            reminder_after_file_tool: None,
+            skills_prompt: None,
+            context_limit: 0,
         };
 
         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(),
@@ -474,6 +661,207 @@ mod tests {
         assert_eq!(final_text, "The file contains: mock file content");
     }
 
+    /// Records its calls and returns a fixed summary, so a test can assert compaction ran.
+    struct MockCompactor {
+        calls: Arc<std::sync::atomic::AtomicUsize>,
+        summary: String,
+    }
+
+    #[async_trait]
+    impl crate::engine::compact::Compactor for MockCompactor {
+        async fn summarize(&self, _messages: &[WireMessage]) -> Result<String, ProviderError> {
+            self.calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+            Ok(self.summary.clone())
+        }
+    }
+
+    async fn seed_session_with_user_text(store: &Store, model: &ModelRef, text: &str) -> SessionId {
+        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: text.into(),
+                    synthetic: false,
+                },
+            })
+            .await
+            .unwrap();
+        session_id
+    }
+
+    fn run_config_with_limit(model: ModelRef, context_limit: u64) -> RunConfig {
+        RunConfig {
+            agent_name: "orchestrator".into(),
+            agent_prompt: "You are a helpful assistant.".into(),
+            model,
+            temperature: None,
+            max_steps: 10,
+            instructions: Vec::new(),
+            cost: None,
+            inject_job_board: false,
+            reminder_turn_start: None,
+            reminder_after_file_tool: None,
+            skills_prompt: None,
+            context_limit,
+        }
+    }
+
+    async fn compaction_parts(
+        store: &Store,
+        session_id: &SessionId,
+    ) -> Vec<(crate::types::MessageId, String)> {
+        let messages = store.messages(session_id.clone()).await.unwrap();
+        let mut found = Vec::new();
+        for message in messages {
+            for part in store.parts(message.id.clone()).await.unwrap() {
+                if let PartBody::Compaction {
+                    replaces_up_to,
+                    summary,
+                } = part.body
+                {
+                    found.push((replaces_up_to, summary));
+                }
+            }
+        }
+        found
+    }
+
+    #[tokio::test]
+    async fn crossing_the_threshold_compacts_then_continues() {
+        let store = Store::open_in_memory().unwrap();
+        let bus = EventBus::new();
+        let model = ModelRef::new("mock", "mock-model");
+        let session_id = seed_session_with_user_text(&store, &model, "do a long task").await;
+
+        // Step 1: a tool-less "continue" whose usage (200) blows past 90% of a 100-token window.
+        // Step 2: a final answer, run against the compacted history.
+        let provider = MockProvider::scripted(vec![
+            vec![
+                Ok(LlmEvent::TextStart { id: "t1".into() }),
+                Ok(LlmEvent::TextDelta {
+                    id: "t1".into(),
+                    text: "working".into(),
+                }),
+                Ok(LlmEvent::TextEnd { id: "t1".into() }),
+                Ok(LlmEvent::Finish {
+                    reason: FinishReason::ToolCalls,
+                    usage: usage(200, 0),
+                }),
+            ],
+            vec![
+                Ok(LlmEvent::TextStart { id: "t2".into() }),
+                Ok(LlmEvent::TextDelta {
+                    id: "t2".into(),
+                    text: "done".into(),
+                }),
+                Ok(LlmEvent::TextEnd { id: "t2".into() }),
+                Ok(LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: usage(5, 1),
+                }),
+            ],
+        ]);
+
+        let cwd = tempfile::tempdir().unwrap();
+        let mut ctx = make_ctx(
+            store.clone(),
+            bus,
+            session_id.clone(),
+            cwd.path().to_path_buf(),
+        )
+        .await;
+        let calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        ctx.compactor = Some(Arc::new(MockCompactor {
+            calls: calls.clone(),
+            summary: "COMPACTED SUMMARY".into(),
+        }));
+
+        let run_config = run_config_with_limit(model, 100);
+        let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;
+        assert!(matches!(outcome, RunOutcome::Stopped));
+        assert_eq!(
+            calls.load(std::sync::atomic::Ordering::SeqCst),
+            1,
+            "compacted once"
+        );
+
+        let parts = compaction_parts(&store, &session_id).await;
+        assert_eq!(parts.len(), 1, "one compaction marker written");
+        assert_eq!(parts[0].1, "COMPACTED SUMMARY");
+    }
+
+    #[tokio::test]
+    async fn context_overflow_recovers_by_compacting() {
+        let store = Store::open_in_memory().unwrap();
+        let bus = EventBus::new();
+        let model = ModelRef::new("mock", "mock-model");
+        let session_id = seed_session_with_user_text(&store, &model, "hello").await;
+        // A prior assistant turn (still "in tool calls") so the loop keeps going into step 1 and
+        // there is real history to compact when the overflow hits.
+        let mut prior =
+            Message::new_assistant(session_id.clone(), model.clone(), "orchestrator", 1);
+        prior.finished = Some(FinishReason::ToolCalls);
+        store.upsert_message(prior.clone()).await.unwrap();
+        store
+            .upsert_part(Part {
+                id: crate::types::PartId::new(),
+                message_id: prior.id.clone(),
+                session_id: session_id.clone(),
+                idx: 0,
+                body: PartBody::Text {
+                    text: "earlier work".into(),
+                    synthetic: false,
+                },
+            })
+            .await
+            .unwrap();
+
+        // Step 1: hard context overflow (no output). Step 2: succeeds post-compaction.
+        let provider = MockProvider::scripted(vec![
+            vec![Err(ProviderError::ContextOverflow)],
+            vec![
+                Ok(LlmEvent::TextStart { id: "t".into() }),
+                Ok(LlmEvent::TextDelta {
+                    id: "t".into(),
+                    text: "recovered".into(),
+                }),
+                Ok(LlmEvent::TextEnd { id: "t".into() }),
+                Ok(LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: usage(1, 1),
+                }),
+            ],
+        ]);
+
+        let cwd = tempfile::tempdir().unwrap();
+        let mut ctx = make_ctx(
+            store.clone(),
+            bus,
+            session_id.clone(),
+            cwd.path().to_path_buf(),
+        )
+        .await;
+        let calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        ctx.compactor = Some(Arc::new(MockCompactor {
+            calls: calls.clone(),
+            summary: "RECAP".into(),
+        }));
+
+        let run_config = run_config_with_limit(model, 0); // threshold off; only overflow triggers
+        let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;
+        assert!(matches!(outcome, RunOutcome::Stopped));
+        assert_eq!(calls.load(std::sync::atomic::Ordering::SeqCst), 1);
+        assert_eq!(compaction_parts(&store, &session_id).await.len(), 1);
+    }
+
     #[tokio::test]
     async fn provider_error_on_first_event_is_reported_and_run_errors() {
         let store = Store::open_in_memory().unwrap();
@@ -517,6 +905,12 @@ mod tests {
             temperature: None,
             max_steps: 10,
             instructions: Vec::new(),
+            cost: None,
+            inject_job_board: false,
+            reminder_turn_start: None,
+            reminder_after_file_tool: None,
+            skills_prompt: None,
+            context_limit: 0,
         };
 
         let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;
@@ -526,4 +920,193 @@ mod tests {
         let messages = store.messages(session_id).await.unwrap();
         assert_eq!(messages.len(), 1);
     }
+
+    /// Records the last request it was asked to stream so tests can assert on prompt content.
+    struct CapturingProvider {
+        last: StdMutex<Option<LlmRequest>>,
+    }
+
+    #[async_trait]
+    impl Provider for CapturingProvider {
+        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> {
+            *self.last.lock().unwrap() = Some(req);
+            let events = vec![
+                Ok(LlmEvent::TextStart { id: "t".into() }),
+                Ok(LlmEvent::TextDelta {
+                    id: "t".into(),
+                    text: "ok".into(),
+                }),
+                Ok(LlmEvent::TextEnd { id: "t".into() }),
+                Ok(LlmEvent::Finish {
+                    reason: FinishReason::Stop,
+                    usage: usage(1, 1),
+                }),
+            ];
+            Ok(Box::pin(futures::stream::iter(events)))
+        }
+    }
+
+    #[tokio::test]
+    async fn job_board_is_injected_into_the_last_user_message() {
+        use crate::engine::jobs::{JobBoard, LaunchSpec};
+
+        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: "carry on".into(),
+                    synthetic: false,
+                },
+            })
+            .await
+            .unwrap();
+
+        // A board with one running job for this session.
+        let board = std::sync::Arc::new(
+            JobBoard::load(store.clone(), bus.clone(), &session_id, 2)
+                .await
+                .unwrap(),
+        );
+        board
+            .register_launch(
+                LaunchSpec {
+                    task_id: "t1".into(),
+                    parent_session: session_id.clone(),
+                    child_session: SessionId::new(),
+                    agent: "explorer".into(),
+                    description: "map auth".into(),
+                    objective: Some("map the auth flow".into()),
+                },
+                1,
+            )
+            .await
+            .unwrap();
+
+        let cwd = tempfile::tempdir().unwrap();
+        let mut ctx = make_ctx(store, bus, session_id, cwd.path().to_path_buf()).await;
+        ctx.job_board = Some(board);
+        let run_config = RunConfig {
+            agent_name: "orchestrator".into(),
+            agent_prompt: "You orchestrate.".into(),
+            model,
+            temperature: None,
+            max_steps: 1,
+            instructions: Vec::new(),
+            cost: None,
+            inject_job_board: true,
+            reminder_turn_start: None,
+            reminder_after_file_tool: None,
+            skills_prompt: None,
+            context_limit: 0,
+        };
+
+        let provider = std::sync::Arc::new(CapturingProvider {
+            last: StdMutex::new(None),
+        });
+        let outcome = run_session(provider.clone(), ctx, &run_config, || 2).await;
+        assert!(matches!(outcome, RunOutcome::Stopped));
+
+        let req = provider.last.lock().unwrap().clone().expect("a request");
+        let last_user = req
+            .messages
+            .iter()
+            .rev()
+            .find(|m| m.role == WireRole::User)
+            .expect("a user message");
+        let text: String = last_user
+            .content
+            .iter()
+            .filter_map(|c| match c {
+                WireContent::Text { text } => Some(text.as_str()),
+                _ => None,
+            })
+            .collect::<Vec<_>>()
+            .join("\n");
+        assert!(text.contains("Background Job Board"), "got: {text}");
+        assert!(text.contains("exp-1"), "got: {text}");
+        assert!(text.contains("map the auth flow"), "got: {text}");
+    }
+
+    #[tokio::test]
+    async fn turn_start_reminder_is_injected_into_the_request() {
+        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: "do the thing".into(),
+                    synthetic: false,
+                },
+            })
+            .await
+            .unwrap();
+
+        let cwd = tempfile::tempdir().unwrap();
+        let ctx = make_ctx(store, bus, session_id, cwd.path().to_path_buf()).await;
+        let run_config = RunConfig {
+            agent_name: "orchestrator".into(),
+            agent_prompt: "You orchestrate.".into(),
+            model,
+            temperature: None,
+            max_steps: 1,
+            instructions: Vec::new(),
+            cost: None,
+            inject_job_board: false,
+            reminder_turn_start: Some("REMEMBER: stay on task.".into()),
+            reminder_after_file_tool: None,
+            skills_prompt: None,
+            context_limit: 0,
+        };
+
+        let provider = std::sync::Arc::new(CapturingProvider {
+            last: StdMutex::new(None),
+        });
+        let outcome = run_session(provider.clone(), ctx, &run_config, || 2).await;
+        assert!(matches!(outcome, RunOutcome::Stopped));
+
+        let req = provider.last.lock().unwrap().clone().expect("a request");
+        let last_user = req
+            .messages
+            .iter()
+            .rev()
+            .find(|m| m.role == WireRole::User)
+            .expect("a user message");
+        let has_reminder = last_user.content.iter().any(
+            |c| matches!(c, WireContent::Text { text } if text.contains("REMEMBER: stay on task.")),
+        );
+        assert!(has_reminder, "turn-start reminder should be injected");
+    }
 }

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

@@ -1,10 +1,14 @@
+pub mod compact;
 pub mod doomloop;
+pub mod jobs;
 pub mod processor;
 pub mod retry;
 #[path = "loop.rs"]
 pub mod session_loop;
 pub mod system;
 
+pub use compact::Compactor;
 pub use doomloop::DoomLoopGuard;
+pub use jobs::{ContextFile, JobBoard, JobRecord, JobState};
 pub use processor::{process_step, StepContext, StepError, StepOutcome, StepResult};
 pub use session_loop::{run_session, RunConfig};

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

@@ -10,10 +10,14 @@ 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::tool::{
+    ContextReporter, MetadataSink, PermissionHandle, SubagentSpawner, Tool, ToolCtx, ToolError,
+    ToolRegistry,
+};
 use crate::types::{Message, MessageId, Part, PartBody, PartId, SessionId, TokenUsage, ToolState};
 
 use super::doomloop::DoomLoopGuard;
+use super::jobs::JobBoard;
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum StepResult {
@@ -26,7 +30,12 @@ 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,
+    /// Whether a file-mutating tool (`edit`/`write`) ran this step — drives the optional
+    /// `after_file_tool` reminder injection on the next turn.
+    pub used_file_tool: bool,
 }
 
 pub struct StepError {
@@ -45,6 +54,9 @@ pub struct StepContext {
     pub permissions: Arc<PermissionService>,
     pub static_rules: Ruleset,
     pub extra_rules: Arc<Mutex<Ruleset>>,
+    /// Parent-effective ruleset for a subagent session; empty for a root session. Enables
+    /// permission intersection on this session's tool calls.
+    pub parent_rules: Ruleset,
     pub session_id: SessionId,
     pub cwd: PathBuf,
     /// Session's `tool-output` spill directory (see `tool::truncate`).
@@ -52,6 +64,19 @@ pub struct StepContext {
     pub cancel: CancellationToken,
     /// Wall-clock for `created_at` stamps — passed in so tests stay deterministic.
     pub now: i64,
+    /// Lets the `task` tool spawn subagents. `None` disables delegation (headless/tests).
+    pub spawner: Option<Arc<dyn SubagentSpawner>>,
+    /// This session's background job board (as a parent). Injected into requests when the
+    /// running agent can delegate; `None` disables the board.
+    pub job_board: Option<Arc<JobBoard>>,
+    /// Present in subagent sessions: reports read files to this session's job on the parent
+    /// board. `None` for root sessions (nothing to report to).
+    pub context_reporter: Option<Arc<dyn ContextReporter>>,
+    /// Language-server diagnostics source shared by the session's edit/write tool calls.
+    pub diagnostics: Option<Arc<dyn crate::lsp::DiagnosticsSource>>,
+    /// Summarizes history when the context nears the model's window. `None` disables
+    /// auto-compaction (headless/tests, or when no `small_model` is configured).
+    pub compactor: Option<Arc<dyn super::compact::Compactor>>,
 }
 
 struct FlushTracker {
@@ -81,9 +106,13 @@ impl FlushTracker {
     }
 }
 
+/// Tool names that mutate files — after one runs, the optional `after_file_tool` reminder fires.
+const FILE_TOOLS: &[&str] = &["edit", "write"];
+
 struct Run<'a> {
     ctx: &'a StepContext,
     assistant: Option<Message>,
+    used_file_tool: bool,
     next_idx: u32,
     active_text: Option<PartId>,
     active_reasoning: Option<PartId>,
@@ -100,6 +129,7 @@ impl<'a> Run<'a> {
         Self {
             ctx,
             assistant: None,
+            used_file_tool: false,
             next_idx: 0,
             active_text: None,
             active_reasoning: None,
@@ -369,6 +399,9 @@ impl<'a> Run<'a> {
         input: serde_json::Value,
         doomloop: &mut DoomLoopGuard,
     ) -> Result<(), ProviderError> {
+        if FILE_TOOLS.contains(&name.as_str()) {
+            self.used_file_tool = true;
+        }
         let part_id = self.pending_tools.remove(&call_id).unwrap_or_default();
         let running = Part {
             id: part_id.clone(),
@@ -481,7 +514,8 @@ impl<'a> Run<'a> {
             self.ctx.static_rules.clone(),
             self.ctx.extra_rules.clone(),
             call_cancel.clone(),
-        );
+        )
+        .with_parent_rules(self.ctx.parent_rules.clone());
         let tool_ctx = ToolCtx {
             session_id: self.ctx.session_id.clone(),
             message_id: self.message_id(),
@@ -491,6 +525,9 @@ impl<'a> Run<'a> {
             cancel: call_cancel.clone(),
             ask,
             metadata: metadata_sink,
+            spawner: self.ctx.spawner.clone(),
+            context_reporter: self.ctx.context_reporter.clone(),
+            diagnostics: self.ctx.diagnostics.clone(),
         };
 
         let result = tokio::select! {
@@ -513,6 +550,7 @@ impl<'a> Run<'a> {
         &mut self,
         reason: FinishReason,
         usage: TokenUsage,
+        cost: f64,
     ) -> Result<StepResult, ProviderError> {
         let part = Part {
             id: PartId::new(),
@@ -521,7 +559,7 @@ impl<'a> Run<'a> {
             idx: self.next_idx,
             body: PartBody::StepFinish {
                 usage,
-                cost: 0.0,
+                cost,
                 reason: reason.clone(),
             },
         };
@@ -567,10 +605,12 @@ pub async fn process_step(
     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 {
@@ -582,7 +622,9 @@ pub async fn process_step(
                     result: StepResult::Stop,
                     message_id: run.assistant.as_ref().map(|m| m.id.clone()),
                     usage,
+                    cost: step_cost,
                     aborted: true,
+                    used_file_tool: run.used_file_tool,
                 });
             }
         };
@@ -630,7 +672,8 @@ pub async fn process_step(
                 usage: finish_usage,
             } => {
                 usage = finish_usage;
-                match run.on_finish(reason, finish_usage).await {
+                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(())
@@ -652,6 +695,8 @@ pub async fn process_step(
         result,
         message_id: run.assistant.as_ref().map(|m| m.id.clone()),
         usage,
+        cost: step_cost,
         aborted: false,
+        used_file_tool: run.used_file_tool,
     })
 }

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

@@ -25,10 +25,18 @@ pub fn env_header(cwd: &Path) -> String {
     header
 }
 
-/// Ordered system prompt blocks: environment header, agent prompt, then project
+/// Ordered system prompt blocks: environment header, agent prompt, the optional skills
-/// instructions (e.g. AGENTS.md contents). Order matches `02-engine.md`.
+/// listing, 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> {
+pub fn assemble(
+    env_header: String,
+    agent_prompt: &str,
+    skills: Option<&str>,
+    instructions: &[String],
+) -> Vec<String> {
     let mut blocks = vec![env_header, agent_prompt.to_string()];
+    if let Some(skills) = skills {
+        blocks.push(skills.to_string());
+    }
     blocks.extend(instructions.iter().cloned());
     blocks
 }
@@ -49,8 +57,20 @@ mod tests {
         let blocks = assemble(
             "ENV".to_string(),
             "AGENT",
+            None,
             &["AGENTS.md contents".to_string()],
         );
         assert_eq!(blocks, vec!["ENV", "AGENT", "AGENTS.md contents"]);
     }
+
+    #[test]
+    fn assemble_inserts_skills_after_agent_before_instructions() {
+        let blocks = assemble(
+            "ENV".to_string(),
+            "AGENT",
+            Some("SKILLS"),
+            &["INSTR".to_string()],
+        );
+        assert_eq!(blocks, vec!["ENV", "AGENT", "SKILLS", "INSTR"]);
+    }
 }

crates/harness-core/src/lib.rs

@@ -1,7 +1,9 @@
+pub mod agent;
 pub mod config;
 pub mod engine;
 pub mod event;
 pub mod llm;
+pub mod lsp;
 pub mod permission;
 pub mod store;
 pub mod tool;

crates/harness-core/src/lsp.rs

@@ -0,0 +1,46 @@
+//! Seam trait for language-server diagnostics, implemented by `harness-lsp` and consumed by
+//! the edit/write tools. Kept in core so `harness-tools` never links the LSP crate directly.
+//! See `docs/09-integrations.md`.
+
+use std::path::Path;
+use std::time::Duration;
+
+use async_trait::async_trait;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Severity {
+    Error,
+    Warning,
+    Info,
+    Hint,
+}
+
+/// One diagnostic reported by a language server. Line/character are 1-based for display.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Diagnostic {
+    pub line: u32,
+    pub character: u32,
+    pub severity: Severity,
+    pub message: String,
+    pub source: Option<String>,
+}
+
+impl Diagnostic {
+    /// `{file}: L{line}: {message}` — the one-line form appended to tool output.
+    pub fn display_line(&self, file: &str) -> String {
+        format!("{file}: L{}: {}", self.line, self.message)
+    }
+}
+
+/// A source of file diagnostics (an LSP pool). All methods are best-effort: failures are
+/// swallowed (logged) so a broken language server never fails a tool call.
+#[async_trait]
+pub trait DiagnosticsSource: Send + Sync {
+    /// Ensure a server for `path`'s language is running and told about the file's current
+    /// contents (spawn-if-needed + didOpen/didChange).
+    async fn touch(&self, path: &Path);
+
+    /// Diagnostics for `path`, waiting up to `wait` for the server to (re)publish after a
+    /// change. Returns whatever is known on timeout.
+    async fn diagnostics(&self, path: &Path, wait: Duration) -> Vec<Diagnostic>;
+}

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

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

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

@@ -44,6 +44,37 @@ pub fn evaluate(stack: &[&Ruleset], permission: &str, pattern: &str) -> Action {
     result
 }
 
+impl Action {
+    /// How restrictive this verdict is: `Deny` > `Ask` > `Allow`. Used to intersect a
+    /// parent and child verdict when a subagent runs (docs/04-multiagent.md).
+    fn restrictiveness(self) -> u8 {
+        match self {
+            Action::Allow => 0,
+            Action::Ask => 1,
+            Action::Deny => 2,
+        }
+    }
+}
+
+/// Evaluate `permission`/`pattern` against a parent-effective stack and a child stack
+/// independently, returning the **more restrictive** of the two verdicts
+/// (`deny > ask > allow`). A subagent's tool call must satisfy both the rules it inherits
+/// from its spawning chain and its own agent ruleset.
+pub fn evaluate_intersected(
+    parent_stack: &[&Ruleset],
+    child_stack: &[&Ruleset],
+    permission: &str,
+    pattern: &str,
+) -> Action {
+    let parent = evaluate(parent_stack, permission, pattern);
+    let child = evaluate(child_stack, permission, pattern);
+    if child.restrictiveness() >= parent.restrictiveness() {
+        child
+    } else {
+        parent
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -112,4 +143,49 @@ mod tests {
     fn empty_stack_defaults_to_ask() {
         assert_eq!(evaluate(&[], "bash", "ls"), Action::Ask);
     }
+
+    #[test]
+    fn intersected_takes_the_more_restrictive_verdict() {
+        let parent_allow: Ruleset = vec![rule("edit", "*", Action::Allow)];
+        let child_deny: Ruleset = vec![rule("edit", "*", Action::Deny)];
+        // Child denies what the parent would allow → deny wins.
+        assert_eq!(
+            evaluate_intersected(&[&parent_allow], &[&child_deny], "edit", "main.rs"),
+            Action::Deny
+        );
+        // Symmetric: parent denies what the child would allow → deny still wins.
+        assert_eq!(
+            evaluate_intersected(&[&child_deny], &[&parent_allow], "edit", "main.rs"),
+            Action::Deny
+        );
+    }
+
+    #[test]
+    fn intersected_ask_beats_allow_but_loses_to_deny() {
+        let allow: Ruleset = vec![rule("bash", "*", Action::Allow)];
+        let ask: Ruleset = vec![rule("bash", "*", Action::Ask)];
+        let deny: Ruleset = vec![rule("bash", "*", Action::Deny)];
+        assert_eq!(
+            evaluate_intersected(&[&allow], &[&ask], "bash", "ls"),
+            Action::Ask
+        );
+        assert_eq!(
+            evaluate_intersected(&[&ask], &[&deny], "bash", "ls"),
+            Action::Deny
+        );
+    }
+
+    #[test]
+    fn intersected_allows_only_when_both_allow() {
+        let allow: Ruleset = vec![rule("read", "*", Action::Allow)];
+        assert_eq!(
+            evaluate_intersected(&[&allow], &[&allow], "read", "src/a.rs"),
+            Action::Allow
+        );
+        // Empty child stack defaults to Ask, which is more restrictive than parent Allow.
+        assert_eq!(
+            evaluate_intersected(&[&allow], &[], "read", "src/a.rs"),
+            Action::Ask
+        );
+    }
 }

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

@@ -8,7 +8,7 @@ use ulid::Ulid;
 use crate::event::{AppEvent, EventBus, PermissionRequest};
 use crate::types::SessionId;
 
-use super::rule::{evaluate, Action, Rule, Ruleset};
+use super::rule::{evaluate, evaluate_intersected, Action, Rule, Ruleset};
 
 pub struct AskInput {
     pub permission: String,
@@ -73,6 +73,24 @@ impl PermissionService {
         }
     }
 
+    /// Like [`ask`](Self::ask), but for a subagent: the verdict is the more restrictive of
+    /// the `parent_stack` (rules inherited from the spawning chain) and `child_stack` (the
+    /// subagent's own rules). Used so a child can never widen what its parent forbids.
+    pub async fn ask_intersected(
+        &self,
+        session_id: &SessionId,
+        parent_stack: &[&Ruleset],
+        child_stack: &[&Ruleset],
+        input: AskInput,
+        cancel: &CancellationToken,
+    ) -> Result<AskDecision, AskError> {
+        match evaluate_intersected(parent_stack, child_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(

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

@@ -1,6 +1,7 @@
 use rusqlite::{params, Connection};
 use tokio::sync::oneshot;
 
+use crate::engine::jobs::JobRecord;
 use crate::types::{Message, MessageId, Part, Session, SessionId};
 
 use super::api::StoreError;
@@ -13,9 +14,13 @@ 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>>),
+    UpsertJob(JobRecord, Reply<()>),
+    DeleteJob(String, Reply<()>),
+    JobsForParent(SessionId, Reply<Vec<JobRecord>>),
 }
 
 fn init_schema(conn: &Connection) -> rusqlite::Result<()> {
@@ -69,6 +74,15 @@ 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
@@ -99,6 +113,34 @@ fn list_parts(conn: &Connection, message_id: &MessageId) -> Result<Vec<Part>, St
         .collect()
 }
 
+fn upsert_job(conn: &Connection, job: &JobRecord) -> Result<(), StoreError> {
+    let data = serde_json::to_string(job)?;
+    conn.execute(
+        "INSERT INTO job (task_id, parent_session_id, data) VALUES (?1, ?2, ?3)
+         ON CONFLICT(task_id) DO UPDATE SET data = ?3",
+        params![job.task_id, job.parent_session.as_ref(), data],
+    )?;
+    Ok(())
+}
+
+fn delete_job(conn: &Connection, task_id: &str) -> Result<(), StoreError> {
+    conn.execute("DELETE FROM job WHERE task_id = ?1", params![task_id])?;
+    Ok(())
+}
+
+fn list_jobs_for_parent(
+    conn: &Connection,
+    parent: &SessionId,
+) -> Result<Vec<JobRecord>, StoreError> {
+    let mut stmt = conn.prepare("SELECT data FROM job WHERE parent_session_id = ?1")?;
+    let rows = stmt
+        .query_map(params![parent.as_ref()], |row| row.get::<_, String>(0))?
+        .collect::<Result<Vec<_>, _>>()?;
+    rows.iter()
+        .map(|data| serde_json::from_str(data).map_err(StoreError::from))
+        .collect()
+}
+
 /// Runs on a dedicated OS thread; the async facade in `api.rs` talks to it over `mpsc`.
 pub fn run(conn: Connection, mut rx: tokio::sync::mpsc::Receiver<StoreCmd>) {
     if let Err(e) = init_schema(&conn) {
@@ -116,6 +158,9 @@ 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));
             }
@@ -125,6 +170,15 @@ pub fn run(conn: Connection, mut rx: tokio::sync::mpsc::Receiver<StoreCmd>) {
             StoreCmd::Parts(message_id, reply) => {
                 let _ = reply.send(list_parts(&conn, &message_id));
             }
+            StoreCmd::UpsertJob(job, reply) => {
+                let _ = reply.send(upsert_job(&conn, &job));
+            }
+            StoreCmd::DeleteJob(task_id, reply) => {
+                let _ = reply.send(delete_job(&conn, &task_id));
+            }
+            StoreCmd::JobsForParent(parent, reply) => {
+                let _ = reply.send(list_jobs_for_parent(&conn, &parent));
+            }
         }
     }
 }

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

@@ -3,6 +3,7 @@ use std::path::Path;
 use rusqlite::Connection;
 use tokio::sync::{mpsc, oneshot};
 
+use crate::engine::jobs::JobRecord;
 use crate::types::{Message, MessageId, Part, Session, SessionId};
 
 use super::actor::{self, StoreCmd};
@@ -77,6 +78,11 @@ 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
     }
@@ -89,6 +95,19 @@ impl Store {
     pub async fn parts(&self, message_id: MessageId) -> Result<Vec<Part>, StoreError> {
         self.call(|reply| StoreCmd::Parts(message_id, reply)).await
     }
+
+    pub async fn upsert_job(&self, job: JobRecord) -> Result<(), StoreError> {
+        self.call(|reply| StoreCmd::UpsertJob(job, reply)).await
+    }
+
+    pub async fn delete_job(&self, task_id: String) -> Result<(), StoreError> {
+        self.call(|reply| StoreCmd::DeleteJob(task_id, reply)).await
+    }
+
+    pub async fn jobs_for_parent(&self, parent: SessionId) -> Result<Vec<JobRecord>, StoreError> {
+        self.call(|reply| StoreCmd::JobsForParent(parent, reply))
+            .await
+    }
 }
 
 #[cfg(test)]

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

@@ -10,6 +10,58 @@ use tokio_util::sync::CancellationToken;
 use crate::permission::{AskDecision, AskError, AskInput, PermissionService, Ruleset};
 use crate::types::{MessageId, SessionId};
 
+/// A request from the `task` tool to run a subagent. The spawner (owned by the composition
+/// root) resolves the agent, enforces the depth limit, applies permission intersection, and
+/// runs the child session foreground or background. See `docs/04-multiagent.md`.
+pub struct SpawnRequest {
+    pub parent_session_id: SessionId,
+    pub parent_message_id: MessageId,
+    pub agent: String,
+    pub description: String,
+    pub prompt: String,
+    /// Alias or task id of a completed job to reuse (continue its child session).
+    pub reuse_task_id: Option<String>,
+    pub background: bool,
+    /// The tool call's cancellation token — used for foreground child runs. Background runs
+    /// are childed from the parent session's run token by the spawner instead.
+    pub cancel: CancellationToken,
+}
+
+#[derive(Debug)]
+pub struct SpawnOutcome {
+    pub child_session_id: SessionId,
+    pub background: bool,
+    /// Board alias assigned to a background launch.
+    pub alias: Option<String>,
+    /// Final assistant text of a foreground run.
+    pub final_text: Option<String>,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum SpawnError {
+    #[error("unknown subagent {0:?}, or it is not usable as a subagent")]
+    InvalidAgent(String),
+    #[error("subagent depth limit reached — do this work yourself instead of delegating further")]
+    DepthExceeded,
+    #[error("cannot reuse {0:?}: no completed job with that alias for this session")]
+    ReuseNotFound(String),
+    #[error("{0}")]
+    Other(String),
+}
+
+#[async_trait]
+pub trait SubagentSpawner: Send + Sync {
+    async fn spawn(&self, req: SpawnRequest) -> Result<SpawnOutcome, SpawnError>;
+}
+
+/// Lets a child session report the files it read to its job board entry, so a completed
+/// specialist advertises what it already looked at (docs/04-multiagent.md). Present only in
+/// subagent sessions; the spawner wires it to the right board + job.
+#[async_trait]
+pub trait ContextReporter: Send + Sync {
+    async fn report_file(&self, path: String, lines: u32);
+}
+
 #[derive(Debug, thiserror::Error)]
 pub enum ToolError {
     #[error("permission denied")]
@@ -60,6 +112,9 @@ pub struct PermissionHandle {
     session_id: SessionId,
     static_rules: Ruleset,
     extra_rules: Arc<Mutex<Ruleset>>,
+    /// Parent-effective ruleset for a subagent session; empty for a root session. When
+    /// non-empty, verdicts are intersected so a child can only ever be *more* restricted.
+    parent_rules: Ruleset,
     cancel: CancellationToken,
 }
 
@@ -76,10 +131,17 @@ impl PermissionHandle {
             session_id,
             static_rules,
             extra_rules,
+            parent_rules: Vec::new(),
             cancel,
         }
     }
 
+    /// Sets the parent-effective ruleset so this handle intersects verdicts (subagent runs).
+    pub fn with_parent_rules(mut self, parent_rules: Ruleset) -> Self {
+        self.parent_rules = parent_rules;
+        self
+    }
+
     pub async fn ask(
         &self,
         permission: impl Into<String>,
@@ -88,21 +150,29 @@ impl PermissionHandle {
         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 child_stack: [&Ruleset; 2] = [&self.static_rules, &extra_snapshot];
-        let decision = self
+        let input = AskInput {
-            .service
-            .ask(
-                &self.session_id,
-                &stack,
-                AskInput {
             permission: permission.into(),
             pattern: pattern.into(),
             always_pattern: always_pattern.into(),
             metadata,
-                },
+        };
+        let decision = if self.parent_rules.is_empty() {
+            self.service
+                .ask(&self.session_id, &child_stack, input, &self.cancel)
+                .await?
+        } else {
+            let parent_stack: [&Ruleset; 1] = [&self.parent_rules];
+            self.service
+                .ask_intersected(
+                    &self.session_id,
+                    &parent_stack,
+                    &child_stack,
+                    input,
                     &self.cancel,
                 )
-            .await?;
+                .await?
+        };
         if let AskDecision::AllowedAlways(rule) = decision {
             self.extra_rules.lock().unwrap().push(rule);
         }
@@ -120,6 +190,14 @@ pub struct ToolCtx {
     pub cancel: CancellationToken,
     pub ask: PermissionHandle,
     pub metadata: MetadataSink,
+    /// Present when the engine can spawn subagents (the `task` tool's capability). `None`
+    /// in headless/test contexts with no orchestration wired in.
+    pub spawner: Option<Arc<dyn SubagentSpawner>>,
+    /// Present in subagent sessions: lets the read tool report files to the job board.
+    pub context_reporter: Option<Arc<dyn ContextReporter>>,
+    /// Language-server diagnostics source (edit/write surface errors after a change). `None`
+    /// disables LSP integration.
+    pub diagnostics: Option<Arc<dyn crate::lsp::DiagnosticsSource>>,
 }
 
 #[derive(Debug)]

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-lsp/Cargo.toml

@@ -6,6 +6,13 @@ license.workspace = true
 
 [dependencies]
 harness-core = { workspace = true }
+tokio = { workspace = true }
+async-trait = { workspace = true }
+serde_json = { workspace = true }
+tracing = { workspace = true }
+
+[dev-dependencies]
+tempfile = { workspace = true }
 
 [lints]
 workspace = true

crates/harness-lsp/src/client.rs

@@ -0,0 +1,325 @@
+//! Minimal JSON-RPC client over an LSP server's child stdio: `Content-Length` framing, a
+//! request-id → oneshot map, and a diagnostics store keyed by document URI. Only the handful
+//! of methods the edit/write flow needs are implemented (docs/09-integrations.md).
+
+use std::collections::HashMap;
+use std::path::Path;
+use std::sync::atomic::{AtomicI64, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+
+use harness_core::lsp::{Diagnostic, Severity};
+use serde_json::{json, Value};
+use tokio::io::{AsyncReadExt, AsyncWriteExt, BufReader};
+use tokio::process::{Child, Command};
+use tokio::sync::{oneshot, Notify};
+
+type Pending = Arc<Mutex<HashMap<i64, oneshot::Sender<Value>>>>;
+type DiagStore = Arc<Mutex<HashMap<String, Vec<Diagnostic>>>>;
+
+/// A running language server plus the state needed to talk to it.
+pub struct LspClient {
+    outgoing: tokio::sync::mpsc::UnboundedSender<String>,
+    next_id: AtomicI64,
+    pending: Pending,
+    diagnostics: DiagStore,
+    diag_notify: Arc<Notify>,
+    // Kept alive so the child is killed on drop (`kill_on_drop`).
+    _child: Child,
+}
+
+impl LspClient {
+    /// Spawns `command args`, performs the `initialize`/`initialized` handshake rooted at
+    /// `root`, and returns a ready client.
+    pub async fn spawn(command: &str, args: &[String], root: &Path) -> std::io::Result<Self> {
+        let mut child = Command::new(command)
+            .args(args)
+            .current_dir(root)
+            .stdin(std::process::Stdio::piped())
+            .stdout(std::process::Stdio::piped())
+            .stderr(std::process::Stdio::null())
+            .kill_on_drop(true)
+            .spawn()?;
+
+        let stdin = child.stdin.take().expect("piped stdin");
+        let stdout = child.stdout.take().expect("piped stdout");
+
+        let (outgoing, mut out_rx) = tokio::sync::mpsc::unbounded_channel::<String>();
+        let pending: Pending = Arc::default();
+        let diagnostics: DiagStore = Arc::default();
+        let diag_notify = Arc::new(Notify::new());
+
+        // Writer task: frame and forward outgoing payloads.
+        tokio::spawn(async move {
+            let mut stdin = stdin;
+            while let Some(payload) = out_rx.recv().await {
+                let frame = format!("Content-Length: {}\r\n\r\n{}", payload.len(), payload);
+                if stdin.write_all(frame.as_bytes()).await.is_err() {
+                    break;
+                }
+                let _ = stdin.flush().await;
+            }
+        });
+
+        // Reader task: parse frames, route responses, collect diagnostics, ack server requests.
+        {
+            let pending = pending.clone();
+            let diagnostics = diagnostics.clone();
+            let diag_notify = diag_notify.clone();
+            let outgoing_ack = outgoing.clone();
+            tokio::spawn(async move {
+                let mut reader = BufReader::new(stdout);
+                while let Some(msg) = read_message(&mut reader).await {
+                    dispatch(msg, &pending, &diagnostics, &diag_notify, &outgoing_ack);
+                }
+            });
+        }
+
+        let client = Self {
+            outgoing,
+            next_id: AtomicI64::new(1),
+            pending,
+            diagnostics,
+            diag_notify,
+            _child: child,
+        };
+
+        client.initialize(root).await?;
+        Ok(client)
+    }
+
+    async fn initialize(&self, root: &Path) -> std::io::Result<()> {
+        let root_uri = path_to_uri(root);
+        let params = json!({
+            "processId": std::process::id(),
+            "rootUri": root_uri,
+            "workspaceFolders": [{ "uri": root_uri, "name": "root" }],
+            "capabilities": {
+                "textDocument": {
+                    "publishDiagnostics": { "relatedInformation": false }
+                }
+            },
+            "clientInfo": { "name": "ai-harness" }
+        });
+        // A slow server (rust-analyzer indexing) can take a while to answer initialize.
+        let _ = self
+            .request("initialize", params, Duration::from_secs(30))
+            .await;
+        self.notify("initialized", json!({}));
+        Ok(())
+    }
+
+    fn send(&self, payload: Value) {
+        if let Ok(text) = serde_json::to_string(&payload) {
+            let _ = self.outgoing.send(text);
+        }
+    }
+
+    pub fn notify(&self, method: &str, params: Value) {
+        self.send(json!({ "jsonrpc": "2.0", "method": method, "params": params }));
+    }
+
+    async fn request(&self, method: &str, params: Value, timeout: Duration) -> Option<Value> {
+        let id = self.next_id.fetch_add(1, Ordering::Relaxed);
+        let (tx, rx) = oneshot::channel();
+        self.pending.lock().unwrap().insert(id, tx);
+        self.send(json!({ "jsonrpc": "2.0", "id": id, "method": method, "params": params }));
+        match tokio::time::timeout(timeout, rx).await {
+            Ok(Ok(value)) => Some(value),
+            _ => {
+                self.pending.lock().unwrap().remove(&id);
+                None
+            }
+        }
+    }
+
+    pub fn did_open(&self, path: &Path, language_id: &str, version: i32, text: &str) {
+        self.notify(
+            "textDocument/didOpen",
+            json!({
+                "textDocument": {
+                    "uri": path_to_uri(path),
+                    "languageId": language_id,
+                    "version": version,
+                    "text": text,
+                }
+            }),
+        );
+    }
+
+    pub fn did_change(&self, path: &Path, version: i32, text: &str) {
+        self.notify(
+            "textDocument/didChange",
+            json!({
+                "textDocument": { "uri": path_to_uri(path), "version": version },
+                "contentChanges": [{ "text": text }],
+            }),
+        );
+    }
+
+    /// Clears the stored diagnostics for `path` so the next `wait_diagnostics` observes a fresh
+    /// publish rather than a stale one.
+    pub fn clear(&self, path: &Path) {
+        self.diagnostics.lock().unwrap().remove(&path_to_uri(path));
+    }
+
+    /// Waits up to `wait` for the server to publish diagnostics for `path`, returning whatever
+    /// is stored on timeout (possibly empty).
+    pub async fn wait_diagnostics(&self, path: &Path, wait: Duration) -> Vec<Diagnostic> {
+        let uri = path_to_uri(path);
+        let deadline = tokio::time::Instant::now() + wait;
+        loop {
+            if let Some(diags) = self.diagnostics.lock().unwrap().get(&uri) {
+                return diags.clone();
+            }
+            let notified = self.diag_notify.notified();
+            let remaining = deadline.saturating_duration_since(tokio::time::Instant::now());
+            if remaining.is_zero() {
+                break;
+            }
+            if tokio::time::timeout(remaining, notified).await.is_err() {
+                break;
+            }
+        }
+        self.diagnostics
+            .lock()
+            .unwrap()
+            .get(&uri)
+            .cloned()
+            .unwrap_or_default()
+    }
+}
+
+fn dispatch(
+    msg: Value,
+    pending: &Pending,
+    diagnostics: &DiagStore,
+    diag_notify: &Arc<Notify>,
+    outgoing: &tokio::sync::mpsc::UnboundedSender<String>,
+) {
+    let method = msg.get("method").and_then(|m| m.as_str());
+    let id = msg.get("id");
+
+    match (method, id) {
+        // Server → client request: ack with a null result so the server can proceed
+        // (e.g. client/registerCapability, window/workDoneProgress/create).
+        (Some(_), Some(id)) => {
+            let reply = json!({ "jsonrpc": "2.0", "id": id, "result": Value::Null });
+            if let Ok(text) = serde_json::to_string(&reply) {
+                let _ = outgoing.send(text);
+            }
+        }
+        // Notification from the server.
+        (Some("textDocument/publishDiagnostics"), None) => {
+            if let Some(params) = msg.get("params") {
+                if let Some((uri, diags)) = parse_diagnostics(params) {
+                    diagnostics.lock().unwrap().insert(uri, diags);
+                    diag_notify.notify_waiters();
+                }
+            }
+        }
+        (Some(_), None) => {}
+        // Response to one of our requests.
+        (None, Some(id)) => {
+            if let Some(id) = id.as_i64() {
+                if let Some(tx) = pending.lock().unwrap().remove(&id) {
+                    let result = msg.get("result").cloned().unwrap_or(Value::Null);
+                    let _ = tx.send(result);
+                }
+            }
+        }
+        (None, None) => {}
+    }
+}
+
+fn parse_diagnostics(params: &Value) -> Option<(String, Vec<Diagnostic>)> {
+    let uri = params.get("uri")?.as_str()?.to_string();
+    let items = params.get("diagnostics")?.as_array()?;
+    let diags = items
+        .iter()
+        .filter_map(|d| {
+            let start = d.get("range")?.get("start")?;
+            Some(Diagnostic {
+                line: start.get("line")?.as_u64().unwrap_or(0) as u32 + 1,
+                character: start.get("character")?.as_u64().unwrap_or(0) as u32 + 1,
+                severity: match d.get("severity").and_then(|s| s.as_u64()) {
+                    Some(1) => Severity::Error,
+                    Some(2) => Severity::Warning,
+                    Some(3) => Severity::Info,
+                    _ => Severity::Hint,
+                },
+                message: d.get("message")?.as_str().unwrap_or("").to_string(),
+                source: d
+                    .get("source")
+                    .and_then(|s| s.as_str())
+                    .map(|s| s.to_string()),
+            })
+        })
+        .collect();
+    Some((uri, diags))
+}
+
+/// Reads one `Content-Length`-framed JSON-RPC message, or `None` at EOF.
+async fn read_message<R: tokio::io::AsyncBufRead + Unpin>(reader: &mut R) -> Option<Value> {
+    use tokio::io::AsyncBufReadExt;
+    let mut content_length: Option<usize> = None;
+    loop {
+        let mut line = String::new();
+        let n = reader.read_line(&mut line).await.ok()?;
+        if n == 0 {
+            return None; // EOF
+        }
+        let trimmed = line.trim_end();
+        if trimmed.is_empty() {
+            break; // end of headers
+        }
+        if let Some(value) = trimmed.strip_prefix("Content-Length:") {
+            content_length = value.trim().parse().ok();
+        }
+    }
+    let len = content_length?;
+    let mut buf = vec![0u8; len];
+    reader.read_exact(&mut buf).await.ok()?;
+    serde_json::from_slice(&buf).ok()
+}
+
+/// `file://` URI for an absolute path (best-effort; assumes UTF-8, no percent-encoding needed
+/// for the local paths we handle).
+fn path_to_uri(path: &Path) -> String {
+    let s = path.to_string_lossy();
+    if s.starts_with('/') {
+        format!("file://{s}")
+    } else {
+        format!("file:///{s}")
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn parses_publish_diagnostics_into_one_based_positions() {
+        let params = json!({
+            "uri": "file:///tmp/a.rs",
+            "diagnostics": [{
+                "range": {"start": {"line": 4, "character": 8}, "end": {"line": 4, "character": 12}},
+                "severity": 1,
+                "message": "cannot find value `x`",
+                "source": "rustc"
+            }]
+        });
+        let (uri, diags) = parse_diagnostics(&params).unwrap();
+        assert_eq!(uri, "file:///tmp/a.rs");
+        assert_eq!(diags.len(), 1);
+        assert_eq!(diags[0].line, 5); // 0-based 4 → 1-based 5
+        assert_eq!(diags[0].character, 9);
+        assert_eq!(diags[0].severity, Severity::Error);
+        assert_eq!(diags[0].source.as_deref(), Some("rustc"));
+    }
+
+    #[test]
+    fn path_to_uri_prefixes_file_scheme() {
+        assert_eq!(path_to_uri(Path::new("/tmp/a.rs")), "file:///tmp/a.rs");
+    }
+}

crates/harness-lsp/src/lib.rs

@@ -1 +1,241 @@
-// LSP client pool + diagnostics service land here in M5.
+//! LSP diagnostics pool: maps a file extension to a language server, lazily spawns one server
+//! per language, and answers the core `DiagnosticsSource` seam. Built-in servers are only
+//! offered when their binary is on `PATH`; config can add or override them.
+//! See `docs/09-integrations.md`.
+
+mod client;
+
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+
+use async_trait::async_trait;
+use harness_core::lsp::{Diagnostic, DiagnosticsSource};
+
+use client::LspClient;
+
+/// One language server: which binary to run and which extensions it handles.
+#[derive(Debug, Clone)]
+pub struct ServerConfig {
+    pub name: String,
+    pub command: String,
+    pub args: Vec<String>,
+    pub extensions: Vec<String>,
+}
+
+/// Built-in servers, tried only when their binary exists on `PATH`.
+fn builtin_servers() -> Vec<ServerConfig> {
+    vec![
+        ServerConfig {
+            name: "rust".into(),
+            command: "rust-analyzer".into(),
+            args: vec![],
+            extensions: vec!["rs".into()],
+        },
+        ServerConfig {
+            name: "typescript".into(),
+            command: "typescript-language-server".into(),
+            args: vec!["--stdio".into()],
+            extensions: vec!["ts".into(), "tsx".into(), "js".into(), "jsx".into()],
+        },
+        ServerConfig {
+            name: "go".into(),
+            command: "gopls".into(),
+            args: vec![],
+            extensions: vec!["go".into()],
+        },
+        ServerConfig {
+            name: "python".into(),
+            command: "pyright-langserver".into(),
+            args: vec!["--stdio".into()],
+            extensions: vec!["py".into()],
+        },
+    ]
+}
+
+/// LSP `languageId` for a file extension (falls back to the extension itself).
+fn language_id_for_ext(ext: &str) -> &str {
+    match ext {
+        "rs" => "rust",
+        "ts" => "typescript",
+        "tsx" => "typescriptreact",
+        "js" => "javascript",
+        "jsx" => "javascriptreact",
+        "go" => "go",
+        "py" => "python",
+        other => other,
+    }
+}
+
+/// Returns true if `command` is an absolute existing file or resolvable on `PATH`.
+fn binary_exists(command: &str) -> bool {
+    let p = Path::new(command);
+    if p.is_absolute() {
+        return p.is_file();
+    }
+    let Some(paths) = std::env::var_os("PATH") else {
+        return false;
+    };
+    std::env::split_paths(&paths).any(|dir| dir.join(command).is_file())
+}
+
+enum Slot {
+    Ready(Arc<LspClient>),
+    Failed,
+}
+
+pub struct LspPool {
+    root: PathBuf,
+    servers: Vec<ServerConfig>,
+    /// One slot per server name; absent until first spawn attempt.
+    clients: tokio::sync::Mutex<HashMap<String, Slot>>,
+    /// Open documents → last-sent version, so `didChange` bumps monotonically.
+    open: Mutex<HashMap<PathBuf, i32>>,
+}
+
+impl LspPool {
+    /// Builds a pool: available built-ins plus any `config` servers (which override built-ins
+    /// by name; an empty command disables a built-in).
+    pub fn new(root: PathBuf, config: Vec<ServerConfig>) -> Self {
+        let mut by_name: HashMap<String, ServerConfig> = HashMap::new();
+        for server in builtin_servers() {
+            if binary_exists(&server.command) {
+                by_name.insert(server.name.clone(), server);
+            }
+        }
+        for server in config {
+            if server.command.is_empty() {
+                by_name.remove(&server.name);
+            } else {
+                by_name.insert(server.name.clone(), server);
+            }
+        }
+        Self {
+            root,
+            servers: by_name.into_values().collect(),
+            clients: tokio::sync::Mutex::new(HashMap::new()),
+            open: Mutex::new(HashMap::new()),
+        }
+    }
+
+    fn server_for(&self, path: &Path) -> Option<&ServerConfig> {
+        let ext = path.extension().and_then(|e| e.to_str())?;
+        self.servers
+            .iter()
+            .find(|s| s.extensions.iter().any(|e| e == ext))
+    }
+
+    /// Gets or lazily spawns the client for `server`. Caches a failure so we don't respawn a
+    /// broken server on every edit.
+    async fn client_for(&self, server: &ServerConfig) -> Option<Arc<LspClient>> {
+        let mut clients = self.clients.lock().await;
+        match clients.get(&server.name) {
+            Some(Slot::Ready(c)) => return Some(c.clone()),
+            Some(Slot::Failed) => return None,
+            None => {}
+        }
+        match LspClient::spawn(&server.command, &server.args, &self.root).await {
+            Ok(client) => {
+                let client = Arc::new(client);
+                clients.insert(server.name.clone(), Slot::Ready(client.clone()));
+                Some(client)
+            }
+            Err(e) => {
+                tracing::warn!(server = %server.name, error = %e, "LSP server failed to start");
+                clients.insert(server.name.clone(), Slot::Failed);
+                None
+            }
+        }
+    }
+}
+
+#[async_trait]
+impl DiagnosticsSource for LspPool {
+    async fn touch(&self, path: &Path) {
+        let Some(server) = self.server_for(path).cloned() else {
+            return;
+        };
+        let Ok(text) = tokio::fs::read_to_string(path).await else {
+            return;
+        };
+        let Some(client) = self.client_for(&server).await else {
+            return;
+        };
+        let ext = path
+            .extension()
+            .and_then(|e| e.to_str())
+            .unwrap_or_default();
+        let language_id = language_id_for_ext(ext);
+
+        let existing_version = {
+            let mut open = self.open.lock().unwrap();
+            match open.get_mut(path) {
+                Some(version) => {
+                    *version += 1;
+                    Some(*version)
+                }
+                None => {
+                    open.insert(path.to_path_buf(), 1);
+                    None
+                }
+            }
+        };
+
+        // Fresh diagnostics only reflect the current contents — drop any stale set first.
+        client.clear(path);
+        match existing_version {
+            None => client.did_open(path, language_id, 1, &text),
+            Some(version) => client.did_change(path, version, &text),
+        }
+    }
+
+    async fn diagnostics(&self, path: &Path, wait: Duration) -> Vec<Diagnostic> {
+        let Some(server) = self.server_for(path).cloned() else {
+            return Vec::new();
+        };
+        let Some(client) = self.client_for(&server).await else {
+            return Vec::new();
+        };
+        client.wait_diagnostics(path, wait).await
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn language_ids_map_common_extensions() {
+        assert_eq!(language_id_for_ext("rs"), "rust");
+        assert_eq!(language_id_for_ext("tsx"), "typescriptreact");
+        assert_eq!(language_id_for_ext("unknown"), "unknown");
+    }
+
+    #[test]
+    fn config_server_overrides_builtin_by_name() {
+        let pool = LspPool::new(
+            std::env::temp_dir(),
+            vec![ServerConfig {
+                name: "rust".into(),
+                command: "my-custom-ra".into(),
+                args: vec!["--flag".into()],
+                extensions: vec!["rs".into()],
+            }],
+        );
+        let server = pool.server_for(Path::new("/x/a.rs")).unwrap();
+        assert_eq!(server.command, "my-custom-ra");
+    }
+
+    #[test]
+    fn no_server_for_unknown_extension() {
+        let pool = LspPool::new(std::env::temp_dir(), vec![]);
+        assert!(pool.server_for(Path::new("/x/a.zzz")).is_none());
+    }
+
+    #[test]
+    fn binary_exists_finds_a_known_tool() {
+        assert!(binary_exists("sh"));
+        assert!(!binary_exists("definitely-not-a-real-binary-xyz"));
+    }
+}

crates/harness-lsp/tests/rust_analyzer.rs

@@ -0,0 +1,38 @@
+//! End-to-end LSP test against a real `rust-analyzer`. Ignored by default (requires the binary
+//! on PATH and is slow — RA indexes the project). Run with:
+//!   cargo test -p harness-lsp --test rust_analyzer -- --ignored --nocapture
+
+use std::time::Duration;
+
+use harness_core::lsp::{DiagnosticsSource, Severity};
+use harness_lsp::LspPool;
+
+#[tokio::test]
+#[ignore = "requires rust-analyzer on PATH; slow"]
+async fn rust_analyzer_reports_a_type_error() {
+    // Minimal cargo project with a deliberate type error.
+    let dir = tempfile::tempdir().unwrap();
+    std::fs::write(
+        dir.path().join("Cargo.toml"),
+        "[package]\nname = \"probe\"\nversion = \"0.1.0\"\nedition = \"2021\"\n",
+    )
+    .unwrap();
+    std::fs::create_dir_all(dir.path().join("src")).unwrap();
+    let main_rs = dir.path().join("src/main.rs");
+    std::fs::write(
+        &main_rs,
+        "fn main() {\n    let x: i32 = \"not an integer\";\n    let _ = x;\n}\n",
+    )
+    .unwrap();
+
+    let pool = LspPool::new(dir.path().to_path_buf(), vec![]);
+    pool.touch(&main_rs).await;
+    // Generous wait: rust-analyzer must index the workspace before it reports anything.
+    let diags = pool.diagnostics(&main_rs, Duration::from_secs(60)).await;
+
+    println!("diagnostics: {diags:#?}");
+    assert!(
+        diags.iter().any(|d| d.severity == Severity::Error),
+        "expected at least one error diagnostic, got: {diags:?}"
+    );
+}

crates/harness-mcp/Cargo.toml

@@ -6,6 +6,17 @@ license.workspace = true
 
 [dependencies]
 harness-core = { workspace = true }
+rmcp = { workspace = true, features = ["client", "transport-child-process"] }
+tokio = { workspace = true }
+async-trait = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+tracing = { workspace = true }
+thiserror = { workspace = true }
+
+[dev-dependencies]
+tokio-util = { workspace = true }
+tempfile = { workspace = true }
 
 [lints]
 workspace = true

crates/harness-mcp/src/lib.rs

@@ -1 +1,264 @@
-// MCP stdio client → Tool adapters land here in M5.
+//! MCP stdio client → `Tool` adapters (M5). For each configured server we spawn the child
+//! over rmcp's `TokioChildProcess` transport, `initialize`, `list_tools`, and wrap every
+//! remote tool as an [`McpTool`] named `{server}_{tool}`. Servers are gated behind the `mcp`
+//! permission key. See `docs/09-integrations.md`.
+//!
+//! Out of scope for v1 (matching the doc): resources, prompts, sampling, and non-stdio
+//! transports.
+
+use std::collections::HashMap;
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use rmcp::model::{CallToolRequestParam, RawContent};
+use rmcp::service::RunningService;
+use rmcp::transport::TokioChildProcess;
+use rmcp::{RoleClient, ServiceExt};
+use tokio::process::Command;
+
+/// Sanitized-name cap so a `{server}_{tool}` name stays a legal tool identifier.
+const MAX_NAME_LEN: usize = 64;
+
+/// One configured MCP server: the child command plus its environment.
+#[derive(Debug, Clone)]
+pub struct ServerConfig {
+    pub command: String,
+    pub args: Vec<String>,
+    pub env: HashMap<String, String>,
+}
+
+#[derive(Debug, thiserror::Error)]
+enum ConnectError {
+    #[error("spawn/transport failed: {0}")]
+    Transport(#[from] std::io::Error),
+    #[error("MCP service error: {0}")]
+    Service(#[from] rmcp::service::ServiceError),
+}
+
+/// Connects every configured server and returns ready tool adapters. A server that fails to
+/// start (or list its tools) is logged and skipped — the rest are unaffected, and the session
+/// still runs with whatever connected. `named_servers` is `{server_name: config}`.
+pub async fn connect_all(named_servers: HashMap<String, ServerConfig>) -> Vec<Arc<dyn Tool>> {
+    let mut tools: Vec<Arc<dyn Tool>> = Vec::new();
+    for (name, config) in named_servers {
+        match connect(&name, &config).await {
+            Ok(mut server_tools) => {
+                tracing::info!(server = %name, count = server_tools.len(), "MCP server connected");
+                tools.append(&mut server_tools);
+            }
+            Err(e) => {
+                tracing::warn!(server = %name, error = %e, "MCP server failed to start; skipping");
+            }
+        }
+    }
+    tools
+}
+
+async fn connect(name: &str, config: &ServerConfig) -> Result<Vec<Arc<dyn Tool>>, ConnectError> {
+    let mut command = Command::new(&config.command);
+    command.args(&config.args);
+    for (key, value) in &config.env {
+        command.env(key, value);
+    }
+    // rmcp sets stdin/stdout to piped and kill-on-drop; the child dies with the `RunningService`.
+    let transport = TokioChildProcess::new(&mut command)?;
+    let service = Arc::new(().serve(transport).await?);
+
+    let remote_tools = service.peer().list_all_tools().await?;
+    let adapters = remote_tools
+        .into_iter()
+        .map(|tool| {
+            let full_name = qualified_name(name, &tool.name);
+            let parameters = serde_json::Value::Object((*tool.input_schema).clone());
+            Arc::new(McpTool {
+                full_name,
+                remote_name: tool.name.to_string(),
+                description: tool.description.to_string(),
+                parameters,
+                service: service.clone(),
+            }) as Arc<dyn Tool>
+        })
+        .collect();
+    Ok(adapters)
+}
+
+/// `{server}_{tool}` sanitized to `[a-zA-Z0-9_-]` and capped at [`MAX_NAME_LEN`] chars.
+fn qualified_name(server: &str, tool: &str) -> String {
+    let mut name: String = format!("{server}_{tool}")
+        .chars()
+        .map(|c| {
+            if c.is_ascii_alphanumeric() || c == '_' || c == '-' {
+                c
+            } else {
+                '_'
+            }
+        })
+        .collect();
+    name.truncate(MAX_NAME_LEN);
+    name
+}
+
+/// A single remote MCP tool exposed to the engine as a `Tool`. Holds a shared handle to the
+/// server's `RunningService` (kept alive for the whole session so the child stays up).
+struct McpTool {
+    /// Engine-facing name: sanitized `{server}_{tool}`; also the permission pattern.
+    full_name: String,
+    /// The server's own tool name, sent back verbatim in `call_tool`.
+    remote_name: String,
+    description: String,
+    parameters: serde_json::Value,
+    service: Arc<RunningService<RoleClient, ()>>,
+}
+
+#[async_trait]
+impl Tool for McpTool {
+    fn name(&self) -> &str {
+        &self.full_name
+    }
+
+    fn description(&self) -> &str {
+        &self.description
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        self.parameters.clone()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        ctx.ask
+            .ask(
+                "mcp",
+                self.full_name.clone(),
+                self.full_name.clone(),
+                input.clone(),
+            )
+            .await?;
+
+        let arguments = match input {
+            serde_json::Value::Object(map) => Some(map),
+            serde_json::Value::Null => None,
+            other => {
+                return Err(ToolError::Invalid(format!(
+                    "MCP tool arguments must be a JSON object, got {other}"
+                )))
+            }
+        };
+
+        let result = self
+            .service
+            .peer()
+            .call_tool(CallToolRequestParam {
+                name: self.remote_name.clone().into(),
+                arguments,
+            })
+            .await
+            .map_err(|e| ToolError::Other(e.to_string()))?;
+
+        let mut text = String::new();
+        let mut image_index = 0;
+        for content in &result.content {
+            match &content.raw {
+                RawContent::Text(t) => {
+                    if !text.is_empty() {
+                        text.push('\n');
+                    }
+                    text.push_str(&t.text);
+                }
+                RawContent::Image(image) => {
+                    let note = save_image(&ctx.data_dir, &self.full_name, image_index, image).await;
+                    if !text.is_empty() {
+                        text.push('\n');
+                    }
+                    text.push_str(&note);
+                    image_index += 1;
+                }
+                RawContent::Resource(resource) => {
+                    let embedded = resource_text(resource);
+                    if !embedded.is_empty() {
+                        if !text.is_empty() {
+                            text.push('\n');
+                        }
+                        text.push_str(&embedded);
+                    }
+                }
+            }
+        }
+
+        // MCP surfaces tool-level failures as `is_error` with the message in `content`; map
+        // that to a tool error so the model sees it as a failed call rather than a result.
+        if result.is_error.unwrap_or(false) {
+            return Err(ToolError::Other(if text.is_empty() {
+                "MCP tool reported an error".to_string()
+            } else {
+                text
+            }));
+        }
+
+        Ok(ToolOutput::new(self.full_name.clone(), text))
+    }
+}
+
+/// Writes an image payload to the session data dir and returns a one-line note for the tool
+/// output. Best-effort: a write failure still yields a note (without a path).
+async fn save_image(
+    data_dir: &PathBuf,
+    tool_name: &str,
+    index: usize,
+    image: &rmcp::model::RawImageContent,
+) -> String {
+    let ext = image.mime_type.rsplit('/').next().unwrap_or("bin");
+    let file_name = format!("{tool_name}-image-{index}.{ext}.b64");
+    let path = data_dir.join(&file_name);
+    let saved = tokio::fs::create_dir_all(data_dir).await.is_ok()
+        && tokio::fs::write(&path, &image.data).await.is_ok();
+    if saved {
+        format!(
+            "[image: {} ({} base64 bytes) saved to {}]",
+            image.mime_type,
+            image.data.len(),
+            path.display()
+        )
+    } else {
+        format!(
+            "[image: {} ({} base64 bytes, not saved)]",
+            image.mime_type,
+            image.data.len()
+        )
+    }
+}
+
+/// Best-effort text extraction from an embedded resource (text resources only in v1).
+fn resource_text(resource: &rmcp::model::RawEmbeddedResource) -> String {
+    match &resource.resource {
+        rmcp::model::ResourceContents::TextResourceContents { text, .. } => text.clone(),
+        _ => String::new(),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn qualified_name_prefixes_and_sanitizes() {
+        assert_eq!(qualified_name("fs", "read_file"), "fs_read_file");
+        assert_eq!(
+            qualified_name("my.server", "do/thing"),
+            "my_server_do_thing"
+        );
+    }
+
+    #[test]
+    fn qualified_name_caps_length() {
+        let long_tool = "t".repeat(100);
+        let name = qualified_name("srv", &long_tool);
+        assert_eq!(name.len(), MAX_NAME_LEN);
+        assert!(name.starts_with("srv_t"));
+    }
+}

crates/harness-mcp/tests/echo.rs

@@ -0,0 +1,153 @@
+//! End-to-end MCP integration test: spawn a real stdio MCP server (a small Python fixture),
+//! connect through the real rmcp client, and verify a discovered tool is callable and gated
+//! behind the `mcp` permission key. This is the M5 milestone's ✅ for MCP.
+
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex};
+
+use harness_core::event::{AppEvent, EventBus};
+use harness_core::permission::{PermissionReply, PermissionService};
+use harness_core::tool::{MetadataSink, PermissionHandle, ToolCtx, ToolError};
+use harness_core::types::{MessageId, SessionId};
+use harness_mcp::{connect_all, ServerConfig};
+use tokio_util::sync::CancellationToken;
+
+/// A permission frontend that replies `Once` to every ask and records the `permission`/`pattern`
+/// of each, so a test can assert the call was actually gated.
+fn recording_auto_approve(
+    bus: EventBus,
+    service: Arc<PermissionService>,
+) -> Arc<Mutex<Vec<(String, String)>>> {
+    let asks = Arc::new(Mutex::new(Vec::new()));
+    let asks_task = asks.clone();
+    // Subscribe before spawning: a subscription created inside the task could miss the ask
+    // (tokio broadcast only delivers to receivers that exist at publish time).
+    let mut rx = bus.subscribe();
+    tokio::spawn(async move {
+        while let Ok(event) = rx.recv().await {
+            if let AppEvent::PermissionAsked { request } = event {
+                asks_task
+                    .lock()
+                    .unwrap()
+                    .push((request.permission.clone(), request.pattern.clone()));
+                service.reply(&request.id, PermissionReply::Once);
+            }
+        }
+    });
+    asks
+}
+
+struct Harness {
+    ctx_data_dir: std::path::PathBuf,
+    service: Arc<PermissionService>,
+}
+
+impl Harness {
+    fn ctx(&self) -> ToolCtx {
+        let (metadata, _rx) = MetadataSink::channel();
+        ToolCtx {
+            session_id: SessionId::new(),
+            message_id: MessageId::new(),
+            call_id: "call_1".into(),
+            data_dir: self.ctx_data_dir.clone(),
+            cwd: std::env::temp_dir(),
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                self.service.clone(),
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
+        }
+    }
+}
+
+fn fixture_server() -> HashMap<String, ServerConfig> {
+    let script = concat!(env!("CARGO_MANIFEST_DIR"), "/tests/fixtures/echo_server.py");
+    HashMap::from([(
+        "fix".to_string(),
+        ServerConfig {
+            command: "python3".to_string(),
+            args: vec![script.to_string()],
+            env: HashMap::new(),
+        },
+    )])
+}
+
+#[tokio::test]
+async fn discovers_and_calls_a_real_mcp_tool_with_permission() {
+    let tools = connect_all(fixture_server()).await;
+    let names: Vec<_> = tools.iter().map(|t| t.name().to_string()).collect();
+    assert!(
+        names.contains(&"fix_echo".to_string()),
+        "expected fix_echo among {names:?}"
+    );
+    assert!(names.contains(&"fix_boom".to_string()));
+
+    let echo = tools.iter().find(|t| t.name() == "fix_echo").unwrap();
+    // Schema passes through untouched from the server.
+    assert_eq!(echo.parameters()["properties"]["text"]["type"], "string");
+
+    let bus = EventBus::new();
+    let service = Arc::new(PermissionService::new(bus.clone()));
+    let asks = recording_auto_approve(bus, service.clone());
+    let dir = tempfile::tempdir().unwrap();
+    let harness = Harness {
+        ctx_data_dir: dir.path().to_path_buf(),
+        service,
+    };
+
+    let out = echo
+        .execute(serde_json::json!({"text": "hi there"}), harness.ctx())
+        .await
+        .expect("echo call succeeds");
+    assert_eq!(out.output, "hi there");
+
+    // The call was gated on the `mcp` key with the qualified tool name as the pattern.
+    let recorded = asks.lock().unwrap().clone();
+    assert_eq!(recorded, vec![("mcp".to_string(), "fix_echo".to_string())]);
+}
+
+#[tokio::test]
+async fn tool_error_result_maps_to_tool_error() {
+    let tools = connect_all(fixture_server()).await;
+    let boom = tools.iter().find(|t| t.name() == "fix_boom").unwrap();
+
+    let bus = EventBus::new();
+    let service = Arc::new(PermissionService::new(bus.clone()));
+    let _asks = recording_auto_approve(bus, service.clone());
+    let dir = tempfile::tempdir().unwrap();
+    let harness = Harness {
+        ctx_data_dir: dir.path().to_path_buf(),
+        service,
+    };
+
+    let err = boom
+        .execute(serde_json::json!({}), harness.ctx())
+        .await
+        .expect_err("boom reports an error result");
+    match err {
+        ToolError::Other(msg) => assert_eq!(msg, "kaboom"),
+        other => panic!("expected ToolError::Other, got {other:?}"),
+    }
+}
+
+#[tokio::test]
+async fn a_failed_server_is_skipped_not_fatal() {
+    let servers = HashMap::from([(
+        "broken".to_string(),
+        ServerConfig {
+            command: "definitely-not-a-real-binary-xyz".to_string(),
+            args: vec![],
+            env: HashMap::new(),
+        },
+    )]);
+    // No panic, no tools — the missing server is logged and skipped.
+    let tools = connect_all(servers).await;
+    assert!(tools.is_empty());
+}

crates/harness-mcp/tests/fixtures/echo_server.py

@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+"""Minimal MCP stdio server fixture for harness-mcp integration tests.
+
+Speaks newline-delimited JSON-RPC (the framing rmcp's child-process transport uses) and
+implements just enough of the protocol to be discovered and called: `initialize`,
+`notifications/initialized`, `tools/list`, and `tools/call`. Exposes one tool, `echo`,
+which returns its `text` argument, plus `boom`, which returns an error result.
+"""
+import json
+import sys
+
+PROTOCOL_VERSION = "2024-11-05"
+
+TOOLS = [
+    {
+        "name": "echo",
+        "description": "Returns the text it is given.",
+        "inputSchema": {
+            "type": "object",
+            "properties": {"text": {"type": "string"}},
+            "required": ["text"],
+        },
+    },
+    {
+        "name": "boom",
+        "description": "Always fails.",
+        "inputSchema": {"type": "object", "properties": {}},
+    },
+]
+
+
+def reply(msg_id, result):
+    sys.stdout.write(json.dumps({"jsonrpc": "2.0", "id": msg_id, "result": result}) + "\n")
+    sys.stdout.flush()
+
+
+def main():
+    # readline() rather than `for line in sys.stdin`: the latter's read-ahead buffer blocks
+    # until it fills, which would stall the JSON-RPC handshake line-by-line.
+    while True:
+        line = sys.stdin.readline()
+        if line == "":  # EOF: parent closed stdin
+            break
+        line = line.strip()
+        if not line:
+            continue
+        msg = json.loads(line)
+        method = msg.get("method")
+        msg_id = msg.get("id")
+
+        if method == "initialize":
+            reply(msg_id, {
+                "protocolVersion": PROTOCOL_VERSION,
+                "capabilities": {"tools": {}},
+                "serverInfo": {"name": "echo-fixture", "version": "0.1.0"},
+            })
+        elif method == "notifications/initialized":
+            pass  # notification: no response
+        elif method == "tools/list":
+            reply(msg_id, {"tools": TOOLS})
+        elif method == "tools/call":
+            params = msg.get("params") or {}
+            name = params.get("name")
+            args = params.get("arguments") or {}
+            if name == "echo":
+                reply(msg_id, {
+                    "content": [{"type": "text", "text": args.get("text", "")}],
+                    "isError": False,
+                })
+            elif name == "boom":
+                reply(msg_id, {
+                    "content": [{"type": "text", "text": "kaboom"}],
+                    "isError": True,
+                })
+            else:
+                reply(msg_id, {
+                    "content": [{"type": "text", "text": f"unknown tool {name}"}],
+                    "isError": True,
+                })
+        elif msg_id is not None:
+            # Unknown request: empty result keeps the client happy.
+            reply(msg_id, {})
+
+
+if __name__ == "__main__":
+    main()

crates/harness-providers/Cargo.toml

@@ -18,9 +18,11 @@ 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/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/mod.rs

@@ -1 +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

@@ -1,6 +1,14 @@
 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 +1,220 @@
+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";
+/// OpenCode Zen's OpenAI-compatible gateway (chat-completions only).
+const OPENCODE_BASE_URL: &str = "https://opencode.ai/zen/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 {
+    id: String,
+    api_key: String,
+    base_url: String,
+    /// When set, always speak chat-completions regardless of model name — required for
+    /// OpenAI-compatible gateways (e.g. OpenCode Zen) that don't implement `/responses`.
+    chat_only: bool,
+    client: reqwest::Client,
+}
+
+impl OpenAiProvider {
+    pub fn new(api_key: impl Into<String>) -> Self {
+        Self {
+            id: "openai".to_string(),
+            api_key: api_key.into(),
+            base_url: DEFAULT_BASE_URL.to_string(),
+            chat_only: false,
+            client: reqwest::Client::new(),
+        }
+    }
+
+    pub fn with_base_url(api_key: impl Into<String>, base_url: impl Into<String>) -> Self {
+        Self {
+            id: "openai".to_string(),
+            api_key: api_key.into(),
+            base_url: base_url.into(),
+            chat_only: false,
+            client: reqwest::Client::new(),
+        }
+    }
+
+    /// An OpenCode Zen provider: id `opencode`, chat-completions only, defaulting to Zen's
+    /// gateway. Pass `base_url = None` to use the default endpoint.
+    pub fn opencode(api_key: impl Into<String>, base_url: Option<String>) -> Self {
+        Self {
+            id: "opencode".to_string(),
+            api_key: api_key.into(),
+            base_url: base_url.unwrap_or_else(|| OPENCODE_BASE_URL.to_string()),
+            chat_only: true,
+            client: reqwest::Client::new(),
+        }
+    }
+
+    /// The wire format for `model`, honoring `chat_only`.
+    fn flavor(&self, model: &str) -> ApiFlavor {
+        if self.chat_only {
+            ApiFlavor::Chat
+        } else {
+            flavor_for(model)
+        }
+    }
+
+    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 {
+        &self.id
+    }
+
+    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 self.flavor(&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 self.flavor(&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 opencode_provider_is_chat_only_and_ids_as_opencode() {
+        let p = OpenAiProvider::opencode("k", None);
+        assert_eq!(p.id(), "opencode");
+        assert_eq!(p.base_url, OPENCODE_BASE_URL);
+        // Even a gpt-*/o* model must route to chat completions on a chat-only gateway.
+        assert_eq!(p.flavor("gpt-5.5"), ApiFlavor::Chat);
+        assert_eq!(p.flavor("claude-sonnet-5"), ApiFlavor::Chat);
+    }
+
+    #[test]
+    fn opencode_honors_custom_base_url() {
+        let p = OpenAiProvider::opencode("k", Some("https://example.test/v1".into()));
+        assert_eq!(p.base_url, "https://example.test/v1");
+    }
+
+    #[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-tools/src/bash.rs

@@ -149,6 +149,9 @@ mod tests {
                 CancellationToken::new(),
             ),
             metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
         }
     }
 

crates/harness-tools/src/diagnostics.rs

@@ -0,0 +1,73 @@
+//! Shared LSP-diagnostics reporting for the edit/write tools. After a file is written we ask
+//! the (optional) diagnostics source to re-analyze it and append any error-severity items to
+//! the tool output so the model sees mistakes it just introduced. Best-effort: no source, a
+//! slow server, or a timeout all just mean "no diagnostics" — never a tool failure.
+
+use std::path::Path;
+use std::time::Duration;
+
+use harness_core::lsp::Severity;
+use harness_core::tool::{ToolCtx, ToolOutput};
+
+/// docs/09-integrations.md: wait up to 1.5s for the server to (re)publish after the change.
+const DIAGNOSTICS_WAIT: Duration = Duration::from_millis(1500);
+
+/// Touches `path` in the language server and appends error-severity diagnostics to `output`
+/// (both as a human-readable block in the text and the full set in metadata under `diagnostics`).
+pub async fn append_diagnostics(
+    ctx: &ToolCtx,
+    path: &Path,
+    display_name: &str,
+    output: &mut ToolOutput,
+) {
+    let Some(source) = &ctx.diagnostics else {
+        return;
+    };
+    source.touch(path).await;
+    let diagnostics = source.diagnostics(path, DIAGNOSTICS_WAIT).await;
+    if diagnostics.is_empty() {
+        return;
+    }
+
+    let errors: Vec<_> = diagnostics
+        .iter()
+        .filter(|d| d.severity == Severity::Error)
+        .collect();
+    if !errors.is_empty() {
+        output
+            .output
+            .push_str("\n\nLSP errors detected in this file, please fix:");
+        for diag in &errors {
+            output.output.push('\n');
+            output.output.push_str(&diag.display_line(display_name));
+        }
+    }
+
+    // Full set (all severities) into metadata for the TUI.
+    let items: Vec<serde_json::Value> = diagnostics
+        .iter()
+        .map(|d| {
+            serde_json::json!({
+                "line": d.line,
+                "character": d.character,
+                "severity": severity_str(d.severity),
+                "message": d.message,
+                "source": d.source,
+            })
+        })
+        .collect();
+    if let serde_json::Value::Object(map) = &mut output.metadata {
+        map.insert("diagnostics".into(), serde_json::Value::Array(items));
+    } else {
+        output.metadata = serde_json::json!({ "diagnostics": items });
+    }
+}
+
+fn severity_str(severity: Severity) -> &'static str {
+    match severity {
+        Severity::Error => "error",
+        Severity::Warning => "warning",
+        Severity::Info => "info",
+        Severity::Hint => "hint",
+    }
+}

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

@@ -196,8 +196,9 @@ impl Tool for EditTool {
             .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());
+        let mut output = ToolOutput::new(pattern.clone(), "Edit applied successfully.".to_string());
         output.metadata = serde_json::json!({"diff": diff, "added": added, "removed": removed});
+        crate::diagnostics::append_diagnostics(&ctx, &path, &pattern, &mut output).await;
         Ok(output)
     }
 }
@@ -232,6 +233,9 @@ mod tests {
                 CancellationToken::new(),
             ),
             metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
         }
     }
 

crates/harness-tools/src/glob.rs

@@ -130,6 +130,9 @@ mod tests {
                 CancellationToken::new(),
             ),
             metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
         }
     }
 

crates/harness-tools/src/grep.rs

@@ -159,6 +159,9 @@ mod tests {
                 CancellationToken::new(),
             ),
             metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
         }
     }
 

crates/harness-tools/src/lib.rs

@@ -1,9 +1,12 @@
 mod bash;
+mod diagnostics;
 mod edit;
 mod glob;
 mod grep;
 mod paths;
 mod read;
+mod skill;
+mod task;
 mod write;
 
 pub use bash::BashTool;
@@ -11,6 +14,8 @@ pub use edit::EditTool;
 pub use glob::GlobTool;
 pub use grep::GrepTool;
 pub use read::ReadTool;
+pub use skill::SkillTool;
+pub use task::TaskTool;
 pub use write::WriteTool;
 
 use std::sync::Arc;
@@ -26,3 +31,18 @@ pub fn register_builtins(registry: &mut ToolRegistry) {
     registry.register(Arc::new(GlobTool));
     registry.register(Arc::new(GrepTool));
 }
+
+/// Registers the multiagent `task` tool (M4). Kept separate from [`register_builtins`] so
+/// non-orchestrating contexts can omit delegation; the tool no-ops with an error if the
+/// session has no spawner wired in.
+pub fn register_task_tool(registry: &mut ToolRegistry) {
+    registry.register(Arc::new(TaskTool));
+}
+
+/// Registers the `skill` tool (M5) over a loaded skill set. No-op when there are no skills,
+/// so the tool is only advertised when something can be loaded.
+pub fn register_skill_tool(registry: &mut ToolRegistry, skills: &[harness_core::config::SkillDef]) {
+    if !skills.is_empty() {
+        registry.register(Arc::new(SkillTool::new(skills)));
+    }
+}

crates/harness-tools/src/read.rs

@@ -102,6 +102,13 @@ impl Tool for ReadTool {
                 "(empty file or offset past end)".to_string(),
             ));
         }
+
+        // In a subagent session, advertise this read on the job board.
+        if let Some(reporter) = &ctx.context_reporter {
+            let reported = paths::relative_pattern(&ctx.cwd, &path);
+            reporter.report_file(reported, numbered.len() as u32).await;
+        }
+
         Ok(ToolOutput::new(
             params.file_path.clone(),
             numbered.join("\n"),
@@ -139,6 +146,9 @@ mod tests {
                 CancellationToken::new(),
             ),
             metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
         }
     }
 

crates/harness-tools/src/skill.rs

@@ -0,0 +1,154 @@
+//! The `skill` tool (M5): the system prompt advertises each skill's name + description; when
+//! the model decides a skill is relevant it calls this tool with the skill name to pull the
+//! full instructions on demand. See `docs/09-integrations.md`.
+
+use std::collections::HashMap;
+
+use async_trait::async_trait;
+use harness_core::config::SkillDef;
+use harness_core::tool::{Tool, ToolCtx, ToolError, ToolOutput};
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+#[derive(Debug, Deserialize, JsonSchema)]
+struct SkillParams {
+    /// The name of the skill to load, as advertised in the system prompt.
+    name: String,
+}
+
+/// Serves skill bodies by name. Built from the loaded skill set; if there are no skills the
+/// caller simply doesn't register the tool.
+pub struct SkillTool {
+    /// name → (description, body).
+    skills: HashMap<String, (String, String)>,
+    /// Sorted names, for a stable "unknown skill" hint.
+    names: Vec<String>,
+    description: String,
+}
+
+impl SkillTool {
+    pub fn new(skills: &[SkillDef]) -> Self {
+        let mut names: Vec<String> = skills.iter().map(|s| s.name.clone()).collect();
+        names.sort();
+        let map = skills
+            .iter()
+            .map(|s| (s.name.clone(), (s.description.clone(), s.body.clone())))
+            .collect();
+        let description = format!(
+            "Load the full instructions for a named skill before doing the related work. \
+             Available skills: {}.",
+            names.join(", ")
+        );
+        Self {
+            skills: map,
+            names,
+            description,
+        }
+    }
+}
+
+#[async_trait]
+impl Tool for SkillTool {
+    fn name(&self) -> &str {
+        "skill"
+    }
+
+    fn description(&self) -> &str {
+        &self.description
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        serde_json::to_value(schemars::schema_for!(SkillParams)).unwrap()
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        _ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let params: SkillParams =
+            serde_json::from_value(input).map_err(|e| ToolError::Invalid(e.to_string()))?;
+        match self.skills.get(&params.name) {
+            Some((_description, body)) => Ok(ToolOutput::new(
+                format!("skill: {}", params.name),
+                body.clone(),
+            )),
+            None => Err(ToolError::Invalid(format!(
+                "unknown skill {:?}; available: {}",
+                params.name,
+                self.names.join(", ")
+            ))),
+        }
+    }
+}
+
+#[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() -> 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: "c1".into(),
+            data_dir: std::env::temp_dir(),
+            cwd: std::env::temp_dir(),
+            cancel: CancellationToken::new(),
+            ask: PermissionHandle::new(
+                service,
+                SessionId::new(),
+                Vec::new(),
+                Arc::new(Mutex::new(Vec::new())),
+                CancellationToken::new(),
+            ),
+            metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
+        }
+    }
+
+    fn sample() -> Vec<SkillDef> {
+        vec![SkillDef {
+            name: "formatter".into(),
+            description: "format code".into(),
+            body: "Run cargo fmt.".into(),
+        }]
+    }
+
+    #[tokio::test]
+    async fn returns_skill_body_by_name() {
+        let tool = SkillTool::new(&sample());
+        let out = tool
+            .execute(serde_json::json!({"name": "formatter"}), ctx())
+            .await
+            .unwrap();
+        assert_eq!(out.output, "Run cargo fmt.");
+    }
+
+    #[tokio::test]
+    async fn unknown_skill_is_an_input_error() {
+        let tool = SkillTool::new(&sample());
+        let err = tool
+            .execute(serde_json::json!({"name": "nope"}), ctx())
+            .await
+            .unwrap_err();
+        assert!(matches!(err, ToolError::Invalid(_)));
+    }
+
+    #[test]
+    fn description_lists_available_skills() {
+        let tool = SkillTool::new(&sample());
+        assert!(tool.description().contains("formatter"));
+    }
+}

crates/harness-tools/src/task.rs

@@ -0,0 +1,140 @@
+//! The `task` tool: delegate work to a specialist subagent, foreground or background.
+//!
+//! This tool is deliberately thin — it validates input, gates on a `task/<agent>` permission,
+//! and hands off to the engine's `SubagentSpawner` (owned by the composition root), which
+//! resolves the agent, enforces the depth limit, applies permission intersection, and runs
+//! the child session. See `docs/04-multiagent.md`.
+
+use async_trait::async_trait;
+use serde::Deserialize;
+use serde_json::json;
+
+use harness_core::tool::{
+    invalid_input, SpawnError, SpawnRequest, Tool, ToolCtx, ToolError, ToolOutput,
+};
+
+#[derive(Debug, Deserialize)]
+struct TaskInput {
+    /// Short human-facing label for the subtask (shown on the job board).
+    description: String,
+    /// The full instruction handed to the subagent.
+    prompt: String,
+    /// Which specialist to run (must be a subagent-capable agent).
+    subagent_type: String,
+    /// Alias or task id of a completed job to continue instead of starting fresh.
+    #[serde(default)]
+    task_id: Option<String>,
+    /// Run in the background and return immediately (tracked on the job board).
+    #[serde(default)]
+    background: bool,
+}
+
+pub struct TaskTool;
+
+#[async_trait]
+impl Tool for TaskTool {
+    fn name(&self) -> &str {
+        "task"
+    }
+
+    fn description(&self) -> &str {
+        "Delegate a self-contained unit of work to a specialist subagent. Set `background: \
+         true` to launch it without blocking (track progress on the job board); reuse a \
+         completed subagent by passing its `task_id`/alias to continue the same session."
+    }
+
+    fn parameters(&self) -> serde_json::Value {
+        json!({
+            "type": "object",
+            "properties": {
+                "description": { "type": "string", "description": "Short label for the subtask." },
+                "prompt": { "type": "string", "description": "Full instruction for the subagent." },
+                "subagent_type": { "type": "string", "description": "Specialist to run." },
+                "task_id": { "type": "string", "description": "Alias/id of a completed job to reuse." },
+                "background": { "type": "boolean", "description": "Launch without blocking." }
+            },
+            "required": ["description", "prompt", "subagent_type"]
+        })
+    }
+
+    async fn execute(
+        &self,
+        input: serde_json::Value,
+        ctx: ToolCtx,
+    ) -> Result<ToolOutput, ToolError> {
+        let args: TaskInput = serde_json::from_value(input).map_err(|e| invalid_input(self, e))?;
+
+        let Some(spawner) = ctx.spawner.clone() else {
+            return Err(ToolError::Other(
+                "subagent delegation is not available in this session".into(),
+            ));
+        };
+
+        // Gate on task/<agent>. `Always` grants blanket delegation to this specialist.
+        ctx.ask
+            .ask(
+                "task",
+                &args.subagent_type,
+                &args.subagent_type,
+                json!({
+                    "agent": args.subagent_type,
+                    "description": args.description,
+                    "background": args.background,
+                }),
+            )
+            .await?;
+
+        let req = SpawnRequest {
+            parent_session_id: ctx.session_id.clone(),
+            parent_message_id: ctx.message_id.clone(),
+            agent: args.subagent_type.clone(),
+            description: args.description.clone(),
+            prompt: args.prompt,
+            reuse_task_id: args.task_id,
+            background: args.background,
+            cancel: ctx.cancel.clone(),
+        };
+
+        let outcome = spawner.spawn(req).await.map_err(map_spawn_error)?;
+
+        if outcome.background {
+            let alias = outcome.alias.unwrap_or_default();
+            Ok(ToolOutput {
+                title: format!("launched {} ({alias})", args.subagent_type),
+                output: format!(
+                    "Launched background task {alias} ({}). Check the job board; do not poll — \
+                     wait for completion.",
+                    outcome.child_session_id
+                ),
+                metadata: json!({
+                    "child_session": outcome.child_session_id,
+                    "agent": args.subagent_type,
+                    "alias": alias,
+                    "background": true,
+                }),
+            })
+        } else {
+            let text = outcome.final_text.unwrap_or_default();
+            Ok(ToolOutput {
+                title: format!("{} — {}", args.subagent_type, args.description),
+                output: text,
+                metadata: json!({
+                    "child_session": outcome.child_session_id,
+                    "agent": args.subagent_type,
+                    "background": false,
+                }),
+            })
+        }
+    }
+}
+
+fn map_spawn_error(err: SpawnError) -> ToolError {
+    match err {
+        // Depth/agent problems are the model's to fix — surface as tool errors it can read
+        // and act on, not hard failures.
+        SpawnError::DepthExceeded => ToolError::Other(err.to_string()),
+        SpawnError::InvalidAgent(_) => ToolError::Invalid(err.to_string()),
+        SpawnError::ReuseNotFound(_) => ToolError::Invalid(err.to_string()),
+        SpawnError::Other(msg) => ToolError::Other(msg),
+    }
+}

crates/harness-tools/src/write.rs

@@ -80,6 +80,7 @@ impl Tool for WriteTool {
             format!("wrote {} bytes", params.content.len()),
         );
         output.metadata = serde_json::json!({"diff": diff, "added": added, "removed": removed});
+        crate::diagnostics::append_diagnostics(&ctx, &path, &params.file_path, &mut output).await;
         Ok(output)
     }
 }
@@ -114,6 +115,9 @@ mod tests {
                 CancellationToken::new(),
             ),
             metadata,
+            spawner: None,
+            context_reporter: None,
+            diagnostics: None,
         }
     }
 

crates/harness-tui/src/app.rs

@@ -31,7 +31,7 @@ pub struct App {
 
 impl App {
     pub async fn new(cwd: PathBuf) -> anyhow::Result<Self> {
-        let engine = EngineHandle::init(cwd)?;
+        let engine = EngineHandle::init(cwd).await?;
         let bus_rx = engine.bus().subscribe();
 
         let config = engine.config();

crates/harness-tui/src/input.rs

@@ -9,7 +9,16 @@ use crate::state::{AppState, ModalState};
 #[derive(Debug)]
 pub enum InputAction {
     None,
-    Submit { text: String },
+    Submit {
+        text: String,
+    },
+    /// A user-defined slash command: `name` (no leading `/`), its `args`, and the `raw` input
+    /// to fall back to submitting verbatim if no such command is defined.
+    RunCommand {
+        name: String,
+        args: String,
+        raw: String,
+    },
     Abort,
     Quit,
     LoadSessions,
@@ -23,6 +32,10 @@ pub enum InputAction {
     SessionPickerUp,
     SessionPickerDown,
     SessionPickerSelect,
+    OpenJobs,
+    JobsUp,
+    JobsDown,
+    JobsDrillIn,
 }
 
 /// Translate a crossterm event into an action and/or mutate `state` directly.
@@ -49,10 +62,21 @@ 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::JobsPane { .. } => handle_jobs_pane_key(key, state),
         ModalState::None => handle_normal_key(key, state),
     }
 }
 
+fn handle_jobs_pane_key(key: KeyEvent, _state: &mut AppState) -> InputAction {
+    match key.code {
+        KeyCode::Up => InputAction::JobsUp,
+        KeyCode::Down => InputAction::JobsDown,
+        KeyCode::Enter => InputAction::JobsDrillIn,
+        KeyCode::Esc => InputAction::CloseModal,
+        _ => InputAction::None,
+    }
+}
+
 fn handle_permission_key(key: KeyEvent, _state: &mut AppState) -> InputAction {
     match key.code {
         KeyCode::Char('y') | KeyCode::Char('Y') => {
@@ -119,6 +143,7 @@ fn handle_normal_key(key: KeyEvent, state: &mut AppState) -> InputAction {
             }
         }
         KeyCode::Char('s') if ctrl => InputAction::LoadSessions,
+        KeyCode::Char('j') if ctrl => InputAction::OpenJobs,
         KeyCode::Up => {
             let (row, _) = state.input.cursor();
             if row == 0 {
@@ -158,8 +183,15 @@ fn parse_slash_command(text: &str) -> Option<InputAction> {
         "/model" if !rest.is_empty() => Some(InputAction::SetModel(rest)),
         "/agent" if !rest.is_empty() => Some(InputAction::SetAgent(rest)),
         "/sessions" => Some(InputAction::LoadSessions),
+        "/jobs" => Some(InputAction::OpenJobs),
         "/quit" => Some(InputAction::Quit),
-        _ => None,
+        // Any other `/word` is treated as a user-defined command, resolved against the engine's
+        // loaded commands when applied; if none matches, the raw text is submitted as-is.
+        other => Some(InputAction::RunCommand {
+            name: other.trim_start_matches('/').to_string(),
+            args: rest,
+            raw: trimmed.to_string(),
+        }),
     }
 }
 
@@ -173,6 +205,30 @@ pub async fn apply_action(action: InputAction, state: &mut AppState, engine: &En
                 }
             }
         }
+        InputAction::RunCommand { name, args, raw } => {
+            let Some(session_id) = state.session_id.clone() else {
+                return;
+            };
+            match engine.command(&name) {
+                Some(cmd) => {
+                    let text = cmd.expand(&args);
+                    // A command may switch model/agent for this one run only.
+                    let model_ref = cmd.model.clone().unwrap_or_else(|| state.model_ref.clone());
+                    if let Err(e) = engine
+                        .prompt_with(session_id, text, &model_ref, cmd.agent.as_deref())
+                        .await
+                    {
+                        tracing::error!(error = %e, "command prompt failed");
+                    }
+                }
+                // Unknown command: submit the original text as an ordinary message.
+                None => {
+                    if let Err(e) = engine.prompt(session_id, raw, &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);
@@ -187,6 +243,8 @@ pub async fn apply_action(action: InputAction, state: &mut AppState, engine: &En
                 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"),
@@ -233,6 +291,45 @@ pub async fn apply_action(action: InputAction, state: &mut AppState, engine: &En
                 }
             }
         }
+        InputAction::OpenJobs => {
+            if let Some(session_id) = state.session_id.clone() {
+                match engine.jobs(session_id).await {
+                    Ok(jobs) => {
+                        state.set_jobs(jobs);
+                        state.open_jobs_pane();
+                    }
+                    Err(e) => tracing::error!(error = %e, "failed to load jobs"),
+                }
+            }
+        }
+        InputAction::JobsUp => {
+            if let ModalState::JobsPane { selected } = &mut state.modal {
+                *selected = selected.saturating_sub(1);
+            }
+            state.dirty = true;
+        }
+        InputAction::JobsDown => {
+            let job_count = state.jobs.len();
+            if let ModalState::JobsPane { selected } = &mut state.modal {
+                if *selected + 1 < job_count {
+                    *selected += 1;
+                }
+            }
+            state.dirty = true;
+        }
+        InputAction::JobsDrillIn => {
+            if let Some(child) = state.selected_job_child() {
+                match engine.get_session(child).await {
+                    Ok(Some(session)) => {
+                        if let Err(e) = modal::select_session(state, engine, session).await {
+                            tracing::error!(error = %e, "failed to open subtask session");
+                        }
+                    }
+                    Ok(None) => tracing::warn!("subtask session no longer exists"),
+                    Err(e) => tracing::error!(error = %e, "failed to load subtask session"),
+                }
+            }
+        }
         InputAction::None => {}
     }
 }
@@ -259,4 +356,33 @@ mod tests {
         assert!(state.dirty, "a keystroke must request a redraw");
         assert_eq!(state.input.lines().join("\n"), "x");
     }
+
+    #[test]
+    fn builtin_slash_commands_still_parse() {
+        assert!(matches!(
+            parse_slash_command("/new"),
+            Some(InputAction::NewSession)
+        ));
+        assert!(matches!(
+            parse_slash_command("/model openai/gpt-5"),
+            Some(InputAction::SetModel(m)) if m == "openai/gpt-5"
+        ));
+    }
+
+    #[test]
+    fn unknown_slash_becomes_a_run_command() {
+        match parse_slash_command("/deploy prod now") {
+            Some(InputAction::RunCommand { name, args, raw }) => {
+                assert_eq!(name, "deploy");
+                assert_eq!(args, "prod now");
+                assert_eq!(raw, "/deploy prod now");
+            }
+            other => panic!("expected RunCommand, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn non_slash_text_is_not_a_command() {
+        assert!(parse_slash_command("hello world").is_none());
+    }
 }

crates/harness-tui/src/main.rs

@@ -50,7 +50,7 @@ async fn run_headless(args: &[String]) -> i32 {
         }
     };
 
-    let app = match harness_app::App::init(cwd) {
+    let app = match harness_app::App::init(cwd).await {
         Ok(app) => app,
         Err(e) => {
             eprintln!("error: {e}");

crates/harness-tui/src/modal.rs

@@ -26,13 +26,20 @@ pub async fn select_session(
     session: Session,
 ) -> Result<(), harness_app::AppError> {
     let session_id = session.id.clone();
-    let messages = engine.session_messages(session_id).await?;
+    // Repair a session that crashed mid-run before loading it, so the transcript shows the
+    // aborted turn rather than a tool call frozen "running".
+    engine.repair_session(&session_id).await?;
+    let messages = engine.session_messages(session_id.clone()).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);
+    // Surface any subtasks this session spawned (drill-in and session-switch both land here).
+    if let Ok(jobs) = engine.jobs(session_id).await {
+        state.set_jobs(jobs);
+    }
     state.modal = ModalState::None;
     Ok(())
 }

crates/harness-tui/src/render.rs

@@ -33,6 +33,9 @@ pub fn render(frame: &mut Frame, state: &mut AppState) {
         ModalState::SessionPicker { sessions, selected } => {
             render_session_picker(frame, sessions, *selected, area);
         }
+        ModalState::JobsPane { selected } => {
+            render_jobs_pane(frame, &state.jobs, *selected, area);
+        }
         ModalState::None => {}
     }
 }
@@ -169,13 +172,27 @@ fn render_status(frame: &mut Frame, state: &AppState, area: Rect) {
     let hints = if state.ctrl_c_pressed {
         "Press Ctrl+C again to quit"
     } else {
-        "Ctrl+S: sessions | Esc: abort | Ctrl+C: quit"
+        "Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C: quit"
     };
-    let text = format!("{spinner}{status} · {hints}");
+    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,
@@ -228,6 +245,86 @@ fn render_session_picker(
     frame.render_widget(paragraph, popup);
 }
 
+fn render_jobs_pane(
+    frame: &mut Frame,
+    jobs: &[harness_core::engine::JobRecord],
+    selected: usize,
+    area: Rect,
+) {
+    let popup = centered_rect(70, 70, area);
+    frame.render_widget(Clear, popup);
+
+    let mut text: Vec<Line<'static>> =
+        vec![Line::from("Background jobs").style(Style::new().add_modifier(Modifier::BOLD))];
+
+    if jobs.is_empty() {
+        text.push(Line::default());
+        text.push(Line::from("No subtasks spawned yet.").style(Style::new().fg(Color::DarkGray)));
+    } else {
+        for (i, job) in jobs.iter().enumerate() {
+            let marker = if i == selected { "> " } else { "  " };
+            let (icon, color) = job_state_style(job.state);
+            let header = format!(
+                "{marker}{icon} {} · {} · {}",
+                job.alias,
+                job.agent,
+                job_state_label(job.state),
+            );
+            let style = if i == selected {
+                Style::new().bg(Color::Blue).fg(Color::White)
+            } else {
+                Style::new().fg(color)
+            };
+            text.push(Line::from(Span::styled(header, style)));
+            if let Some(objective) = &job.objective {
+                text.push(
+                    Line::from(format!("    {objective}")).style(Style::new().fg(Color::Gray)),
+                );
+            }
+            if !job.context_files.is_empty() {
+                let files: Vec<&str> = job
+                    .context_files
+                    .iter()
+                    .take(8)
+                    .map(|f| f.path.as_str())
+                    .collect();
+                text.push(
+                    Line::from(format!("    read: {}", files.join(", ")))
+                        .style(Style::new().fg(Color::DarkGray)),
+                );
+            }
+        }
+        text.push(Line::default());
+        text.push(
+            Line::from("Enter: open subtask · Esc: close").style(Style::new().fg(Color::DarkGray)),
+        );
+    }
+
+    let block = Block::default().borders(Borders::ALL).title(" jobs ");
+    let paragraph = Paragraph::new(text).block(block).wrap(Wrap { trim: false });
+    frame.render_widget(paragraph, popup);
+}
+
+fn job_state_label(state: harness_core::engine::JobState) -> &'static str {
+    use harness_core::engine::JobState;
+    match state {
+        JobState::Running => "running",
+        JobState::Completed => "completed",
+        JobState::Error => "error",
+        JobState::Cancelled => "cancelled",
+    }
+}
+
+fn job_state_style(state: harness_core::engine::JobState) -> (&'static str, Color) {
+    use harness_core::engine::JobState;
+    match state {
+        JobState::Running => ("⚙", Color::Yellow),
+        JobState::Completed => ("✓", Color::Green),
+        JobState::Error => ("✗", Color::Red),
+        JobState::Cancelled => ("∅", Color::DarkGray),
+    }
+}
+
 fn centered_rect(percent_x: u16, percent_y: u16, r: Rect) -> Rect {
     let popup_layout = Layout::default()
         .direction(Direction::Vertical)
@@ -473,6 +570,59 @@ mod tests {
         insta::assert_snapshot!(buffer_to_string(terminal.backend()));
     }
 
+    #[test]
+    fn snapshot_jobs_pane() {
+        use harness_core::engine::{ContextFile, JobRecord, JobState};
+
+        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.jobs = vec![
+            JobRecord {
+                task_id: "t1".to_string(),
+                alias: "exp-1".to_string(),
+                parent_session: SessionId("ses_test_001".to_string()),
+                child_session: SessionId("ses_child_001".to_string()),
+                agent: "explorer".to_string(),
+                description: "map auth".to_string(),
+                objective: Some("map the auth flow".to_string()),
+                state: JobState::Completed,
+                reconciled: true,
+                result_summary: Some("done".to_string()),
+                context_files: vec![ContextFile {
+                    path: "src/auth.rs".to_string(),
+                    lines: 42,
+                }],
+                launched_at: 1,
+                updated_at: 2,
+                last_used_at: 2,
+            },
+            JobRecord {
+                task_id: "t2".to_string(),
+                alias: "fix-1".to_string(),
+                parent_session: SessionId("ses_test_001".to_string()),
+                child_session: SessionId("ses_child_002".to_string()),
+                agent: "fixer".to_string(),
+                description: "patch bug".to_string(),
+                objective: Some("fix the null deref".to_string()),
+                state: JobState::Running,
+                reconciled: false,
+                result_summary: None,
+                context_files: Vec::new(),
+                launched_at: 3,
+                updated_at: 3,
+                last_used_at: 3,
+            },
+        ];
+        state.modal = ModalState::JobsPane { 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);

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

@@ -1,6 +1,6 @@
 ---
 source: crates/harness-tui/src/render.rs
-assertion_line: 277
+assertion_line: 381
 expression: buffer_to_string(terminal.backend())
 ---
             new session · orchestrator · anthropic/claude-sonnet-4-5
@@ -26,4 +26,4 @@ expression: buffer_to_string(terminal.backend())
 ┌ input ───────────────────────────────────────────────────────────────────────┐
 │                                                                              │
 └──────────────────────────────────────────────────────────────────────────────┘
-idle · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

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

@@ -0,0 +1,29 @@
+---
+source: crates/harness-tui/src/render.rs
+assertion_line: 621
+expression: buffer_to_string(terminal.backend())
+---
+            new session · orchestrator · anthropic/claude-sonnet-4-5
+┌ chat ────────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+│                                                                              │
+│           ┌ jobs ────────────────────────────────────────────────┐           │
+│           │Background jobs                                       │           │
+│           │  ✓ exp-1 · explorer · completed                      │           │
+│           │    map the auth flow                                 │           │
+│           │    read: src/auth.rs                                 │           │
+│           │> ⚙ fix-1 · fixer · running                           │           │
+│           │    fix the null deref                                │           │
+│           │                                                      │           │
+│           │Enter: open subtask · Esc: close                      │           │
+│           │                                                      │           │
+│           │                                                      │           │
+│           │                                                      │           │
+│           │                                                      │           │
+│           │                                                      │           │
+│           │                                                      │           │
+└───────────└──────────────────────────────────────────────────────┘───────────┘
+┌ input ───────────────────────────────────────────────────────────────────────┐
+│                                                                              │
+└──────────────────────────────────────────────────────────────────────────────┘
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

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

@@ -1,6 +1,6 @@
 ---
 source: crates/harness-tui/src/render.rs
-assertion_line: 397
+assertion_line: 511
 expression: buffer_to_string(terminal.backend())
 ---
             new session · orchestrator · anthropic/claude-sonnet-4-5
@@ -26,4 +26,4 @@ expression: buffer_to_string(terminal.backend())
 ┌ input ───────────────────────────────────────────────────────────────────────┐
 │                                                                              │
 └──────────────────────────────────────────────────────────────────────────────┘
-idle · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

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

@@ -1,6 +1,6 @@
 ---
 source: crates/harness-tui/src/render.rs
-assertion_line: 449
+assertion_line: 568
 expression: buffer_to_string(terminal.backend())
 ---
             new session · orchestrator · anthropic/claude-sonnet-4-5
@@ -26,4 +26,4 @@ expression: buffer_to_string(terminal.backend())
 ┌ input ───────────────────────────────────────────────────────────────────────┐
 │                                                                              │
 └──────────────────────────────────────────────────────────────────────────────┘
-idle · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

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

@@ -1,6 +1,6 @@
 ---
 source: crates/harness-tui/src/render.rs
-assertion_line: 330
+assertion_line: 438
 expression: buffer_to_string(terminal.backend())
 ---
             new session · orchestrator · anthropic/claude-sonnet-4-5
@@ -26,4 +26,4 @@ expression: buffer_to_string(terminal.backend())
 ┌ input ───────────────────────────────────────────────────────────────────────┐
 │                                                                              │
 └──────────────────────────────────────────────────────────────────────────────┘
-idle · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

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

@@ -1,6 +1,6 @@
 ---
 source: crates/harness-tui/src/render.rs
-assertion_line: 376
+assertion_line: 488
 expression: buffer_to_string(terminal.backend())
 ---
             new session · orchestrator · anthropic/claude-sonnet-4-5
@@ -26,4 +26,4 @@ expression: buffer_to_string(terminal.backend())
 ┌ input ───────────────────────────────────────────────────────────────────────┐
 │                                                                              │
 └──────────────────────────────────────────────────────────────────────────────┘
-idle · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

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

@@ -1,6 +1,6 @@
 ---
 source: crates/harness-tui/src/render.rs
-assertion_line: 353
+assertion_line: 463
 expression: buffer_to_string(terminal.backend())
 ---
             new session · orchestrator · anthropic/claude-sonnet-4-5
@@ -26,4 +26,4 @@ expression: buffer_to_string(terminal.backend())
 ┌ input ───────────────────────────────────────────────────────────────────────┐
 │                                                                              │
 └──────────────────────────────────────────────────────────────────────────────┘
-idle · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

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

@@ -1,6 +1,6 @@
 ---
 source: crates/harness-tui/src/render.rs
-assertion_line: 306
+assertion_line: 412
 expression: buffer_to_string(terminal.backend())
 ---
             new session · orchestrator · anthropic/claude-sonnet-4-5
@@ -26,4 +26,4 @@ expression: buffer_to_string(terminal.backend())
 ┌ input ───────────────────────────────────────────────────────────────────────┐
 │                                                                              │
 └──────────────────────────────────────────────────────────────────────────────┘
-idle · Ctrl+S: sessions | Esc: abort | Ctrl+C: quit
+idle · 0 tok · $0.0000 · Ctrl+S: sessions | Ctrl+J: jobs | Esc: abort | Ctrl+C:

crates/harness-tui/src/state.rs

@@ -1,4 +1,5 @@
 use harness_app::EngineHandle;
+use harness_core::engine::JobRecord;
 use harness_core::event::{AppEvent, PermissionRequest, RunOutcome};
 use harness_core::permission::PermissionReply;
 use harness_core::types::{Message, Part, PartBody, PartId, Role, Session, SessionId, ToolState};
@@ -86,6 +87,10 @@ pub enum ModalState {
         sessions: Vec<Session>,
         selected: usize,
     },
+    /// Background job board for the current session, with a cursor for drill-in.
+    JobsPane {
+        selected: usize,
+    },
 }
 
 /// All mutable UI state lives here.
@@ -105,6 +110,13 @@ pub struct AppState {
     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,
+    /// Background jobs spawned by the current session, newest activity last. Populated on
+    /// session load and kept live via `JobUpdated` events; surfaced in the jobs pane.
+    pub jobs: Vec<JobRecord>,
 }
 
 impl AppState {
@@ -125,6 +137,9 @@ impl AppState {
             dirty: true,
             ctrl_c_pressed: false,
             render_width: 0,
+            session_cost: 0.0,
+            session_tokens: 0,
+            jobs: Vec::new(),
         }
     }
 
@@ -147,7 +162,11 @@ impl AppState {
         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();
+        // Jobs are reloaded for the newly-selected session by the caller.
+        self.jobs.clear();
 
         let mut messages = messages;
         messages.sort_by_key(|m| m.created_at);
@@ -171,6 +190,8 @@ impl AppState {
                     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;
                 }
             }
@@ -248,13 +269,49 @@ impl AppState {
                 }
                 self.dirty = true;
             }
-            AppEvent::JobUpdated { .. }
+            AppEvent::JobUpdated { job } => {
-            | AppEvent::AuthPrompt { .. }
+                if let Ok(record) = serde_json::from_value::<JobRecord>(job.0) {
-            | AppEvent::ServerNotice { .. } => {
+                    if self.session_id.as_ref() == Some(&record.parent_session) {
+                        self.upsert_job(record);
                         self.dirty = true;
                     }
                 }
             }
+            AppEvent::AuthPrompt { .. } | AppEvent::ServerNotice { .. } => {
+                self.dirty = true;
+            }
+        }
+    }
+
+    /// Replaces this session's job list (e.g. after loading a session).
+    pub fn set_jobs(&mut self, jobs: Vec<JobRecord>) {
+        self.jobs = jobs;
+        self.dirty = true;
+    }
+
+    /// Inserts or replaces a job by `task_id`, preserving list order for stable rendering.
+    fn upsert_job(&mut self, record: JobRecord) {
+        if let Some(existing) = self.jobs.iter_mut().find(|j| j.task_id == record.task_id) {
+            *existing = record;
+        } else {
+            self.jobs.push(record);
+        }
+    }
+
+    /// Opens the jobs pane (cursor at the top).
+    pub fn open_jobs_pane(&mut self) {
+        self.modal = ModalState::JobsPane { selected: 0 };
+        self.dirty = true;
+    }
+
+    /// Child session of the currently-selected job in the jobs pane, for drill-in.
+    pub fn selected_job_child(&self) -> Option<SessionId> {
+        if let ModalState::JobsPane { selected } = &self.modal {
+            self.jobs.get(*selected).map(|j| j.child_session.clone())
+        } else {
+            None
+        }
+    }
 
     pub fn scroll_up(&mut self, n: u16) {
         self.scroll_offset = self.scroll_offset.saturating_sub(n);