Agents: drop stale pre-compaction usage snapshots

CHANGELOG.md

@@ -32,6 +32,7 @@ Docs: https://docs.openclaw.ai
 - TUI/Status: request immediate renders after setting `sending`/`waiting` activity states so in-flight runs always show visible progress indicators instead of appearing idle until completion. (#21549) Thanks @13Guinness.
 - Agents/Fallbacks: treat JSON payloads with `type: "api_error"` + `"Internal server error"` as transient failover errors so Anthropic 500-style failures trigger model fallback. (#23193) Thanks @jarvis-lane.
 - Agents/Transcripts: validate assistant tool-call names (syntax/length + registered tool allowlist) before transcript persistence and during replay sanitization so malformed failover tool names no longer poison sessions with repeated provider HTTP 400 errors. (#23324) Thanks @johnsantry.
+- Agents/Compaction: strip stale assistant usage snapshots from pre-compaction turns when replaying history after a compaction summary so context-token estimation no longer reuses pre-compaction totals and immediately re-triggers destructive follow-up compactions. (#19127) Thanks @tedwatson.
 - Agents/Diagnostics: include resolved lifecycle error text in `embedded run agent end` warnings so UI/TUI “Connection error” runs expose actionable provider failure reasons in gateway logs. (#23054) Thanks @Raize.
 - Plugins/Hooks: run legacy `before_agent_start` once per agent turn and reuse that result across model-resolve and prompt-build compatibility paths, preventing duplicate hook side effects (for example duplicate external API calls). (#23289) Thanks @ksato8710.
 - Models/Config: default missing Anthropic provider/model `api` fields to `anthropic-messages` during config validation so custom relay model entries are preserved instead of being dropped by runtime model registry validation. (#23332) Thanks @bigbigmonkey123.

src/agents/pi-embedded-runner.sanitize-session-history.test.ts

@@ -158,6 +158,102 @@ describe("sanitizeSessionHistory", () => {
     expect(first.content as string).toContain("sourceSession=agent:main:req");
   });
 
+  it("drops stale assistant usage snapshots kept before latest compaction summary", async () => {
+    vi.mocked(helpers.isGoogleModelApi).mockReturnValue(false);
+
+    const messages = [
+      { role: "user", content: "old context" },
+      {
+        role: "assistant",
+        content: [{ type: "text", text: "old answer" }],
+        stopReason: "stop",
+        usage: {
+          input: 191_919,
+          output: 2_000,
+          cacheRead: 0,
+          cacheWrite: 0,
+          totalTokens: 193_919,
+          cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+        },
+      },
+      {
+        role: "compactionSummary",
+        summary: "compressed",
+        tokensBefore: 191_919,
+        timestamp: new Date().toISOString(),
+      },
+    ] as unknown as AgentMessage[];
+
+    const result = await sanitizeSessionHistory({
+      messages,
+      modelApi: "openai-responses",
+      provider: "openai",
+      sessionManager: mockSessionManager,
+      sessionId: TEST_SESSION_ID,
+    });
+
+    const staleAssistant = result.find((message) => message.role === "assistant") as
+      | (AgentMessage & { usage?: unknown })
+      | undefined;
+    expect(staleAssistant).toBeDefined();
+    expect(staleAssistant?.usage).toBeUndefined();
+  });
+
+  it("preserves fresh assistant usage snapshots created after latest compaction summary", async () => {
+    vi.mocked(helpers.isGoogleModelApi).mockReturnValue(false);
+
+    const messages = [
+      {
+        role: "assistant",
+        content: [{ type: "text", text: "pre-compaction answer" }],
+        stopReason: "stop",
+        usage: {
+          input: 120_000,
+          output: 3_000,
+          cacheRead: 0,
+          cacheWrite: 0,
+          totalTokens: 123_000,
+          cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+        },
+      },
+      {
+        role: "compactionSummary",
+        summary: "compressed",
+        tokensBefore: 123_000,
+        timestamp: new Date().toISOString(),
+      },
+      { role: "user", content: "new question" },
+      {
+        role: "assistant",
+        content: [{ type: "text", text: "fresh answer" }],
+        stopReason: "stop",
+        usage: {
+          input: 1_000,
+          output: 250,
+          cacheRead: 0,
+          cacheWrite: 0,
+          totalTokens: 1_250,
+          cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+        },
+      },
+    ] as unknown as AgentMessage[];
+
+    const result = await sanitizeSessionHistory({
+      messages,
+      modelApi: "openai-responses",
+      provider: "openai",
+      sessionManager: mockSessionManager,
+      sessionId: TEST_SESSION_ID,
+    });
+
+    const assistants = result.filter((message) => message.role === "assistant") as Array<
+      AgentMessage & { usage?: unknown }
+    >;
+    expect(assistants).toHaveLength(2);
+    expect(assistants[0]?.usage).toBeUndefined();
+    expect(assistants[1]?.usage).toBeDefined();
+  });
+
   it("keeps reasoning-only assistant messages for openai-responses", async () => {
     setNonGoogleModelApi();
 

src/agents/pi-embedded-runner/google.ts

@@ -214,6 +214,35 @@ function annotateInterSessionUserMessages(messages: AgentMessage[]): AgentMessag
   return touched ? out : messages;
 }
 
+function stripStaleAssistantUsageBeforeLatestCompaction(messages: AgentMessage[]): AgentMessage[] {
+  let latestCompactionSummaryIndex = -1;
+  for (let i = 0; i < messages.length; i += 1) {
+    if (messages[i]?.role === "compactionSummary") {
+      latestCompactionSummaryIndex = i;
+    }
+  }
+  if (latestCompactionSummaryIndex <= 0) {
+    return messages;
+  }
+
+  const out = [...messages];
+  let touched = false;
+  for (let i = 0; i < latestCompactionSummaryIndex; i += 1) {
+    const candidate = out[i] as (AgentMessage & { usage?: unknown }) | undefined;
+    if (!candidate || candidate.role !== "assistant") {
+      continue;
+    }
+    if (!candidate.usage || typeof candidate.usage !== "object") {
+      continue;
+    }
+    const candidateRecord = candidate as unknown as Record<string, unknown>;
+    const { usage: _droppedUsage, ...rest } = candidateRecord;
+    out[i] = rest as unknown as AgentMessage;
+    touched = true;
+  }
+  return touched ? out : messages;
+}
+
 function findUnsupportedSchemaKeywords(schema: unknown, path: string): string[] {
   if (!schema || typeof schema !== "object") {
     return [];
@@ -466,6 +495,8 @@ export async function sanitizeSessionHistory(params: {
     ? sanitizeToolUseResultPairing(sanitizedToolCalls)
     : sanitizedToolCalls;
   const sanitizedToolResults = stripToolResultDetails(repairedTools);
+  const sanitizedCompactionUsage =
+    stripStaleAssistantUsageBeforeLatestCompaction(sanitizedToolResults);
 
   const isOpenAIResponsesApi =
     params.modelApi === "openai-responses" || params.modelApi === "openai-codex-responses";
@@ -480,8 +511,8 @@ export async function sanitizeSessionHistory(params: {
       })
     : false;
   const sanitizedOpenAI = isOpenAIResponsesApi
-    ? downgradeOpenAIReasoningBlocks(sanitizedToolResults)
+    ? downgradeOpenAIReasoningBlocks(sanitizedCompactionUsage)
-    : sanitizedToolResults;
+    : sanitizedCompactionUsage;
 
   if (hasSnapshot && (!priorSnapshot || modelChanged)) {
     appendModelSnapshot(params.sessionManager, {