fix: 继续收敛 codex trace 残余噪声 (#68)

Co-authored-by: Codex <codex@pikas.tech>

docs/reference/spec-v01-agentrun-runner.md

@@ -62,7 +62,7 @@ claimed -> lease_lost
 规则：
 
 - runner 必须先 register，再 claim run；claim 失败不能继续调用 backend。
-- lease heartbeat 必须可观察；过期或冲突时写入 failure event 或明确退出原因。
+- lease heartbeat 必须通过 manager lease/status 可观察；不得把周期性心跳或 backend running tick 写成 durable trace event 刷屏。长 turn 只在 `backend-turn-finished` 中输出有界 progress 摘要；过期或冲突时写入 failure event 或明确退出原因。
 - command 只能从 manager poll；不得从本地文件或临时参数伪造正式 command。
 - runner 的普通 poll 只选择 pending `turn`；当 backend adapter 暴露 active turn control 后，runner 才在同 run 内轮询 pending `steer` command，ack 后调用 backend 的 steer 能力并单独终结该 steer command。active turn 结束后到达的 steer 必须结构化 blocked，不得启动新 turn，也不得把 run 标为 terminal。
 - backend 产生的所有可见输出必须先经过 adapter normalization 和 redaction，再 append 到 manager；backend_status 至少包含 redacted profile/backendKind/protocol 摘要。

docs/reference/spec-v01-backend-adapter.md

@@ -69,7 +69,7 @@ Adapter 输出给 runner 的 event 类型至少包括：
 
 事件必须有上限和分页友好形态。大型日志、完整 stdout 或完整 trace 应进入 logPath 或后续 artifact，不得一次性塞入单个 event 造成输出爆炸。
 
-Codex app-server 的低价值内部 notification 必须在 AgentRun adapter 层收敛，不得要求 HWLAB Web/CLI 或其他消费侧自行过滤。以下事件默认不作为 durable trace event 持久化：`item/reasoning/textDelta`、纯 `reasoning` item 的 `item/started|item/completed`、`thread/tokenUsage/updated`、`account/rateLimits/updated`、普通 `warning` 和 `configWarning`。adapter 可以输出一条有界 `backend_status.phase=codex-app-server-notifications-suppressed` 摘要，只包含计数、method 和 item type，不包含 reasoning 文本、Secret、token 或 env value。真实 `agentMessage`、`commandExecution`、`command_output`、error、terminal 和生命周期事件必须继续保留。
+Codex app-server 的低价值内部 notification 必须在 AgentRun adapter 层收敛，不得要求 HWLAB Web/CLI 或其他消费侧自行过滤。以下事件默认不作为 durable trace event 持久化：`item/reasoning/textDelta`、纯 `reasoning` item 的 `item/started|item/completed`、非 `commandExecution` item 的通用 `item/started|item/completed`、`thread/tokenUsage/updated`、`account/rateLimits/updated`、普通 `warning` 和 `configWarning`。adapter 可以输出一条有界 `backend_status.phase=codex-app-server-notifications-suppressed` 摘要，只包含计数、method 和 item type，不包含 reasoning 文本、Secret、token 或 env value。真实 `agentMessage`、`commandExecution`、`command_output`、error、terminal 和关键生命周期事件必须继续保留。
 
 ## Failure Mapping
 

src/backend/codex-stdio.ts

@@ -606,7 +606,7 @@ function normalizeCodexNotification(message: JsonRecord, suppressed: SuppressedN
   if (method === "item/started" || method === "item/completed") {
     const item = asRecordAt(params, "item");
     const itemType = typeof item.type === "string" ? item.type : "unknown";
-    if (isSuppressedCodexItemType(itemType)) {
+    if (itemType !== "commandExecution" || isSuppressedCodexItemType(itemType)) {
       recordSuppressedNotification(suppressed, method, itemType);
       return { events: [] };
     }

src/runner/run-once.ts

@@ -153,7 +153,7 @@ async function executeCommand(api: RunnerManagerApi, options: RunnerOnceOptions,
   await api.appendEvent(options.runId, { type: "backend_status", payload: { phase: "backend-turn-started", commandId: command.id, attemptId, runnerId: runner.id, backendProfile: options.backendProfile ?? null, workspaceReady: Boolean(workspacePath) } });
   const abortController = new AbortController();
   const stopCancelWatch = watchCancellation(api, options.runId, command.id, abortController);
-  const stopBackendProgress = startBackendProgress(api, options.runId, command.id, attemptId, runner.id, options.backendProfile ?? null);
+  const backendProgress = startBackendProgress();
   let stopSteerWatch: (() => void) | undefined;
   try {
     const latestRun = await api.getRun(options.runId);
@@ -184,8 +184,8 @@ async function executeCommand(api: RunnerManagerApi, options: RunnerOnceOptions,
     return await reportCommandFailure(api, options.runId, command.id, runner, attemptId, failure, "runner:execute");
   } finally {
     stopSteerWatch?.();
-    stopBackendProgress();
-    await appendBestEffort(api, options.runId, { type: "backend_status", payload: { phase: "backend-turn-finished", commandId: command.id, attemptId, runnerId: runner.id } });
+    const progressSummary = backendProgress.stop();
+    await appendBestEffort(api, options.runId, { type: "backend_status", payload: { phase: "backend-turn-finished", commandId: command.id, attemptId, runnerId: runner.id, ...progressSummary } });
     stopCancelWatch();
   }
 }
@@ -307,20 +307,21 @@ function startHeartbeat(api: RunnerManagerApi, runId: string, runnerId: string,
   };
 }
 
-function startBackendProgress(api: RunnerManagerApi, runId: string, commandId: string, attemptId: string, runnerId: string, backendProfile: string | null): () => void {
+function startBackendProgress(): { stop: () => JsonRecord } {
   let stopped = false;
   let ticks = 0;
   const startedAt = Date.now();
-  const emit = async (): Promise<void> => {
+  const tick = (): void => {
     if (stopped) return;
     ticks += 1;
-    await appendBestEffort(api, runId, { type: "backend_status", payload: { phase: "backend-turn-running", commandId, attemptId, runnerId, backendProfile, elapsedMs: Date.now() - startedAt, ticks } });
   };
-  const timer = setInterval(() => { void emit(); }, 10_000);
-  void emit();
-  return () => {
-    stopped = true;
-    clearInterval(timer);
+  const timer = setInterval(tick, 10_000);
+  return {
+    stop: () => {
+      stopped = true;
+      clearInterval(timer);
+      return { elapsedMs: Date.now() - startedAt, progressTicks: ticks, progressEventsPrinted: false };
+    },
   };
 }
 

src/selftest/cases/30-codex-stdio.ts

@@ -125,6 +125,7 @@ const selfTest: SelfTestCase = async (context) => {
     assert.equal(noisyResult.terminalStatus, "completed", "noisy reasoning turn should complete");
     const noisyEvents = await client.get(`/api/v1/runs/${noisy.runId}/events?afterSeq=0&limit=100`) as { items?: Array<{ type: string; payload: unknown }> };
     const noisyItems = noisyEvents.items ?? [];
+    const noisyPhases = noisyItems.map((event) => eventPayload(event).phase).filter(Boolean);
     assert.equal(noisyItems.some((event) => event.type === "backend_status" && eventPayload(event).phase === "item/reasoning/textDelta"), false, "reasoning textDelta must not be persisted as backend_status");
     assert.equal(noisyItems.some((event) => event.type === "backend_status" && eventPayload(event).phase === "thread/tokenUsage/updated"), false, "token usage update must not be persisted as backend_status");
     assert.equal(noisyItems.some((event) => event.type === "backend_status" && eventPayload(event).phase === "account/rateLimits/updated"), false, "rate limit update must not be persisted as backend_status");
@@ -132,6 +133,10 @@ const selfTest: SelfTestCase = async (context) => {
     assert.equal(noisyItems.some((event) => event.type === "backend_status" && eventPayload(event).phase === "configWarning"), false, "low value config warnings must not be persisted as backend_status");
     assert.equal(noisyItems.some((event) => event.type === "tool_call" && eventPayloadItem(event).type === "reasoning"), false, "reasoning items must not be persisted as tool_call");
     assert.ok(noisyItems.some((event) => event.type === "tool_call" && eventPayload(event).method === "item/started" && eventPayloadItem(event).type === "commandExecution"), "real commandExecution tool call should remain visible");
+    assert.equal(noisyItems.some((event) => event.type === "tool_call" && eventPayloadItem(event).type !== "commandExecution"), false, "non-commandExecution item lifecycle must not be persisted as tool_call");
+    assert.equal(noisyPhases.includes("backend-turn-running"), false, "backend progress ticks must be summarized instead of persisted as durable trace events");
+    const noisyFinished = noisyItems.find((event) => event.type === "backend_status" && eventPayload(event).phase === "backend-turn-finished");
+    assert.equal(eventPayload(noisyFinished ?? { payload: {} }).progressEventsPrinted, false, "backend-turn-finished must declare progress ticks were not printed as events");
     assert.ok(noisyItems.some((event) => event.type === "assistant_message" && eventPayload(event).text === "noise filtered final"), "final assistant_message should remain visible");
     const suppression = noisyItems.find((event) => event.type === "backend_status" && eventPayload(event).phase === "codex-app-server-notifications-suppressed");
     assert.ok(suppression, "suppression summary must be emitted when noisy notifications are filtered");

src/selftest/cases/50-hwlab-manual-dispatch.ts

@@ -123,9 +123,9 @@ console.log(JSON.stringify({ apiVersion: manifest.apiVersion, kind: manifest.kin
     assert.equal(multiEvents.filter((event) => event.type === "backend_status" && event.payload?.phase === "resource-bundle-materialized").length, 1);
     for (const commandId of [multiTurn.commandId, secondCommand.id]) {
       assert.ok(multiEvents.some((event) => event.type === "backend_status" && event.payload?.phase === "backend-turn-started" && event.payload?.commandId === commandId), `command ${commandId} must emit backend-turn-started before waiting on Codex`);
-      assert.ok(multiEvents.some((event) => event.type === "backend_status" && event.payload?.phase === "backend-turn-running" && event.payload?.commandId === commandId), `command ${commandId} must emit backend-turn-running while Codex is active`);
       assert.ok(multiEvents.some((event) => event.type === "backend_status" && event.payload?.phase === "backend-turn-finished" && event.payload?.commandId === commandId), `command ${commandId} must emit backend-turn-finished after Codex returns`);
     }
+    assert.equal(multiEvents.some((event) => event.type === "backend_status" && event.payload?.phase === "backend-turn-running"), false, "backend-turn-running ticks must not be persisted as durable trace events");
     assert.equal(multiEvents.filter((event) => event.type === "backend_status" && event.payload?.phase === "command-terminal").length, 2);
     const secondEnvelope = await client.get(`/api/v1/runs/${multiTurn.runId}/commands/${secondCommand.id}/result`) as JsonRecord;
     assert.equal(secondEnvelope.terminalStatus, "completed");