(MOT-4375) fix(harness,providers): session-total cost in the chip; stop double-billing cached tokens

harness/src/context_snapshot.rs

@@ -134,6 +134,12 @@ pub struct ContextSnapshotV1 {
     /// generation never completed).
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub usage: Option<Usage>,
+    /// Running cost of the whole session in USD, accumulated across every
+    /// generation step. `usage.cost_usd` is one step's bill — on providers
+    /// with steep cache discounts the per-step number swings two orders of
+    /// magnitude, so a chip showing it alone reads as a bouncing total.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub session_cost_usd: Option<f64>,
     pub timestamp: i64,
 }
 
@@ -419,6 +425,7 @@ mod tests {
                 reasoning: None,
                 cost_usd: Some(0.42),
             }),
+            session_cost_usd: Some(1.37),
             timestamp: 1_722_700_000_000,
         };
         let mut value = serde_json::to_value(&snapshot).unwrap();

harness/src/turn_loop.rs

@@ -739,6 +739,24 @@ pub async fn run_step(
     // must never fail a turn that generated successfully.
     if let Some(snapshot) = record.context_snapshot.as_mut() {
         snapshot.usage = outcome.message.usage.clone();
+        // Accumulate the session's running cost on top of the stored
+        // snapshot's total: the loop is the only writer per session, so the
+        // read-back is race-free, and seeding from the store keeps the total
+        // honest across turns and harness restarts.
+        let step_cost = outcome
+            .message
+            .usage
+            .as_ref()
+            .and_then(|u| u.cost_usd)
+            .unwrap_or(0.0);
+        let prior_cost =
+            crate::context_snapshot::get(&deps.iii, &record.session_id, cfg.session_timeout_ms)
+                .await
+                .ok()
+                .flatten()
+                .and_then(|prev| prev.session_cost_usd)
+                .unwrap_or(0.0);
+        snapshot.session_cost_usd = Some(prior_cost + step_cost);
         crate::context_snapshot::exactify(snapshot, &router, gen_system_prompt.as_deref(), &tools)
             .await;
         if let Err(error) =
@@ -2217,6 +2235,7 @@ fn build_context_snapshot(
         model: record.options.model.clone(),
         provider: record.options.provider.clone(),
         estimator: b.estimator,
+        session_cost_usd: None,
         usable: assembled.usable,
         effective_max_output_tokens: assembled.effective_max_output_tokens,
         total: final_request_tokens,

harness/tests/golden/schemas/harness.metrics.json

@@ -53,6 +53,14 @@
               "null"
             ]
           },
+          "session_cost_usd": {
+            "description": "Running cost of the whole session in USD, accumulated across every generation step. `usage.cost_usd` is one step's bill — on providers with steep cache discounts the per-step number swings two orders of magnitude, so a chip showing it alone reads as a bouncing total.",
+            "format": "double",
+            "type": [
+              "number",
+              "null"
+            ]
+          },
           "session_id": {
             "type": "string"
           },

harness/ui/src/context-chip/index.tsx

@@ -282,6 +282,7 @@ function ContextPopover({
           <span>
             last step {formatTokens(usage?.input ?? 0)} in · output{' '}
             {formatTokens(usage?.output ?? 0)}
+            {usage?.cost_usd != null ? <> · {formatCost(usage.cost_usd)}</> : null}
           </span>
         ) : null}
         {cache ? (
@@ -298,8 +299,15 @@ function ContextPopover({
             </span>
           </span>
         ) : null}
-        {usage?.cost_usd != null ? (
-          <span>cost {formatCost(usage.cost_usd)}</span>
+        {snapshot.session_cost_usd != null ? (
+          <span
+            title={
+              'every generation step of this session summed. The per-step ' +
+              'cost on the line above swings with cache hits; this one only grows.'
+            }
+          >
+            session total {formatCost(snapshot.session_cost_usd)}
+          </span>
         ) : null}
       </div>
     </div>

harness/ui/src/lib/metrics.ts

@@ -46,6 +46,8 @@ export interface ContextSnapshot {
   compacted: boolean
   summarized_head_tokens?: number
   usage?: SnapshotUsage
+  /** Running cost of the whole session, summed across generation steps. */
+  session_cost_usd?: number
   timestamp: number
 }
 

provider-kimi/src/sse.rs

@@ -153,20 +153,28 @@ pub fn map_finish_reason(s: &str) -> StopReason {
 /// values — overwriting is correct for both, adding double-counts.
 pub fn merge_usage(raw: &Value, into: &mut Usage) {
     let num = |k: &str| raw.get(k).and_then(Value::as_u64);
-    if let Some(v) = num("prompt_tokens").or_else(|| num("input_tokens")) {
-        into.input = Some(v);
-    }
-    if let Some(v) = num("completion_tokens").or_else(|| num("output_tokens")) {
-        into.output = Some(v);
-    }
+    // `Usage.input` is the cache-MISS slice, disjoint from `cache_read`
+    // (pricing bills the splits additively). The wire's `prompt_tokens` is a
+    // TOTAL that includes the cached slice, so the miss slice is derived —
+    // mapping the total verbatim would bill the cached prefix twice.
+    let mut cached = None;
     for parent in ["prompt_tokens_details", "input_tokens_details"] {
         if let Some(v) = raw
             .pointer(&format!("/{parent}/cached_tokens"))
             .and_then(Value::as_u64)
         {
-            into.cache_read = Some(v);
+            cached = Some(v);
         }
     }
+    if let Some(v) = cached {
+        into.cache_read = Some(v);
+    }
+    if let Some(v) = num("prompt_tokens").or_else(|| num("input_tokens")) {
+        into.input = Some(v.saturating_sub(cached.unwrap_or(0)));
+    }
+    if let Some(v) = num("completion_tokens").or_else(|| num("output_tokens")) {
+        into.output = Some(v);
+    }
     if let Some(v) = raw
         .pointer("/completion_tokens_details/reasoning_tokens")
         .and_then(Value::as_u64)
@@ -394,7 +402,7 @@ mod tests {
         assert_eq!(final_msg.stop_reason, StopReason::End);
         assert_eq!(final_msg.native_stop_reason.as_deref(), Some("stop"));
         let usage = final_msg.usage.unwrap();
-        assert_eq!(usage.input, Some(12));
+        assert_eq!(usage.input, Some(8));
         assert_eq!(usage.output, Some(2));
         assert_eq!(usage.cache_read, Some(4));
         assert_eq!(usage.reasoning, Some(0));

provider-kimi/src/upstream.rs

@@ -214,7 +214,7 @@ mod tests {
         );
         match events.last() {
             Some(AssistantMessageEvent::Done { message }) => {
-                assert_eq!(message.usage.as_ref().unwrap().input, Some(12));
+                assert_eq!(message.usage.as_ref().unwrap().input, Some(8));
                 assert_eq!(message.usage.as_ref().unwrap().output, Some(2));
                 assert_eq!(message.usage.as_ref().unwrap().cache_read, Some(4));
                 assert_eq!(message.native_stop_reason.as_deref(), Some("stop"));

provider-llamacpp/src/sse.rs

@@ -154,20 +154,28 @@ pub fn map_finish_reason(s: &str) -> StopReason {
 /// correct whether it reports once at the end or cumulatively per chunk.
 pub fn merge_usage(raw: &Value, into: &mut Usage) {
     let num = |k: &str| raw.get(k).and_then(Value::as_u64);
-    if let Some(v) = num("prompt_tokens").or_else(|| num("input_tokens")) {
-        into.input = Some(v);
-    }
-    if let Some(v) = num("completion_tokens").or_else(|| num("output_tokens")) {
-        into.output = Some(v);
-    }
+    // `Usage.input` is the cache-MISS slice, disjoint from `cache_read`
+    // (pricing bills the splits additively). The wire's `prompt_tokens` is a
+    // TOTAL that includes the cached slice, so the miss slice is derived —
+    // mapping the total verbatim would bill the cached prefix twice.
+    let mut cached = None;
     for parent in ["prompt_tokens_details", "input_tokens_details"] {
         if let Some(v) = raw
             .pointer(&format!("/{parent}/cached_tokens"))
             .and_then(Value::as_u64)
         {
-            into.cache_read = Some(v);
+            cached = Some(v);
         }
     }
+    if let Some(v) = cached {
+        into.cache_read = Some(v);
+    }
+    if let Some(v) = num("prompt_tokens").or_else(|| num("input_tokens")) {
+        into.input = Some(v.saturating_sub(cached.unwrap_or(0)));
+    }
+    if let Some(v) = num("completion_tokens").or_else(|| num("output_tokens")) {
+        into.output = Some(v);
+    }
     if let Some(v) = raw
         .pointer("/completion_tokens_details/reasoning_tokens")
         .and_then(Value::as_u64)
@@ -427,7 +435,7 @@ mod tests {
         assert_eq!(final_msg.stop_reason, StopReason::End);
         assert_eq!(final_msg.native_stop_reason.as_deref(), Some("stop"));
         let usage = final_msg.usage.unwrap();
-        assert_eq!(usage.input, Some(12));
+        assert_eq!(usage.input, Some(8));
         assert_eq!(usage.output, Some(2));
         assert_eq!(usage.cache_read, Some(4));
         assert_eq!(usage.reasoning, Some(0));

provider-llamacpp/src/upstream.rs

@@ -221,7 +221,7 @@ mod tests {
         );
         match events.last() {
             Some(AssistantMessageEvent::Done { message }) => {
-                assert_eq!(message.usage.as_ref().unwrap().input, Some(12));
+                assert_eq!(message.usage.as_ref().unwrap().input, Some(8));
                 assert_eq!(message.usage.as_ref().unwrap().output, Some(2));
                 assert_eq!(message.usage.as_ref().unwrap().cache_read, Some(4));
                 assert_eq!(message.native_stop_reason.as_deref(), Some("stop"));