Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(maker-core): 同批发出的等待轮询不再误判循环,受邀者会话用分享的模型复核

Dashc06695bc0966
diff --git a/apps/desktop/src/main/maker-host/__tests__/tool-loop-reviewer.test.ts b/apps/desktop/src/main/maker-host/__tests__/tool-loop-reviewer.test.ts
index e4c0e0dae2..ae19328766 100644
--- a/apps/desktop/src/main/maker-host/__tests__/tool-loop-reviewer.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/tool-loop-reviewer.test.ts
@@ -200,6 +200,25 @@ describe('tool loop reviewer', () => {
.validateResponse;
expect(validate('STOP')).toBe(true);
expect(validate('I think it is fine')).toBe(false);
+ const opts = (requestText.mock.calls[0] as unknown as [string, Record<string, unknown>])[1];
+ expect(opts).not.toHaveProperty('providerId');
+ expect(opts).not.toHaveProperty('model');
+ });
+
+ it('pins a routed review to the given provider and model', async () => {
+ const requestText = vi.fn(async () => okResult('CONTINUE'));
+ const reviewer = createToolLoopReviewer({ requestText });
+ const routed: ToolLoopReviewRequest = {
+ ...request,
+ agentKind: 'claude-code',
+ route: { providerId: 'shared-provider', model: 'shared-model' },
+ };
+ await expect(reviewer(routed, { signal: new AbortController().signal })).resolves.toBe('continue');
+ expect(requestText).toHaveBeenCalledWith(expect.any(String), expect.objectContaining({
+ providerId: 'shared-provider',
+ agentKind: 'claude-code',
+ model: 'shared-model',
+ }));
});
it('falls back to stop when no model answers usefully', async () => {
diff --git a/apps/desktop/src/main/maker-host/tool-loop-reviewer.ts b/apps/desktop/src/main/maker-host/tool-loop-reviewer.ts
index 76c7d17e45..43f19699d9 100644
--- a/apps/desktop/src/main/maker-host/tool-loop-reviewer.ts
+++ b/apps/desktop/src/main/maker-host/tool-loop-reviewer.ts
@@ -176,6 +176,10 @@ export function createToolLoopReviewer(deps: ToolLoopReviewerDeps): ToolLoopRevi
systemPrompt: TOOL_LOOP_REVIEW_SYSTEM_PROMPT,
responseInstructions: RESPONSE_INSTRUCTIONS,
validateResponse: (text) => parseToolLoopReviewDecision(text) !== null,
+ // 钉了路线(受邀者会话)就只走它,失败按 stop,不回落到本机辅助模型链。
+ ...(request.route
+ ? { providerId: request.route.providerId, agentKind: request.agentKind, model: request.route.model }
+ : {}),
});
if (!result.ok) return 'stop';
return parseToolLoopReviewDecision(result.text) ?? 'stop';
diff --git a/docs/dev-rules/maker-core-and-agent-behavior.md b/docs/dev-rules/maker-core-and-agent-behavior.md
index b2ebdf2f47..3e881d12a8 100644
--- a/docs/dev-rules/maker-core-and-agent-behavior.md
+++ b/docs/dev-rules/maker-core-and-agent-behavior.md
@@ -51,8 +51,10 @@ Claude Code 在原有 per-sidechain 回调里检测所有模型;Pi / Codex 在
这仍是有界启发式,不是任意长度循环的证明,也不以没有文件改动作为失败依据。
节奏与复核:一次普通调用若距上一次普通调用开始已有至少 30 秒(`sleep` 后再查 CI、
-命令内自带等待或慢推理),视为在等外部进度的节奏轮询,不计入上述重复与窗口判据,
-只续接完全相同结果的连续段。上述判据命中只算“疑似”(`final: false`):host 通过
+命令内自带等待或慢推理),或其结果距上一次普通结果到达已有至少 30 秒(模型一次发出
+多条等待命令、工具再逐条排队执行,发出时刻挨在一起但每条都真等了),视为在等外部进度
+的节奏轮询,不计入上述重复与窗口判据,只续接完全相同结果的连续段;同时发出又同时返回的
+快速重复照常计数。上述判据命中只算“疑似”(`final: false`):host 通过
`MakerDeps.toolLoopReviewer`(Pi / Codex)与 `AgentDeps.toolLoopReviewer`(Claude Code)
注入同一个辅助模型复核入口,由 `agents/shared/tool-loop-review.ts` 的 `ToolLoopMonitor`
在后台复核,Agent 不暂停。复核 continue → 接下来 20 次普通结果不再报疑似;stop、
@@ -71,7 +73,10 @@ Claude Code 在原有 per-sidechain 回调里检测所有模型;Pi / Codex 在
的厂商令牌与私钥、截断私钥块、URL 内嵌凭证、命令行凭证参数、凭证类 HTTP 头,最后以
含字母和数字的 32 位以上连续串兜底,宁可多删),再截断到每段 400 字符,最后改写分隔标签;
顺序不能颠倒,否则跨截断点的凭证会留下认不出的前缀。走共享辅助模型链,回答只接受
-CONTINUE / STOP。
+CONTINUE / STOP。供应商分享的受邀者会话(Claude Code 在分享者电脑上运行)不走分享者的
+辅助模型链:复核请求带 `route`,钉到分享的供应商与会话模型,失败按 stop,不回落;
+受邀者的复核材料只发给它本来就在用的模型。Pi / Codex 的检测在任务所在电脑的 Session 上,
+用的是受邀者自己的复核入口。
回归见 `loop-guard.test.ts`、`tool-loop-review.test.ts`、`session.tool-loop.test.ts`、
Claude Code 的 `upstream-idle-watchdog.test.ts` 与 desktop 的 `tool-loop-reviewer.test.ts`。
diff --git a/packages/maker-core/src/agents/claude-code/__tests__/upstream-idle-watchdog.test.ts b/packages/maker-core/src/agents/claude-code/__tests__/upstream-idle-watchdog.test.ts
index a15feb28af..0eb7fd7786 100644
--- a/packages/maker-core/src/agents/claude-code/__tests__/upstream-idle-watchdog.test.ts
+++ b/packages/maker-core/src/agents/claude-code/__tests__/upstream-idle-watchdog.test.ts
@@ -21,7 +21,7 @@ import path from 'node:path';
import { afterEach, describe, expect, it, vi } from 'vitest';
-import type { AgentDeps } from '../../base-agent.js';
+import type { AgentDeps, DeviceHostedSession } from '../../base-agent.js';
import type { AuthAdapter } from '../../../interfaces/auth-adapter.js';
import type { AgentEvent } from '../../../types/events.js';
import type { Logger } from '../../../interfaces/logger.js';
@@ -143,7 +143,11 @@ async function makeTempDir(): Promise<string> {
return dir;
}
-async function startSessionWithStream(model = 'claude-opus-4-6', toolLoopReviewer?: ToolLoopReviewer) {
+async function startSessionWithStream(
+ model = 'claude-opus-4-6',
+ toolLoopReviewer?: ToolLoopReviewer,
+ hosted?: { providerId: string; deviceHosted: (workingDir: string) => DeviceHostedSession },
+) {
const configDir = await makeTempDir();
process.env.CLAUDE_CONFIG_DIR = configDir;
const workingDir = await makeTempDir();
@@ -168,6 +172,7 @@ async function startSessionWithStream(model = 'claude-opus-4-6', toolLoopReviewe
model,
workingDir,
permissionMode: 'acceptEdits',
+ ...(hosted ? { providerId: hosted.providerId, deviceHosted: hosted.deviceHosted(workingDir) } : {}),
});
const events: AgentEvent[] = [];
@@ -515,6 +520,56 @@ describe('Claude Code tool-loop guard runtime integration', () => {
}
});
+ it('受邀者会话的复核钉到分享的供应商与会话模型', async () => {
+ const reviewer = vi.fn<ToolLoopReviewer>(async () => 'continue');
+ const { handle, stream, events, collected } = await startSessionWithStream('shared/model-a', reviewer, {
+ providerId: 'shared-provider',
+ deviceHosted: (workingDir) => ({
+ workingDir,
+ extraDirs: [],
+ writableDirs: [],
+ platform: 'darwin',
+ shell: 'zsh',
+ isGitRepo: false,
+ tunnelUrl: 'http://127.0.0.1:4000/t/tok/',
+ tunnelToken: 'tok',
+ mcpServers: [],
+ mirrorRoot: path.dirname(workingDir),
+ guest: true,
+ guestProvider: { providerId: 'shared-provider', modelIds: ['shared/model-a'], routeToken: 'route' },
+ }),
+ });
+ vi.useFakeTimers({ toFake: ['setTimeout', 'clearTimeout', 'Date'] });
+ try {
+ await handle.send({ type: 'user', content: 'wait for the run' });
+ for (let i = 0; i < 4; i++) {
+ const id = `guest-bash-${i}`;
+ stream.emit({
+ type: 'assistant', parent_tool_use_id: null,
+ message: { role: 'assistant', content: [{
+ type: 'tool_use', id, name: 'mcp__cindy_exec__Bash', input: { command: 'gh run view 1' },
+ }] },
+ });
+ stream.emit({
+ type: 'user', parent_tool_use_id: null,
+ message: { role: 'user', content: [{ type: 'tool_result', tool_use_id: id, content: 'in_progress' }] },
+ });
+ await pumpUntil(() => events.filter(e => e.type === 'tool_result_full').length === i + 1, 'bash delivered');
+ }
+ expect(reviewer).toHaveBeenCalledOnce();
+ expect(reviewer.mock.calls[0]?.[0]).toMatchObject({
+ agentKind: 'claude-code',
+ route: { providerId: 'shared-provider', model: 'shared/model-a' },
+ });
+ expect(toolLoopError(events)).toBeUndefined();
+ } finally {
+ vi.useRealTimers();
+ stream.end();
+ await handle.close().catch(() => undefined);
+ await collected;
+ }
+ });
+
it('复核放行时相同调用不中断;复核 stop 后才中断', async () => {
const decisions: Array<(decision: 'continue' | 'stop') => void> = [];
const reviewer = vi.fn<ToolLoopReviewer>(() => new Promise((resolve) => decisions.push(resolve)));
diff --git a/packages/maker-core/src/agents/claude-code/index.ts b/packages/maker-core/src/agents/claude-code/index.ts
index f921bd264d..9ac4f59545 100644
--- a/packages/maker-core/src/agents/claude-code/index.ts
+++ b/packages/maker-core/src/agents/claude-code/index.ts
@@ -3068,14 +3068,15 @@ export class ClaudeCodeAgent extends BaseAgent {
// keep exact repetition/rotation guards, not category-only retries.
contractConsecutiveLimit: Number.POSITIVE_INFINITY,
}), {
- // 供应商分享受邀者不复核:复核走本机用户自己的辅助模型,会把受邀者的工具内容发给分享之外的
- // 供应商。疑似循环按缺省直接中断。
- reviewer: hosted?.guest ? undefined : this.deps.toolLoopReviewer,
+ reviewer: this.deps.toolLoopReviewer,
reviewBudget: toolLoopReviewBudget,
context: () => ({
sessionId: sid,
agentKind: 'claude-code',
model: toolLoopGuardModelForScope(parentToolUseId),
+ // 供应商分享受邀者:复核钉到分享的供应商与会话模型。任务内容本来就发给它,
+ // 不能落到本机用户自己的辅助模型链(分享之外的供应商)上。
+ ...(guestProvider ? { route: { providerId: guestProvider.providerId, model: mutableModel } } : {}),
}),
// 复核结果晚到时,guard 可能已随新 turn 重置;只处理仍在位的同一 monitor。
onReviewedStop: (verdict) => {
diff --git a/packages/maker-core/src/agents/shared/loop-guard.test.ts b/packages/maker-core/src/agents/shared/loop-guard.test.ts
index f0785fdc04..38cfd2d5e1 100644
--- a/packages/maker-core/src/agents/shared/loop-guard.test.ts
+++ b/packages/maker-core/src/agents/shared/loop-guard.test.ts
@@ -599,6 +599,35 @@ describe('ToolLoopGuard pacing and final limits', () => {
}
});
+ it('treats waits issued together and run one by one as paced (2026-10 gate polling false positive)', () => {
+ const { guard, advance } = clockGuard();
+ const wait = (i: number) => ({ command: `sleep 58; grep GATE_EXIT gate.log # ${i % 2}` });
+ let id = 0;
+ // 模型一次发出 3-6 条等待命令,工具逐条排队执行:发出时刻相差 1 秒,结果每分钟到一个。
+ for (const batch of [6, 6, 3, 3, 6]) {
+ const ids: string[] = [];
+ for (let i = 0; i < batch; i++) {
+ advance(1_000);
+ ids.push(String(id));
+ guard.onToolUse(String(id), 'Bash', wait(id));
+ id += 1;
+ }
+ for (const toolUseId of ids) {
+ advance(minute);
+ expect(guard.onToolResult(toolUseId, `11:${toolUseId} waiting`).kind).toBe('ok');
+ }
+ }
+ });
+
+ it('still counts identical fast calls issued together and finished together', () => {
+ const { guard, advance } = clockGuard();
+ const verdicts: ToolLoopGuardVerdict[] = [];
+ for (let i = 0; i < 4; i++) guard.onToolUse(String(i), 'read', { path: 'same.ts' });
+ advance(minute);
+ for (let i = 0; i < 4; i++) verdicts.push(guard.onToolResult(String(i), 'unchanged'));
+ expect(verdicts[3]).toMatchObject({ kind: 'hard', reason: 'consecutive', count: 4, final: false });
+ });
+
it('stops identical paced polling after one hour without review', () => {
const { guard, advance } = clockGuard();
let verdict: ToolLoopGuardVerdict = { kind: 'ok' };
diff --git a/packages/maker-core/src/agents/shared/loop-guard.ts b/packages/maker-core/src/agents/shared/loop-guard.ts
index 9bf062517c..c59a531763 100644
--- a/packages/maker-core/src/agents/shared/loop-guard.ts
+++ b/packages/maker-core/src/agents/shared/loop-guard.ts
@@ -36,9 +36,11 @@ const STABLE_READ_DISTINCT_LIMIT = 32;
const STABLE_READ_MIN_REPEATS = 4;
const READ_ONLY_TOOL_NAMES = new Set(['Read', 'Grep', 'Glob', 'read', 'grep', 'find', 'ls']);
/**
- * 节奏判据:本次普通调用距上一次普通调用开始至少这么久,说明模型在等外部进度
- * (sleep 后再查 CI、命令内自带等待、慢推理),不是快速空转。这类结果不计入
- * 第 1-3 层的重复判据,只受下方两条最终上限约束。
+ * 节奏判据:本次普通调用距上一次普通调用开始至少这么久,或本次结果距上一次普通结果
+ * 到达至少这么久,说明模型在等外部进度(sleep 后再查 CI、命令内自带等待、慢推理),
+ * 不是快速空转。看结果间隔是因为模型可能一次发出多条相同的等待命令、再逐条排队执行:
+ * 发出时刻挨在一起,实际每条都等了很久。这类结果不计入第 1-3 层的重复判据,只受下方
+ * 两条最终上限约束。
*/
const DEFAULT_PACED_INTERVAL_MS = 30_000;
/** 最终上限:快速(非节奏)的完全相同调用连续达到此数,不经复核直接判定。 */
@@ -210,7 +212,7 @@ export interface ToolLoopGuardOptions {
rotationDistinctLimit?: number;
/** 第 4 层:同工具同类契约错误连续多少次判止损。 */
contractConsecutiveLimit?: number;
- /** 两次普通调用开始间隔达到此值即视为节奏轮询,不计入重复判据。 */
+ /** 两次普通调用开始间隔或结果到达间隔达到此值即视为节奏轮询,不计入重复判据。 */
pacedIntervalMs?: number;
/** 快速完全相同调用的最终上限(不经复核)。 */
finalConsecutiveLimit?: number;
@@ -254,8 +256,8 @@ export type ToolLoopGuardVerdict =
* 不按调用总数或任意长度的重复序列硬中断:仅凭工具 trace 无法区分合法批处理、
* 稳定状态轮询与死循环,有限窗口也只能移动误判/漏判边界。因此 1-4 层的命中
* 都是"疑似"(final=false),调用方可交给辅助模型复核;复核放行后由
- * acceptCurrentPattern() 暂缓疑似判定。距上一次普通调用 ≥ pacedIntervalMs
- * 才开始的调用视为节奏轮询,不计入重复判据。两条最终上限(快速完全相同
+ * acceptCurrentPattern() 暂缓疑似判定。距上一次普通调用开始或上一次普通结果
+ * 到达 ≥ pacedIntervalMs 的结果视为节奏轮询,不计入重复判据。两条最终上限(快速完全相同
* 连续 finalConsecutiveLimit 次、完全相同结果持续 finalStreakMs)与长只读
* 轮转不经复核直接判定(final=true)。
*
@@ -283,6 +285,7 @@ export class ToolLoopGuard {
// 节奏判据与复核材料
private lastOrdinaryStartedAt: number | null = null;
+ private lastOrdinaryFinishedAt: number | null = null;
private evidence: ToolLoopEvidence[] = [];
/** 复核放行后,接下来还有多少次普通结果不报疑似。 */
private suspicionGraceResults = 0;
@@ -393,9 +396,13 @@ export class ToolLoopGuard {
if (isPollingTool(toolUse.name, toolUse.input, isError)) return { kind: 'ok' };
this.lastResultObservedValue = true;
- const paced = this.lastOrdinaryStartedAt !== null
- && toolUse.startedAt - this.lastOrdinaryStartedAt >= this.pacedIntervalMs;
+ const finishedAt = this.now();
+ const paced = (this.lastOrdinaryStartedAt !== null
+ && toolUse.startedAt - this.lastOrdinaryStartedAt >= this.pacedIntervalMs)
+ || (this.lastOrdinaryFinishedAt !== null
+ && finishedAt - this.lastOrdinaryFinishedAt >= this.pacedIntervalMs);
this.lastOrdinaryStartedAt = Math.max(this.lastOrdinaryStartedAt ?? toolUse.startedAt, toolUse.startedAt);
+ this.lastOrdinaryFinishedAt = finishedAt;
this.recordEvidence(toolUse, output, isError);
const fullFingerprint = fingerprintToolCall(toolUse.name, toolUse.input, output);
const callFingerprint = fingerprintToolCall(toolUse.name, toolUse.input, null);
@@ -592,6 +599,7 @@ export class ToolLoopGuard {
this.streakResults = 0;
this.streakStartedAt = 0;
this.lastOrdinaryStartedAt = null;
+ this.lastOrdinaryFinishedAt = null;
this.evidence = [];
this.suspicionGraceResults = 0;
this.callWindow = [];
diff --git a/packages/maker-core/src/agents/shared/tool-loop-review.ts b/packages/maker-core/src/agents/shared/tool-loop-review.ts
index 5066869744..4c6ec7ce0f 100644
--- a/packages/maker-core/src/agents/shared/tool-loop-review.ts
+++ b/packages/maker-core/src/agents/shared/tool-loop-review.ts
@@ -4,10 +4,25 @@ import type { ToolLoopEvidence, ToolLoopGuard, ToolLoopGuardVerdict } from './lo
export type HardToolLoopVerdict = Extract<ToolLoopGuardVerdict, { kind: 'hard' }>;
-export interface ToolLoopReviewRequest {
+/**
+ * 复核必须走的供应商与模型。受邀者会话只能用分享给它的供应商:复核材料就是它的任务内容,
+ * 只能发给它本来就在用的那个模型,不能落到本机用户自己的辅助模型链上。
+ */
+export interface ToolLoopReviewRoute {
+ providerId: string;
+ model: string;
+}
+
+/** 复核请求里由调用方提供的会话信息。 */
+export interface ToolLoopReviewContext {
sessionId: string;
agentKind: AgentKind;
model: string;
+ /** 缺省 = 走 host 的共享辅助模型链。 */
+ route?: ToolLoopReviewRoute;
+}
+
+export interface ToolLoopReviewRequest extends ToolLoopReviewContext {
verdict: Pick<HardToolLoopVerdict, 'reason' | 'count' | 'toolName'>;
/**
* 最近的普通调用(旧→新)。均为未脱敏原文,input 保留原结构;发送前必须在未转义的
@@ -40,7 +55,7 @@ export interface ToolLoopMonitorOptions {
reviewer?: ToolLoopReviewer;
/** 缺省为本 monitor 独占;传入同一对象即共享每 turn 的复核上限。 */
reviewBudget?: ToolLoopReviewBudget;
- context: () => { sessionId: string; agentKind: AgentKind; model: string };
+ context: () => ToolLoopReviewContext;
/** 复核判定 stop(含失败/超时)时调用;调用方自行确认 turn 仍然有效再中断。 */
onReviewedStop: (verdict: HardToolLoopVerdict) => void;
logger: Logger;