improve(gateway): compose live session stress probes (#122519)

* test(gateway): compose live session stress probes

Amp-Thread-ID: https://ampcode.com/threads/T-019feaaa-c7ed-769e-9f29-a3612bec72e7

* fix(ai): resume after Responses compaction checkpoints

Amp-Thread-ID: https://ampcode.com/threads/T-019feaaa-c7ed-769e-9f29-a3612bec72e7

* test(gateway): compose multi-session subagent probes

Amp-Thread-ID: https://ampcode.com/threads/T-019feaaa-c7ed-769e-9f29-a3612bec72e7

* fix(test): invalidate inlined schema transforms

Amp-Thread-ID: https://ampcode.com/threads/T-019feaaa-c7ed-769e-9f29-a3612bec72e7

* test(ai): cover empty compaction owners

Amp-Thread-ID: https://ampcode.com/threads/T-019feaaa-c7ed-769e-9f29-a3612bec72e7

---------

Co-authored-by: Amp <amp@ampcode.com>

docs/reference/transcript-hygiene.md

@@ -77,9 +77,9 @@ Implementation:
 - Max image side is configurable via `agents.defaults.imageMaxDimensionPx`
   (default: `1200`)
 - Blank text blocks are removed while this pass walks replay content.
-  Assistant turns that become empty are dropped from the replay copy; user
-  and tool-result turns that become empty receive a non-empty
-  omitted-content placeholder.
+  Assistant turns that become empty are dropped unless they own opaque
+  provider replay state; user and tool-result turns that become empty receive
+  a non-empty omitted-content placeholder.
 
 ---
 

packages/ai/src/transports/openai-responses-compaction-replay.test.ts

@@ -928,19 +928,15 @@ describe("OpenAI Responses compaction replay", () => {
     expect(input.map((item) => item.type)).toEqual(["compaction", "message"]);
   });
 
-  it("replays when session and auth identities match", () => {
-    const assistant = createOutput();
-    assistant.providerReplay = compactionState(model, { replayIndex: 0 });
+  it.each(responseConverters)(
+    "$name replays an empty checkpoint owner when request identities match",
+    ({ convert }) => {
+      const assistant = createOutput();
+      assistant.providerReplay = compactionState(model, { replayIndex: 0 });
 
-    const input = convertResponsesMessages(
-      model,
-      { messages: [assistant] },
-      new Set(["openai"]),
-      replayIdentity,
-    );
-
-    expect(input.some((item) => item.type === "compaction")).toBe(true);
-  });
+      expect(convert({ messages: [assistant] }).map((item) => item.type)).toEqual(["compaction"]);
+    },
+  );
 
   it.each(responseConverters)(
     "$name does not replay or prune across a different or missing request identity",

scripts/lib/state-schema-inline-plugin.mts

@@ -21,9 +21,18 @@ export function createStateSchemaInlinePlugin(rootDir = process.cwd()) {
   const schemasByModulePath = new Map(
     STATE_SCHEMA_MODULES.map((schema) => [path.resolve(rootDir, schema.modulePath), schema]),
   );
+  const cacheKeyForSchema = ({ id }: { id: string }) => {
+    const schema = schemasByModulePath.get(path.resolve(id));
+    return schema ? fs.readFileSync(path.resolve(rootDir, schema.schemaPath), "utf8") : undefined;
+  };
 
   return {
     name: STATE_SCHEMA_INLINE_PLUGIN_NAME,
+    configureVitest(context: {
+      experimental_defineCacheKeyGenerator(callback: typeof cacheKeyForSchema): void;
+    }) {
+      context.experimental_defineCacheKeyGenerator(cacheKeyForSchema);
+    },
     load(this: { addWatchFile(id: string): void }, id: string) {
       const schema = schemasByModulePath.get(path.resolve(id));
       if (!schema) {

src/agents/embedded-agent-helpers.sanitize-session-messages-images.removes-empty-assistant-text-blocks-but-preserves.test.ts

@@ -250,6 +250,31 @@ describe("sanitizeSessionMessagesImages", () => {
     expect(out).toHaveLength(1);
     expect(out[0]?.role).toBe("user");
   });
+  it.each([
+    ["full", "length"],
+    ["images-only", "length"],
+    ["full", "error"],
+    ["images-only", "error"],
+  ] as const)(
+    "preserves an empty provider replay owner in %s mode after %s",
+    async (sanitizeMode, stopReason) => {
+      const checkpoint = {
+        ...makeOpenAiResponsesAssistantMessage([{ type: "text", text: "" }], stopReason),
+        providerReplay: {
+          v: 1,
+          type: "opaque-checkpoint",
+          data: "opaque-state",
+          provider: "openai",
+          api: "openai-responses",
+          model: "gpt-5.4",
+        },
+      } satisfies AssistantMessage;
+
+      const out = await sanitizeSessionMessagesImages([checkpoint], "test", { sanitizeMode });
+
+      expect(out).toEqual([{ ...checkpoint, content: [] }]);
+    },
+  );
   it("drops empty assistant error messages", async () => {
     const input = castAgentMessages([
       { role: "user", content: "hello", timestamp: nextTimestamp() } satisfies UserMessage,

src/agents/embedded-agent-helpers/images.ts

@@ -53,8 +53,6 @@ export async function sanitizeSessionMessagesImages(
     };
   } & ImageSanitizationLimits,
 ): Promise<AgentMessage[]> {
-  const sanitizeMode = options?.sanitizeMode ?? "full";
-  const allowNonImageSanitization = sanitizeMode === "full";
   const imageSanitization = {
     maxDimensionPx: options?.maxDimensionPx,
     maxBytes: options?.maxBytes,
@@ -113,7 +111,7 @@ export async function sanitizeSessionMessagesImages(
             imageSanitization,
           )) as unknown as typeof assistantMsg.content;
           const finalContent = dropEmptyTextBlocks(nextContent);
-          if (finalContent.length > 0) {
+          if (finalContent.length > 0 || assistantMsg.providerReplay) {
             out.push({ ...assistantMsg, content: finalContent });
           }
         } else {
@@ -126,28 +124,14 @@ export async function sanitizeSessionMessagesImages(
         const strippedContent = options?.preserveSignatures
           ? content // Keep signatures for Antigravity Claude
           : stripThoughtSignatures(content, options?.sanitizeThoughtSignatures); // Strip for Gemini
-        if (!allowNonImageSanitization) {
-          const nextContent = (await sanitizeContentBlocksImages(
-            dropEmptyTextBlocks(strippedContent) as unknown as ContentBlock[],
-            label,
-            imageSanitization,
-          )) as unknown as typeof assistantMsg.content;
-          if (nextContent.length > 0) {
-            out.push({ ...assistantMsg, content: nextContent });
-          }
-          continue;
-        }
-
-        const filteredContent = dropEmptyTextBlocks(strippedContent);
         const finalContent = (await sanitizeContentBlocksImages(
-          filteredContent as unknown as ContentBlock[],
+          dropEmptyTextBlocks(strippedContent) as unknown as ContentBlock[],
           label,
           imageSanitization,
         )) as unknown as typeof assistantMsg.content;
-        if (finalContent.length === 0) {
-          continue;
+        if (finalContent.length > 0 || assistantMsg.providerReplay) {
+          out.push({ ...assistantMsg, content: finalContent });
         }
-        out.push({ ...assistantMsg, content: finalContent });
         continue;
       }
     }

src/agents/embedded-agent-runner/run.incomplete-turn.empty-response-recovery.test.ts

@@ -134,6 +134,48 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
     expectWarnMessageWith("empty response detected");
   });
 
+  it("continues after an OpenAI Responses compaction-only incomplete turn", async () => {
+    const checkpoint = makeLastAssistant({
+      api: "openai-responses",
+      provider: "openai",
+      model: "gpt-5.6-luna",
+      stopReason: "length",
+      providerReplay: {
+        v: 1,
+        type: "openai-responses-compaction",
+        data: "opaque-checkpoint",
+        provider: "openai",
+        api: "openai-responses",
+        model: "gpt-5.6-luna",
+      },
+    });
+    mockedRunEmbeddedAttempt.mockResolvedValueOnce(
+      makeAttemptResult({
+        assistantTexts: [],
+        currentAttemptAssistant: checkpoint,
+        lastAssistant: checkpoint,
+      }),
+    );
+    mockedRunEmbeddedAttempt.mockResolvedValueOnce(
+      makeAttemptResult({
+        assistantTexts: ["Visible answer after compaction."],
+        lastAssistant: makeLastAssistant({
+          content: [{ type: "text", text: "Visible answer after compaction." }],
+        }),
+      }),
+    );
+
+    await runEmbeddedAgent(
+      makeRunParams("run-provider-compaction-continuation", {
+        provider: "openai",
+        model: "gpt-5.6-luna",
+      }),
+    );
+
+    expect(mockedRunEmbeddedAttempt).toHaveBeenCalledTimes(2);
+    expectWarnMessageWith("compaction interrupted visible final answer");
+  });
+
   it("retries empty Anthropic-compatible stop turns even when the provider is not Kimi", async () => {
     mockedClassifyFailoverReason.mockReturnValue(null);
     mockedResolveModelAsync.mockResolvedValue({

src/agents/embedded-agent-runner/run/terminal-resolution.ts

@@ -341,7 +341,8 @@ export async function resolveEmbeddedRunTerminal(input: {
   if (
     !emptyAssistantReplyIsSilent &&
     !settledTurnFinalizationAttempted &&
-    input.attemptCompactionCount > 0 &&
+    (input.attemptCompactionCount > 0 ||
+      attempt.currentAttemptAssistant?.providerReplay?.type === "openai-responses-compaction") &&
     payloadCount === 0 &&
     !terminalInterrupted &&
     !promptError &&

src/gateway/gateway-codex-harness.live-helpers.test.ts

@@ -62,19 +62,21 @@ describe("gateway codex harness live helpers", () => {
       guardianProbe: false,
       imageProbe: false,
       mcpProbe: false,
+      multiSessionProbe: false,
       resumeStress: false,
       subagentProbe: true,
     };
 
     expect(shouldUseCodexHarnessSubagentOnlyFastPath(base)).toBe(true);
-    expect(shouldUseCodexHarnessSubagentOnlyFastPath({ ...base, resumeStress: true })).toBe(false);
-    expect(shouldUseCodexHarnessSubagentOnlyFastPath({ ...base, compactionStress: true })).toBe(
-      false,
-    );
-    expect(shouldUseCodexHarnessSubagentOnlyFastPath({ ...base, codeModeOnly: true })).toBe(false);
-    expect(shouldUseCodexHarnessSubagentOnlyFastPath({ ...base, explicitOptOut: true })).toBe(
-      false,
-    );
+    for (const flag of [
+      "codeModeOnly",
+      "compactionStress",
+      "explicitOptOut",
+      "multiSessionProbe",
+      "resumeStress",
+    ] as const) {
+      expect(shouldUseCodexHarnessSubagentOnlyFastPath({ ...base, [flag]: true })).toBe(false);
+    }
   });
 
   it("classifies sessions.list timeouts as retryable live Codex errors", () => {

src/gateway/gateway-codex-harness.live-helpers.ts

@@ -96,6 +96,7 @@ export function shouldUseCodexHarnessSubagentOnlyFastPath(params: {
   guardianProbe: boolean;
   imageProbe: boolean;
   mcpProbe: boolean;
+  multiSessionProbe: boolean;
   resumeStress: boolean;
   subagentProbe: boolean;
 }): boolean {
@@ -107,6 +108,7 @@ export function shouldUseCodexHarnessSubagentOnlyFastPath(params: {
     !params.guardianProbe &&
     !params.imageProbe &&
     !params.mcpProbe &&
+    !params.multiSessionProbe &&
     !params.resumeStress &&
     !params.explicitOptOut
   );

src/gateway/gateway-codex-harness.live.test.ts

@@ -148,6 +148,7 @@ const CODEX_HARNESS_SUBAGENT_ONLY = shouldUseCodexHarnessSubagentOnlyFastPath({
   guardianProbe: CODEX_HARNESS_GUARDIAN_PROBE,
   imageProbe: CODEX_HARNESS_IMAGE_PROBE,
   mcpProbe: CODEX_HARNESS_MCP_PROBE,
+  multiSessionProbe: CODEX_HARNESS_MULTI_SESSION_PROBE,
   resumeStress: CODEX_HARNESS_RESUME_STRESS,
   subagentProbe: CODEX_HARNESS_SUBAGENT_PROBE,
 });
@@ -2209,7 +2210,6 @@ describeLive("gateway live (Codex harness)", () => {
                 },
                 workspace,
               });
-              break;
             }
 
             if (CODEX_HARNESS_SUBAGENT_PROBE) {

src/gateway/gateway-openai-long-context.live.test.ts

@@ -503,8 +503,12 @@ describeLive("Gateway OpenAI long-context compaction (live)", () => {
         }
       }
       if (!compactionState?.latest) {
+        const thresholdEvidence =
+          peakPromptTokens > 0
+            ? `peak provider prompt tokens=${peakPromptTokens}, compact threshold=${profile.compactThreshold}`
+            : "provider prompt-token usage unavailable";
         throw new Error(
-          `OpenAI emitted no first-class compaction item after ${profile.maxDenseTurns} dense turns`,
+          `OpenAI emitted no first-class compaction item after ${profile.maxDenseTurns} dense turns; ${thresholdEvidence}`,
         );
       }
       expect(compactionState.latest).toMatchObject({

src/infra/tsdown-config.test.ts

@@ -111,6 +111,12 @@ describe("tsdown config", () => {
     const rootDir = process.cwd();
     const watchedPaths: string[] = [];
     const plugin = createStateSchemaInlinePlugin(rootDir);
+    let cacheKeyGenerator: ((context: { id: string }) => string | undefined) | undefined;
+    plugin.configureVitest({
+      experimental_defineCacheKeyGenerator: (generator) => {
+        cacheKeyGenerator = generator;
+      },
+    });
     const result = plugin.load.call(
       { addWatchFile: (filePath: string) => watchedPaths.push(filePath) },
       path.resolve(rootDir, schema.modulePath),
@@ -126,6 +132,10 @@ describe("tsdown config", () => {
     expect(JSON.parse(match?.[1] ?? "null")).toBe(canonicalSql);
     expect(schema.sourceValue).toBe(canonicalSql);
     expect(watchedPaths).toEqual([schemaPath]);
+    expect(cacheKeyGenerator?.({ id: path.resolve(rootDir, schema.modulePath) })).toBe(
+      canonicalSql,
+    );
+    expect(cacheKeyGenerator?.({ id: path.resolve(rootDir, "src/index.ts") })).toBeUndefined();
   });
 
   it("installs schema inlining only on the unified runtime graph", () => {

test/gateway-openai-compaction-replay.e2e.test.ts

@@ -63,7 +63,9 @@ describe("Gateway OpenAI Responses compaction replay", () => {
       });
       try {
         await runAgentTurn(client, "capture compaction state");
-        expect(modelServer.requests).toHaveLength(1);
+        // The provider can terminate after emitting only a compaction item. The
+        // runner must continue from that checkpoint before completing the turn.
+        expect(modelServer.requests).toHaveLength(2);
 
         const session = await client.request<{
           sessions?: Array<{ key?: string; sessionId?: string }>;
@@ -78,28 +80,31 @@ describe("Gateway OpenAI Responses compaction replay", () => {
           sessionKey: SESSION_KEY,
           storePath: path.join(instance.state.agentDir("main"), "openclaw-agent.sqlite"),
         });
-        const persistedReplay = manager
-          .buildSessionContext()
-          .messages.find((message) => message.role === "assistant")?.providerReplay;
+        const contextMessages = manager.buildSessionContext().messages;
+        const persistedReplay = contextMessages.find(
+          (message) => message.role === "assistant",
+        )?.providerReplay;
         expect(persistedReplay).toMatchObject({
+          v: 1,
           type: "openai-responses-compaction",
           id: COMPACTION_ID,
           data: COMPACTION_DATA,
           provider: "replay-proof",
           api: "openai-responses",
           model: "replay-proof",
+          baseUrlHash: expect.any(String),
           sessionHash: expect.any(String),
         });
         expect(persistedReplay).not.toHaveProperty("authProfileHash");
+        expectCompactionReplay(modelServer.requests[1]?.body.input ?? []);
+        expect(JSON.stringify(modelServer.requests[1]?.body.input)).toContain(
+          "Continue from the compacted transcript",
+        );
 
         await runAgentTurn(client, "replay compaction state");
-        expect(modelServer.requests).toHaveLength(2);
-        const replayInput = modelServer.requests[1]?.body.input ?? [];
-        expect(replayInput).toContainEqual({
-          type: "compaction",
-          id: COMPACTION_ID,
-          encrypted_content: COMPACTION_DATA,
-        });
+        expect(modelServer.requests).toHaveLength(3);
+        const replayInput = modelServer.requests[2]?.body.input ?? [];
+        expectCompactionReplay(replayInput);
         const compactionIndex = replayInput.findIndex(
           (item) =>
             typeof item === "object" &&
@@ -120,7 +125,7 @@ describe("Gateway OpenAI Responses compaction replay", () => {
         ).toBe(true);
         const encodedReplayInput = JSON.stringify(replayInput);
         expect(encodedReplayInput).not.toContain("capture compaction state");
-        expect(encodedReplayInput).toContain("gateway replay response 1");
+        expect(encodedReplayInput).toContain("gateway replay response 2");
         expect(encodedReplayInput).toContain("replay compaction state");
       } finally {
         await disconnectGatewayClient(client);
@@ -188,6 +193,14 @@ async function runAgentTurn(
   return runId;
 }
 
+function expectCompactionReplay(input: unknown[]): void {
+  expect(input).toContainEqual({
+    type: "compaction",
+    id: COMPACTION_ID,
+    encrypted_content: COMPACTION_DATA,
+  });
+}
+
 async function startMockModelServer(): Promise<MockModelServer> {
   const requests: CapturedRequest[] = [];
   const server = createServer((request, response) => {
@@ -241,6 +254,28 @@ async function handleRequest(
 }
 
 function writeModelResponse(response: ServerResponse, sequence: number): void {
+  if (sequence === 1) {
+    const compaction = {
+      type: "compaction",
+      id: COMPACTION_ID,
+      encrypted_content: COMPACTION_DATA,
+    };
+    writeSseEvents(response, [
+      { type: "response.output_item.added", output_index: 0, item: compaction },
+      { type: "response.output_item.done", output_index: 0, item: compaction },
+      {
+        type: "response.incomplete",
+        response: {
+          id: "resp_gateway_replay_1",
+          status: "incomplete",
+          incomplete_details: { reason: "max_output_tokens" },
+          output: [compaction],
+          usage: { input_tokens: 0, output_tokens: 0, total_tokens: 0 },
+        },
+      },
+    ]);
+    return;
+  }
   const text = `gateway replay response ${sequence}`;
   const message = {
     type: "message",
@@ -249,10 +284,7 @@ function writeModelResponse(response: ServerResponse, sequence: number): void {
     status: "completed",
     content: [{ type: "output_text", text, annotations: [] }],
   };
-  const output =
-    sequence === 1
-      ? [{ type: "compaction", id: COMPACTION_ID, encrypted_content: COMPACTION_DATA }, message]
-      : [message];
+  const output = [message];
   const events: MockSseEvent[] = output.flatMap((item, outputIndex) => [
     {
       type: "response.output_item.added",
@@ -270,6 +302,10 @@ function writeModelResponse(response: ServerResponse, sequence: number): void {
       usage: { input_tokens: 10, output_tokens: 2, total_tokens: 12 },
     },
   });
+  writeSseEvents(response, events);
+}
+
+function writeSseEvents(response: ServerResponse, events: MockSseEvent[]): void {
   response.writeHead(200, {
     "content-type": "text/event-stream",
     "cache-control": "no-store",

test/helpers/openai-long-context-live.test.ts

@@ -118,7 +118,7 @@ describe("OpenAI long-context live settings", () => {
       contextWindow: 48_000,
       contextTokens: 48_000,
       maxTokens: 8_192,
-      compactThreshold: 32_000,
+      compactThreshold: 1_000,
     });
     const full = resolveOpenAILongContextLiveSettings(
       {

test/helpers/openai-long-context-live.ts

@@ -48,9 +48,12 @@ const PROFILES = {
     contextWindow: 48_000,
     contextTokens: 48_000,
     maxTokens: 8_192,
-    compactThreshold: 32_000,
+    // Keep the reduced live probe on OpenAI's demonstrated compaction path.
+    // High-threshold Luna probes can cross the configured threshold without
+    // emitting a checkpoint, while the 1k boundary is deterministic.
+    compactThreshold: 1_000,
     denseTurnChars: 120_000,
-    maxDenseTurns: 8,
+    maxDenseTurns: 3,
     defaultToolBytes: 300_000,
     requestTimeoutMs: 2 * 60_000,
     suiteTimeoutMs: 10 * 60_000,
@@ -199,6 +202,9 @@ export function buildOpenAILongContextConfig(params: {
         workspace: params.workspace,
         skipBootstrap: true,
         thinkingDefault: "low",
+        // This suite owns the server-compaction threshold. Embedded proactive
+        // compaction would consume the same history before replay can be proved.
+        compaction: { enabled: false },
         model: { primary: profile.modelRef },
         models: {
           [profile.modelRef]: {
@@ -256,6 +262,11 @@ export function assertOpenAILongContextConfig(
     cfg.secrets?.providers?.default?.source,
     "env",
   );
+  expectConfigValue(
+    "agents.defaults.compaction.enabled",
+    cfg.agents?.defaults?.compaction?.enabled,
+    false,
+  );
   expectConfigValue("models.providers.openai.models.length", provider?.models.length, 1);
   const model = provider?.models[0];
   expectConfigValue("model.id", model?.id, profile.modelId);