Persist context window usage across turns from last task_progress

Stop resetting lastContextWindowUsedTokens between turns. The
result.usage fallback contains accumulated session totals which grow
incorrectly across turns. Once a task_progress arrives, its value is
the accurate context fill level and should be retained until the next
task_progress supersedes it.

packages/server/src/server/agent/providers/claude-agent.test.ts

@@ -358,7 +358,7 @@ describe("ClaudeAgentSession context window usage", () => {
     });
   }
 
-  test("convertUsage includes contextWindowMaxTokens and derives used tokens from result usage", async () => {
+  test("convertUsage includes contextWindowMaxTokens and derives used tokens from result usage as initial fallback", async () => {
     const session = await createSessionForTest();
 
     const usage = session.convertUsage(
@@ -388,7 +388,7 @@ describe("ClaudeAgentSession context window usage", () => {
     });
   });
 
-  test("contextWindowUsedTokens is derived from result usage when task_progress is missing", async () => {
+  test("contextWindowUsedTokens falls back to result usage when no task_progress was received", async () => {
     const session = await createSessionForTest();
 
     const usage = session.convertUsage({
@@ -508,7 +508,7 @@ describe("ClaudeAgentSession context window usage", () => {
     });
   });
 
-  test("contextWindowUsedTokens resets between turns and falls back to result usage", async () => {
+  test("contextWindowUsedTokens persists across turns from last task_progress", async () => {
     const queryFactory = createQueryFactoryForTurns([
       [
         {
@@ -600,20 +600,23 @@ describe("ClaudeAgentSession context window usage", () => {
         contextWindowMaxTokens: 200_000,
         contextWindowUsedTokens: 999,
       });
+      // Turn 2 has no task_progress, so contextWindowUsedTokens retains the
+      // last known value from turn 1 rather than deriving from accumulated
+      // result.usage (which would be incorrect — those are session-level totals).
       expect(secondTurn.usage).toEqual({
         inputTokens: 11,
         cachedInputTokens: 6,
         outputTokens: 8,
         totalCostUsd: 0.1,
         contextWindowMaxTokens: 200_000,
-        contextWindowUsedTokens: 28,
+        contextWindowUsedTokens: 999,
       });
     } finally {
       await session.close();
     }
   });
 
-  test("convertUsage derives used tokens even when modelUsage is missing", async () => {
+  test("convertUsage derives used tokens from result usage as fallback when task_progress is missing", async () => {
     const session = await createSessionForTest();
 
     const usage = session.convertUsage({

packages/server/src/server/agent/providers/claude-agent.ts

@@ -1460,7 +1460,6 @@ class ClaudeAgentSession implements AgentSession {
 
     const sdkMessage = this.toSdkUserMessage(prompt);
     this.lastForegroundPromptText = this.extractPromptText(prompt);
-    this.lastContextWindowUsedTokens = undefined;
     const turnId = this.createTurnId("foreground");
     this.activeForegroundTurnId = turnId;
     this.foregroundHasVisibleActivity = false;
@@ -2304,7 +2303,6 @@ class ClaudeAgentSession implements AgentSession {
     this.notifySubscribers(event);
     this.activeForegroundTurnId = null;
     this.lastForegroundPromptText = null;
-    this.lastContextWindowUsedTokens = undefined;
     this.cancelCurrentTurn = null;
     this.syncTurnState("foreground turn terminal");
   }
@@ -2320,12 +2318,10 @@ class ClaudeAgentSession implements AgentSession {
       if (this.activeForegroundTurnId) {
         this.activeForegroundTurnId = null;
         this.lastForegroundPromptText = null;
-        this.lastContextWindowUsedTokens = undefined;
         this.cancelCurrentTurn = null;
         this.syncTurnState("foreground turn terminal");
       } else if (this.autonomousTurn) {
         this.autonomousTurn = null;
-        this.lastContextWindowUsedTokens = undefined;
         this.syncTurnState("autonomous turn terminal");
       }
     }
@@ -2338,7 +2334,6 @@ class ClaudeAgentSession implements AgentSession {
     this.autonomousTurn = {
       id: this.createTurnId("autonomous"),
     };
-    this.lastContextWindowUsedTokens = undefined;
     this.notifySubscribers({ type: "turn_started", provider: "claude" });
     this.syncTurnState("autonomous turn started");
   }
@@ -2349,7 +2344,6 @@ class ClaudeAgentSession implements AgentSession {
     }
     this.notifySubscribers({ type: "turn_completed", provider: "claude" });
     this.autonomousTurn = null;
-    this.lastContextWindowUsedTokens = undefined;
     this.syncTurnState("autonomous turn completed");
   }
 
@@ -2920,8 +2914,14 @@ class ClaudeAgentSession implements AgentSession {
       usage.contextWindowMaxTokens = contextWindowMaxTokens;
     }
     if (typeof this.lastContextWindowUsedTokens === "number") {
+      // task_progress.total_tokens is the accurate context window fill level.
+      // Prefer it over result.usage which contains accumulated session totals.
       usage.contextWindowUsedTokens = this.lastContextWindowUsedTokens;
     } else if (message.usage) {
+      // Fallback: derive from result.usage when no task_progress has been
+      // received yet. These values are accumulated across all API calls, but
+      // for the first turn they equal the per-call values so the estimate is
+      // reasonable. Once a task_progress arrives it takes over permanently.
       const usageWithCacheCreation = message.usage as typeof message.usage & {
         cache_creation_input_tokens?: number;
       };