M3: model pricing and per-step cost accumulation

Extends the model catalog with per-token pricing and accumulates cost per engine step, threading the running total through the store and up into harness-app so it can be surfaced in the UI.

crates/harness-app/src/lib.rs

@@ -254,6 +254,8 @@ impl EngineHandle {
             temperature: None,
             max_steps: DEFAULT_MAX_STEPS,
             instructions: self.inner.config.instructions.clone(),
+            // Populated from models.dev metadata in the follow-up increment.
+            cost: None,
         };
 
         // Reserve the run slot *before* spawning. If we inserted after spawning, a run that

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

@@ -20,6 +20,33 @@ 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>,
+}
+
+/// 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
@@ -221,17 +248,24 @@ 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) => match outcome.result {
-                StepResult::Continue => continue,
-                StepResult::Stop => return RunOutcome::Stopped,
-                StepResult::Compact => return RunOutcome::Stopped, // stub until M6
-            },
+            Ok(outcome) if outcome.aborted => {
+                accumulate_session_usage(&ctx, &outcome.usage, outcome.cost, now_fn()).await;
+                return RunOutcome::Aborted;
+            }
+            Ok(outcome) => {
+                accumulate_session_usage(&ctx, &outcome.usage, outcome.cost, now_fn()).await;
+                match outcome.result {
+                    StepResult::Continue => continue,
+                    StepResult::Stop => return RunOutcome::Stopped,
+                    StepResult::Compact => return RunOutcome::Stopped, // stub until M6
+                }
+            }
             Err(step_err) if matches!(step_err.source, ProviderError::Cancelled) => {
                 return RunOutcome::Aborted;
             }
@@ -422,11 +456,28 @@ 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()
+            }),
         };
 
         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(),
@@ -517,6 +568,7 @@ mod tests {
             temperature: None,
             max_steps: 10,
             instructions: Vec::new(),
+            cost: None,
         };
 
         let outcome = run_session(Arc::new(provider), ctx, &run_config, || 2).await;

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

@@ -26,6 +26,8 @@ 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,
 }
 
@@ -513,6 +515,7 @@ impl<'a> Run<'a> {
         &mut self,
         reason: FinishReason,
         usage: TokenUsage,
+        cost: f64,
     ) -> Result<StepResult, ProviderError> {
         let part = Part {
             id: PartId::new(),
@@ -521,7 +524,7 @@ impl<'a> Run<'a> {
             idx: self.next_idx,
             body: PartBody::StepFinish {
                 usage,
-                cost: 0.0,
+                cost,
                 reason: reason.clone(),
             },
         };
@@ -567,10 +570,12 @@ 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 {
@@ -582,6 +587,7 @@ 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,
                 });
             }
@@ -630,7 +636,8 @@ 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 +659,7 @@ pub async fn process_step(
         result,
         message_id: run.assistant.as_ref().map(|m| m.id.clone()),
         usage,
+        cost: step_cost,
         aborted: false,
     })
 }

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

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

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

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

crates/harness-core/src/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;
-// this placeholder only carries what harness-core needs to key on.
+/// Per-model pricing in USD per **one million** tokens. Populated from models.dev metadata
+/// 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,
 }