diff --git a/docs/BUG_HISTORY.md b/docs/BUG_HISTORY.md index 3e306f40..233e392d 100644 --- a/docs/BUG_HISTORY.md +++ b/docs/BUG_HISTORY.md @@ -5531,6 +5531,22 @@ - 复发自：无 - 修复版本：未修复 +## BUG-372 | 生时纠正同一轮重复工具调用把已成功写入打成失败 + +- 状态：resolved +- 首次发现：2026-08-25 +- 最近更新：2026-08-25 +- 影响面：`POST /api/rectification/agent`、`runV9AgentTurn` 观察器、公开 NDJSON +- 用户现象：同一轮已读盘、写入证据、跑完诊断和候选比较后，流以 `run.failed` / `repeated_tool_call` 结束，文案「本轮没有完成，状态已记录。」，没有助手叙述。公开事件里每个工具只出现一次 started/completed。 +- 触发条件：`action=message` 的自由文本经历回合；模型在 `rectification-compare-candidates`（或同类只含 caseId 的公开工具）完成后，又发出一次相同工具名与相同参数的 `tool-call`。 +- 根因：BUG-368 P0-3 把「相同工具参数」做成观察器硬上限 1，第二次相同 `toolName + args` 在发布 `tool.activity` started 之前抛错。该码不在自动重试集合，`recoverable=false`。证据与比较已经落库，只是本轮被标失败且跳过 dossier 叙述器。`rectification-set-focus` 已从 Agent 工具列表移除，这条防线不再对应真实循环。 +- 修复：相同公开工具调用视为幂等，跳过重复的 started/phase 收据，不 abort、不 `attempt.reset`。真循环仍由 `maxSteps` 与超时约束。无模型正文时走既有服务器叙述。不改已哈希 Skill `10.0.11`。 +- 验证：`frontend/tests/rectification-v9-agent.test.ts` 锁定重复 read-case 完成叙述且无 `run.failed`；`frontend/tests/rectification-v9-stream.test.ts` 锁定重复 set-focus 继续出回答、诊断后重复 compare 走叙述且无 `attempt.reset`。 +- 防复发：禁止对第二次相同公开 `tool-call` 抛 `repeated_tool_call` 或失败整轮。禁止把幂等工具调用放进自动重试。禁止用观察器 abort 代替 `maxSteps` / timeout。 +- 相关记录：BUG-368、BUG-367 +- 复发自：BUG-368（去掉模型驱动 set-focus 后，仍用 abort-on-second-identical-call 防循环，误杀正常比较重发） +- 修复版本：待发布 + ## BUG-371 | staging publish 的 next build 找不到 createServerSupabaseClient - 状态：resolved @@ -5588,10 +5604,10 @@ - 用户现象：用户说「2020年4月开始实习、6月转正、10月离职」后，界面先刷出残缺英文（`Let me`、`_probe`、`_gain`）。`rectification-set-focus` 连续失败后整轮重跑，已成功的证据再提交一次，只落地实习、另外两条被引文拒。随后聊天里出现一条用户从未输入的「2002 年发生什么了」。 - 触发条件：自由文本经历回合；Agent 在调用工具前输出规划文本；`set-focus` 因重复探针或零信息增益被拒；失败运行的推荐问题被写进历史。 - 根因：三组独立回归。(1) 禁止 `thinking.delta` 后，把每个 step 的 `text-delta` 都公开成 `answer.delta`，工具前规划变成用户正文；再对碎片做英文过滤，句子被剪成残片。(2) `compare-candidates` 已算出下一问，仍让模型自己调 `set-focus`；确定性校验失败后 `attempt.reset` 重放已成功的 Evidence 写入。(3) 未完成运行的 suggestion / 角色映射把「2002 年发生什么了」写成 user 消息。2002 不是引擎从 2020 算出来的。 -- 修复：按 step 缓冲，只发布无工具且 `stop`/`length` 的终端文本；`reasoning-delta` 不下发。Agent 工具列表去掉 `set-focus`，由 compare/read-case 在服务端持久化 `open_question`。`duplicate_focus` 等域错误不整轮重试，相同工具参数最多一次。用户消息记录 `origin`/`clientActionId`/`content_hash`。Evidence quote 用源消息 offset，按条返回 created/already_exists/quote_mismatch。推荐问题只在 `run.completed` 后解析，纠正 UI 仍无 suggestion chip。不改已哈希 Skill `10.0.11`。 +- 修复：按 step 缓冲，只发布无工具且 `stop`/`length` 的终端文本；`reasoning-delta` 不下发。Agent 工具列表去掉 `set-focus`，由 compare/read-case 在服务端持久化 `open_question`。`duplicate_focus` 等域错误不整轮重试。同一流里相同公开工具参数视为幂等，不得 abort 整轮；循环由 `maxSteps` / timeout 约束。用户消息记录 `origin`/`clientActionId`/`content_hash`。Evidence quote 用源消息 offset，按条返回 created/already_exists/quote_mismatch。推荐问题只在 `run.completed` 后解析，纠正 UI 仍无 suggestion chip。不改已哈希 Skill `10.0.11`。 - 验证：`frontend/tests/rectification-step-answer.test.ts`、`frontend/tests/rectification-v9-stream.test.ts`、`frontend/tests/rectification-server-focus.test.ts`、`frontend/tests/rectification-evidence-quote.test.ts`、`frontend/tests/rectification-agentic-entry.test.ts`、`frontend/tests/rectification-v10-tool-contract.test.ts`。 -- 防复发：禁止把含工具调用的 step 的 `text-delta` 发给浏览器。禁止把 `thinking.delta` 改名为 `answer.delta`。禁止模型驱动 `set-focus`。禁止对 `duplicate_focus` / `quote_mismatch` / `zero_information_gain` 做 `attempt.reset`。禁止未完成运行解析或自动提交 suggestion。禁止模型改写 Evidence quote。 -- 相关记录：BUG-345、BUG-354、BUG-357、BUG-367 +- 防复发：禁止把含工具调用的 step 的 `text-delta` 发给浏览器。禁止把 `thinking.delta` 改名为 `answer.delta`。禁止模型驱动 `set-focus`。禁止对 `duplicate_focus` / `quote_mismatch` / `zero_information_gain` 做 `attempt.reset`。禁止对第二次相同公开工具调用 abort 整轮（见 BUG-372）。禁止未完成运行解析或自动提交 suggestion。禁止模型改写 Evidence quote。 +- 相关记录：BUG-345、BUG-354、BUG-357、BUG-367、BUG-372 - 复发自：BUG-367（关掉公开 `thinking.delta` 后，中间 step 的 `text-delta` 被整段当成回答） - 修复版本：fe87a9ecdb71eae4eb79664664cc20c534177c19 diff --git a/frontend/src/lib/rectification-agentic/v9/agent-run.ts b/frontend/src/lib/rectification-agentic/v9/agent-run.ts index 6ad99085..13bf9731 100644 --- a/frontend/src/lib/rectification-agentic/v9/agent-run.ts +++ b/frontend/src/lib/rectification-agentic/v9/agent-run.ts @@ -109,7 +109,6 @@ type AttemptOutcome = Readonly<{ attemptId: string; }>; -const REPEATED_TOOL_CALL_LIMIT = 1; const MAX_ATTEMPTS = 2; const RETRYABLE_ERROR_CODES = new Set([ "stream_aborted", @@ -137,6 +136,22 @@ function first(value: unknown): unknown { return value; } +/** + * Identity for a public tool-call chunk. Mastra may put the model input on + * `args`, `input`, or omit it; missing input collapses to `{}` so a second + * call of the same tool name still looks identical. + */ +function publicToolCallKey(toolName: string, payload: unknown): string { + if (!payload || typeof payload !== "object") return `${toolName}:{}`; + const record = payload as Record; + const args = record.args ?? record.input ?? record.toolArgs ?? {}; + try { + return `${toolName}:${JSON.stringify(args)}`; + } catch { + return `${toolName}:{}`; + } +} + async function rpcOf( accounting: RectificationRpcClient, fn: string, @@ -622,6 +637,12 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise REPEATED_TOOL_CALL_LIMIT) throw new Error("repeated_tool_call"); + skipDuplicateToolCallReceipt = count > 1; } } @@ -644,7 +665,9 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise { - const { options, billing } = runOptions({ +test("a repeated identical tool call is treated as idempotent and does not abort", async () => { + const { options, emitted, billing } = runOptions({ buildAgent: async () => fakeAgentStream([ chunk("start"), chunk("tool-call", { toolName: "skill", args: { name: RECTIFICATION_SKILL_NAME } }), chunk("tool-result", { toolName: "skill" }), - ...Array.from({ length: 2 }, () => chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } })), + chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }), + chunk("tool-result", { toolName: "rectification-read-case" }), + chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }), chunk("finish"), ]) as never, }); const result = await runV9AgentTurn(options); - assert.equal(result.ok, false); - assert.equal(result.errorCode, "repeated_tool_call"); - assert.equal(billing.released, 1); + assert.equal(result.ok, true); + assert.equal(result.errorCode, null); + assert.match(result.answerText, /已经记下|请继续说下一件/); + assert.equal(billing.completed, 1); + assert.equal(billing.released, 0); + assert.equal(emitted.some((event) => event.type === "attempt.reset"), false); + assert.equal(emitted.some((event) => event.type === "run.failed"), false); + assert.equal( + emitted.filter((event) => event.type === "tool.activity" + && (event as { tool?: string; status?: string }).tool === "rectification-read-case" + && (event as { tool?: string; status?: string }).status === "started").length, + 1, + ); }); test("a failed opening does not let the next turn skip the server Skill load gate", async () => { diff --git a/frontend/tests/rectification-v9-stream.test.ts b/frontend/tests/rectification-v9-stream.test.ts index a0f9850d..bfd58688 100644 --- a/frontend/tests/rectification-v9-stream.test.ts +++ b/frontend/tests/rectification-v9-stream.test.ts @@ -1050,7 +1050,11 @@ test("does not retry set-focus with identical arguments", async () => { ...receiptHandlers, get_agentic_rectification_case_dossier: () => dossierFixture(), append_agentic_rectification_turn: () => ({ turn_id: TURN_ID }), - finalize_agentic_rectification_turn: () => ({ turn_id: TURN_ID, status: "failed", idempotent: false }), + finalize_agentic_rectification_turn: (_fn, args) => ({ + turn_id: TURN_ID, + status: args.p_status, + idempotent: false, + }), }); const { options, emitted, billing } = runOptions({ accounting: accounting.client, @@ -1071,10 +1075,66 @@ test("does not retry set-focus with identical arguments", async () => { const result = await runV9AgentTurn(options); - assert.equal(result.ok, false); - assert.equal(result.errorCode, "repeated_tool_call"); + assert.equal(result.ok, true); + assert.equal(result.errorCode, null); + assert.equal(result.answerText, "主问题：请确认这段经历发生在哪个月？"); assert.equal(emitted.some((event) => event.type === "attempt.reset"), false); - assert.deepEqual(billing, { reserved: 1, completed: 0, released: 1 }); + assert.equal(emitted.some((event) => event.type === "run.failed"), false); + assert.deepEqual(billing, { reserved: 1, completed: 1, released: 0 }); +}); + +test("duplicate compare after diagnostics still completes with server narration", async () => { + const executedMethods = [ + "ashtakavarga", + "d1-rashi", + "d10-dashamsa", + "shadbala", + "functional-benefic-malefic", + "arudha-pada", + ]; + const { options, emitted, billing } = runOptions({ + buildAgent: async () => fakeAgentStream([ + chunk("start"), + chunk("tool-call", { toolName: "skill", args: { name: RECTIFICATION_SKILL_NAME } }), + chunk("tool-result", { toolName: "skill" }), + chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }), + chunk("tool-result", { toolName: "rectification-read-case" }), + chunk("tool-call", { toolName: "rectification-record-evidence-batch", args: { caseId: CASE_ID } }), + chunk("tool-result", { toolName: "rectification-record-evidence-batch" }), + chunk("tool-call", { toolName: "rectification-propose-evidence", args: { caseId: CASE_ID } }), + chunk("tool-result", { toolName: "rectification-propose-evidence" }), + chunk("tool-call", { toolName: "rectification-confirm-evidence", args: { caseId: CASE_ID } }), + chunk("tool-result", { toolName: "rectification-confirm-evidence" }), + chunk("tool-call", { toolName: "rectification-read-diagnostics", args: { caseId: CASE_ID } }), + chunk("tool-result", { + toolName: "rectification-read-diagnostics", + result: { executed_methods: executedMethods }, + }), + chunk("tool-call", { toolName: "rectification-compare-candidates", args: { caseId: CASE_ID } }), + chunk("tool-result", { + toolName: "rectification-compare-candidates", + result: { executed_methods: executedMethods }, + }), + chunk("tool-call", { toolName: "rectification-compare-candidates", args: { caseId: CASE_ID } }), + chunk("finish"), + ]) as never, + }); + + const result = await runV9AgentTurn(options); + + assert.equal(result.ok, true); + assert.equal(result.errorCode, null); + assert.match(result.answerText, /已经记下|请继续说下一件/); + assert.equal(emitted.some((event) => event.type === "attempt.reset"), false); + assert.equal(emitted.some((event) => event.type === "run.failed"), false); + assert.equal(emitted.some((event) => event.type === "run.completed"), true); + assert.equal( + emitted.filter((event) => event.type === "tool.activity" + && (event as { tool?: string; status?: string }).tool === "rectification-compare-candidates" + && (event as { tool?: string; status?: string }).status === "started").length, + 1, + ); + assert.deepEqual(billing, { reserved: 1, completed: 1, released: 0 }); }); test("an unclaimed V10 attempt never starts the model", async () => {