minimal-agent 0.11.0 → 0.12.0

package/src/cli/streamJson.ts

@@ -33,7 +33,9 @@ export type { AskAnswer } from '../askBridge.js';
 import { normalizeImageDetail } from '../multimodal.js';
 import type { ImageInputSpec, Message, Provider } from '../types.js';
 import { runWithPlugins } from '../plugins/pluginRunner.js';
-import { buildTransientPromptBundle } from '../prompts/system.js';
+import { buildBaseSystemPrompt, buildTransientPromptBundle } from '../prompts/system.js';
+import { getOrCreateSessionState, sessionStateKey } from '../session/sessionState.js';
+import { hydrateSessionWorktree } from '../tools/worktree/sessionWorktree.js';
 import { ALL_TOOLS } from '../tools/index.js';
 import { createPromptCacheContext } from '../llm/promptCache.js';
 import {
@@ -344,10 +346,16 @@ export async function runStreamJsonMode(
   });
 
   try {
-    const cwd = options.workspace?.cwd ?? getWorkingDir();
-    const workspace = await workspaceIndexService.findByPath(cwd).catch(() => null);
+    const workspaceRoot = options.workspace?.cwd ?? getWorkingDir();
+    const workspace = await workspaceIndexService.findByPath(workspaceRoot).catch(() => null);
     const sessionId = options.workspace?.sessionId ?? 'default';
-    const conversationKey = `workspace:${workspace?.id ?? cwd}:session:${sessionId}`;
+    const sessionState = getOrCreateSessionState(sessionStateKey({
+      surface: 'headless', workspaceId: workspace?.id,
+      cwd: workspace?.id ? undefined : workspaceRoot, sessionId,
+    }));
+    await hydrateSessionWorktree({ sessionState, workspaceId: workspace?.id, sessionId, cwd: workspaceRoot });
+    const cwd = sessionState.activeWorktree?.path ?? workspaceRoot;
+    const conversationKey = `workspace:${workspace?.id ?? workspaceRoot}:session:${sessionId}`;
     const computerRun = options.computerUseMode && options.computerUseMode !== 'disabled'
       ? createCliComputerRun({
           mode: options.computerUseMode,
@@ -371,6 +379,9 @@ export async function runStreamJsonMode(
       computer: computerRun,
     });
     const turnTools = filterToolsForRuntime(ALL_TOOLS, runtimeContext);
+    const stablePrompt = await buildBaseSystemPrompt(cwd, turnTools);
+    if (history[0]?.role === 'system') history[0] = { ...history[0], content: stablePrompt };
+    else history.unshift({ role: 'system', content: stablePrompt });
     const transientBundle = await buildTransientPromptBundle({
       cwd,
       tools: turnTools,
@@ -378,6 +389,7 @@ export async function runStreamJsonMode(
         surface: 'headless',
         cwd,
         workspaceId: workspace?.id,
+        sessionId,
         interactiveAsk: control,
         memoryAccess: runtimeContext.memoryAccess,
         currentUserInput: prompt,
@@ -393,6 +405,7 @@ export async function runStreamJsonMode(
       imageDetail: cliImageOptions.imageDetail,
       imageMaxLongSidePixel: cliImageOptions.imageMaxLongSidePixel,
       runtimeContext,
+      sessionState,
       allowedTools: turnTools.map((tool) => tool.name),
       ...(transientBundle ? { transientBundle } : {}),
       promptCache: createPromptCacheContext(

package/src/config/configDefaults.js

@@ -0,0 +1,25 @@
+/**
+ * 浏览器与 Node/Bun 都可安全加载的纯配置默认值。
+ * 不得在此模块引入 node:*、文件系统或运行时路径逻辑。
+ */
+export const DEFAULT_SUBAGENTS_CONFIG = {
+    enabled: true,
+    maxCallsPerTurn: 6,
+    maxConcurrent: 3,
+    defaultWriteIsolation: 'none',
+    timeoutMs: 1_800_000,
+    exploreMaxTurns: 50,
+    generalMaxTurns: 100,
+};
+export const DEFAULT_BROWSER_CONFIG = {
+    enabled: true,
+    profileId: 'default',
+    developerMode: false,
+    allowedCdpOrigins: [],
+};
+export const DEFAULT_COMPUTER_USE_CONFIG = {
+    enabled: true,
+    imagePolicy: 'single-adaptive',
+    maxImagePixels: 1_600_000,
+    observationFrames: 1,
+};

package/src/config/configDefaults.ts

@@ -0,0 +1,33 @@
+import type {
+  BrowserConfig,
+  ComputerUseConfig,
+  SubagentsConfig,
+} from './configFile.js';
+
+/**
+ * 浏览器与 Node/Bun 都可安全加载的纯配置默认值。
+ * 不得在此模块引入 node:*、文件系统或运行时路径逻辑。
+ */
+export const DEFAULT_SUBAGENTS_CONFIG: SubagentsConfig = {
+  enabled: true,
+  maxCallsPerTurn: 6,
+  maxConcurrent: 3,
+  defaultWriteIsolation: 'none',
+  timeoutMs: 1_800_000,
+  exploreMaxTurns: 50,
+  generalMaxTurns: 100,
+};
+
+export const DEFAULT_BROWSER_CONFIG: BrowserConfig = {
+  enabled: true,
+  profileId: 'default',
+  developerMode: false,
+  allowedCdpOrigins: [],
+};
+
+export const DEFAULT_COMPUTER_USE_CONFIG: ComputerUseConfig = {
+  enabled: true,
+  imagePolicy: 'single-adaptive',
+  maxImagePixels: 1_600_000,
+  observationFrames: 1,
+};

package/src/config/configFile.js

@@ -5,27 +5,8 @@ import { dirname, join } from 'node:path';
 import { execSiblingDir } from '../utils/greenRoot.js';
 import { normalizePromptCacheMode } from '../llm/promptCache.js';
 import { CONFIG_LIMITS, normalizeLimited } from './numericLimits.js';
-export const DEFAULT_SUBAGENTS_CONFIG = {
-    enabled: true,
-    maxCallsPerTurn: 6,
-    maxConcurrent: 3,
-    defaultWriteIsolation: 'none',
-    timeoutMs: 1_800_000,
-    exploreMaxTurns: 50,
-    generalMaxTurns: 100,
-};
-export const DEFAULT_BROWSER_CONFIG = {
-    enabled: true,
-    profileId: 'default',
-    developerMode: false,
-    allowedCdpOrigins: [],
-};
-export const DEFAULT_COMPUTER_USE_CONFIG = {
-    enabled: true,
-    imagePolicy: 'single-adaptive',
-    maxImagePixels: 1_600_000,
-    observationFrames: 1,
-};
+import { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, DEFAULT_SUBAGENTS_CONFIG, } from './configDefaults.js';
+export { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, DEFAULT_SUBAGENTS_CONFIG, } from './configDefaults.js';
 const DEFAULT_CONTEXT_WINDOW = 128_000;
 const LOCK_STALE_MS = 120_000;
 const LOCK_WAIT_MS = 15_000;

package/src/config/configFile.ts

@@ -7,6 +7,17 @@ import { execSiblingDir } from '../utils/greenRoot.js';
 import type { PromptCacheConfig } from '../types.js';
 import { normalizePromptCacheMode } from '../llm/promptCache.js';
 import { CONFIG_LIMITS, normalizeLimited } from './numericLimits.js';
+import {
+  DEFAULT_BROWSER_CONFIG,
+  DEFAULT_COMPUTER_USE_CONFIG,
+  DEFAULT_SUBAGENTS_CONFIG,
+} from './configDefaults.js';
+
+export {
+  DEFAULT_BROWSER_CONFIG,
+  DEFAULT_COMPUTER_USE_CONFIG,
+  DEFAULT_SUBAGENTS_CONFIG,
+} from './configDefaults.js';
 
 export type ImageProtocol = 'auto' | 'chat' | 'images';
 export type ApiFormat = 'openai' | 'anthropic';
@@ -95,30 +106,6 @@ export interface MinimalAgentConfig {
   savedAt: number;
 }
 
-export const DEFAULT_SUBAGENTS_CONFIG: SubagentsConfig = {
-  enabled: true,
-  maxCallsPerTurn: 6,
-  maxConcurrent: 3,
-  defaultWriteIsolation: 'none',
-  timeoutMs: 1_800_000,
-  exploreMaxTurns: 50,
-  generalMaxTurns: 100,
-};
-
-export const DEFAULT_BROWSER_CONFIG: BrowserConfig = {
-  enabled: true,
-  profileId: 'default',
-  developerMode: false,
-  allowedCdpOrigins: [],
-};
-
-export const DEFAULT_COMPUTER_USE_CONFIG: ComputerUseConfig = {
-  enabled: true,
-  imagePolicy: 'single-adaptive',
-  maxImagePixels: 1_600_000,
-  observationFrames: 1,
-};
-
 /**
  * 加模型是提交式弹窗、顾问改字段都是有界枚举/数字——不会出现「输入中途非法值」，
  * 所以 v3 时代的 provider/advisor 草稿桶随 schema 一起删除，只剩 tools（imageGeneration/

package/src/config/configManagement.js

@@ -1,4 +1,4 @@
-import { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, } from './configFile.js';
+import { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, } from './configDefaults.js';
 /**
  * CLI 和 Desktop 共用的脱敏配置视图。这里不返回任何凭据，也不读取 Runtime
  * 状态；它只是把 V4 权威配置稳定地投影成可展示、可比较的管理模型。

package/src/config/configManagement.ts

@@ -1,8 +1,8 @@
 import {
   DEFAULT_BROWSER_CONFIG,
   DEFAULT_COMPUTER_USE_CONFIG,
-  type ModelConfig,
-} from './configFile.js';
+} from './configDefaults.js';
+import type { ModelConfig } from './configFile.js';
 import type { ConfigSnapshot } from './configService.js';
 
 export type ManagementStatus = 'unconfigured' | 'configured' | 'incomplete';

package/src/context/compact.js

@@ -379,6 +379,10 @@ async function runCompaction(messages, provider, signal, chatFn = chat, promptCa
             ...(systemMsg ? [systemMsg] : []),
             {
                 role: 'user',
+                // ⚠️ 这里**故意不带** provider 原始报错：本条消息会进 history，既落盘也会在
+                //   后续轮次重新发给模型，错误体里可能夹带 URL / 请求体 / 凭据。
+                //   排障线索走 `compact_failed` 事件（只到 stderr / UI，不进 history）。
+                //   `compact.test.ts::reports snip fallback without exposing the raw LLM error` 锁定此约定。
                 content: '(LLM 摘要压缩失败，已使用规则式截断继续会话)'
                     + (failureLedger
                         ? `\n\n<user_messages>\n以下是被截掉的那部分对话里**用户的原话**（程序逐字抽取）。摘要不存在，这里就是唯一残留的历史要求。\n${failureLedger}\n</user_messages>`

package/src/context/compact.ts

@@ -491,6 +491,10 @@ async function runCompaction(
       ...(systemMsg ? [systemMsg] : []),
       {
         role: 'user',
+        // ⚠️ 这里**故意不带** provider 原始报错：本条消息会进 history，既落盘也会在
+        //   后续轮次重新发给模型，错误体里可能夹带 URL / 请求体 / 凭据。
+        //   排障线索走 `compact_failed` 事件（只到 stderr / UI，不进 history）。
+        //   `compact.test.ts::reports snip fallback without exposing the raw LLM error` 锁定此约定。
         content: '(LLM 摘要压缩失败，已使用规则式截断继续会话)'
           + (failureLedger
             ? `\n\n<user_messages>\n以下是被截掉的那部分对话里**用户的原话**（程序逐字抽取）。摘要不存在，这里就是唯一残留的历史要求。\n${failureLedger}\n</user_messages>`

package/src/context/contextBundle.js

@@ -8,10 +8,24 @@ export function contextFragment(input) {
         estimatedTokens: input.estimatedTokens ?? countTextTokens(content),
     };
 }
-function clipToTokenBudget(content, budgetTokens) {
+const CLIP_MARKER = '\n[上下文片段因预算截断]';
+/**
+ * 按**渲染后**的 token 数收敛，而不是裸内容。
+ *
+ * ⚠️ 曾经的实现拿裸内容 token 判预算、外层却用 `renderedTokens()`（JSON 转义 + 信封）验收，
+ * 于是 `裸 token ≤ 预算 < 渲染 token` 这个窗口里 clip 会「认为放得下」原样返回，验收再不通过，
+ * 片段就被整段丢进 `omittedIds` —— `clippable` 形同虚设（`clippedIds` 恒空）。
+ * 窗口宽度 = 转义膨胀量，所以 **Windows 上尤其致命**：技能目录里每个 `\` 都要转义成 `\\`，
+ * 27 条技能的窗口从 POSIX 的 ~25 token 涨到 ~92 token，本仓库实测整份 `skills.catalog`
+ * （2643 token）在还剩 1900+ token 预算时被整个扔掉，模型完全看不到任何 Skill。
+ *
+ * 判据必须与验收同口径：这里直接对 `measure(候选内容)` 二分，measure 由调用方传入，
+ * 保证「算的就是最终要花的」。
+ */
+function clipToRenderedBudget(content, budgetTokens, measure) {
     if (budgetTokens <= 0)
         return '';
-    if (countTextTokens(content) <= budgetTokens)
+    if (measure(content) <= budgetTokens)
         return content;
     // Token counter is deliberately dependency-free. Binary search preserves its
     // exact mixed Chinese/ASCII estimate instead of falling back to chars/4.
@@ -19,12 +33,15 @@ function clipToTokenBudget(content, budgetTokens) {
     let high = content.length;
     while (low < high) {
         const mid = Math.ceil((low + high) / 2);
-        if (countTextTokens(content.slice(0, mid)) <= Math.max(1, budgetTokens - 12))
+        const candidate = `${content.slice(0, mid).trimEnd()}${CLIP_MARKER}`;
+        if (measure(candidate) <= budgetTokens)
             low = mid;
         else
             high = mid - 1;
     }
-    return `${content.slice(0, low).trimEnd()}\n[上下文片段因预算截断]`;
+    if (low <= 0)
+        return '';
+    return `${content.slice(0, low).trimEnd()}${CLIP_MARKER}`;
 }
 export function renderContextFragment(fragment) {
     const encoded = JSON.stringify(fragment.content)
@@ -36,6 +53,9 @@ export function renderContextFragment(fragment) {
 function renderedTokens(fragment) {
     return countTextTokens(renderContextFragment(fragment));
 }
+/** 片段之间的分隔符——记账时必须与最终 join 用的是同一个常量。 */
+const FRAGMENT_SEPARATOR = '\n\n';
+const SEPARATOR_TOKENS = countTextTokens(FRAGMENT_SEPARATOR);
 /**
  * Deterministic context budgeting. Required host facts are kept first; optional
  * fragments compete by priority, then original order. No LLM call is involved.
@@ -52,21 +72,26 @@ export function assembleContextBundle(input, budgetTokens = DEFAULT_TRANSIENT_BU
     let remaining = Math.max(0, budgetTokens);
     for (const item of ordered) {
         const fragment = item.fragment;
+        // 片段之间要用 SEPARATOR 拼接，那几个 token 也得记账，否则最终 text 会超出预算
+        // （usedTokens > budgetTokens 会让 loop.ts::transientFor 的余量判断失真）。
+        const separator = included.length > 0 ? SEPARATOR_TOKENS : 0;
+        const budgetForItem = remaining - separator;
         const fullTokens = renderedTokens(fragment);
-        if (fullTokens <= remaining) {
+        if (fullTokens <= budgetForItem) {
             included.push(item);
-            remaining -= fullTokens;
+            remaining -= fullTokens + separator;
             continue;
         }
-        if ((fragment.required || fragment.clippable) && remaining > 24) {
-            const wrapperTokens = renderedTokens({ ...fragment, content: '' });
-            const content = clipToTokenBudget(fragment.content, Math.max(0, remaining - wrapperTokens));
+        if ((fragment.required || fragment.clippable) && budgetForItem > 24) {
+            // 同口径：二分时就按「这段内容渲染进本片段信封后要花多少 token」来量。
+            const content = clipToRenderedBudget(fragment.content, budgetForItem, (candidate) => renderedTokens({ ...fragment, content: candidate }));
             const clipped = contextFragment({ ...fragment, content });
             const clippedTokens = renderedTokens(clipped);
-            if (clipped.content && clippedTokens <= remaining) {
+            if (clipped.content && clippedTokens <= budgetForItem) {
                 included.push({ fragment: clipped, index: item.index });
-                clippedIds.push(fragment.id);
-                remaining -= clippedTokens;
+                if (clipped.content !== fragment.content)
+                    clippedIds.push(fragment.id);
+                remaining -= clippedTokens + separator;
                 continue;
             }
         }
@@ -75,7 +100,7 @@ export function assembleContextBundle(input, budgetTokens = DEFAULT_TRANSIENT_BU
     // Rendering follows semantic input order, not selection priority.
     included.sort((a, b) => a.index - b.index);
     const selected = included.map((item) => item.fragment);
-    const text = selected.map(renderContextFragment).join('\n\n');
+    const text = selected.map(renderContextFragment).join(FRAGMENT_SEPARATOR);
     const usedTokens = countTextTokens(text);
     return {
         fragments: selected,

package/src/context/contextBundle.ts

@@ -8,6 +8,7 @@ export type ContextFragmentKind =
   | 'project'
   | 'skill'
   | 'plugin'
+  | 'task'
   | 'diagnostic';
 
 export type ContextAuthority = 'host' | 'constraint' | 'user-evidence' | 'untrusted-data';
@@ -50,19 +51,40 @@ export function contextFragment(
   };
 }
 
-function clipToTokenBudget(content: string, budgetTokens: number): string {
+const CLIP_MARKER = '\n[上下文片段因预算截断]';
+
+/**
+ * 按**渲染后**的 token 数收敛，而不是裸内容。
+ *
+ * ⚠️ 曾经的实现拿裸内容 token 判预算、外层却用 `renderedTokens()`（JSON 转义 + 信封）验收，
+ * 于是 `裸 token ≤ 预算 < 渲染 token` 这个窗口里 clip 会「认为放得下」原样返回，验收再不通过，
+ * 片段就被整段丢进 `omittedIds` —— `clippable` 形同虚设（`clippedIds` 恒空）。
+ * 窗口宽度 = 转义膨胀量，所以 **Windows 上尤其致命**：技能目录里每个 `\` 都要转义成 `\\`，
+ * 27 条技能的窗口从 POSIX 的 ~25 token 涨到 ~92 token，本仓库实测整份 `skills.catalog`
+ * （2643 token）在还剩 1900+ token 预算时被整个扔掉，模型完全看不到任何 Skill。
+ *
+ * 判据必须与验收同口径：这里直接对 `measure(候选内容)` 二分，measure 由调用方传入，
+ * 保证「算的就是最终要花的」。
+ */
+function clipToRenderedBudget(
+  content: string,
+  budgetTokens: number,
+  measure: (candidate: string) => number,
+): string {
   if (budgetTokens <= 0) return '';
-  if (countTextTokens(content) <= budgetTokens) return content;
+  if (measure(content) <= budgetTokens) return content;
   // Token counter is deliberately dependency-free. Binary search preserves its
   // exact mixed Chinese/ASCII estimate instead of falling back to chars/4.
   let low = 0;
   let high = content.length;
   while (low < high) {
     const mid = Math.ceil((low + high) / 2);
-    if (countTextTokens(content.slice(0, mid)) <= Math.max(1, budgetTokens - 12)) low = mid;
+    const candidate = `${content.slice(0, mid).trimEnd()}${CLIP_MARKER}`;
+    if (measure(candidate) <= budgetTokens) low = mid;
     else high = mid - 1;
   }
-  return `${content.slice(0, low).trimEnd()}\n[上下文片段因预算截断]`;
+  if (low <= 0) return '';
+  return `${content.slice(0, low).trimEnd()}${CLIP_MARKER}`;
 }
 
 export function renderContextFragment(fragment: ContextFragment): string {
@@ -77,6 +99,10 @@ function renderedTokens(fragment: ContextFragment): number {
   return countTextTokens(renderContextFragment(fragment));
 }
 
+/** 片段之间的分隔符——记账时必须与最终 join 用的是同一个常量。 */
+const FRAGMENT_SEPARATOR = '\n\n';
+const SEPARATOR_TOKENS = countTextTokens(FRAGMENT_SEPARATOR);
+
 /**
  * Deterministic context budgeting. Required host facts are kept first; optional
  * fragments compete by priority, then original order. No LLM call is involved.
@@ -98,21 +124,29 @@ export function assembleContextBundle(
 
   for (const item of ordered) {
     const fragment = item.fragment;
+    // 片段之间要用 SEPARATOR 拼接，那几个 token 也得记账，否则最终 text 会超出预算
+    // （usedTokens > budgetTokens 会让 loop.ts::transientFor 的余量判断失真）。
+    const separator = included.length > 0 ? SEPARATOR_TOKENS : 0;
+    const budgetForItem = remaining - separator;
     const fullTokens = renderedTokens(fragment);
-    if (fullTokens <= remaining) {
+    if (fullTokens <= budgetForItem) {
       included.push(item);
-      remaining -= fullTokens;
+      remaining -= fullTokens + separator;
       continue;
     }
-    if ((fragment.required || fragment.clippable) && remaining > 24) {
-      const wrapperTokens = renderedTokens({ ...fragment, content: '' });
-      const content = clipToTokenBudget(fragment.content, Math.max(0, remaining - wrapperTokens));
+    if ((fragment.required || fragment.clippable) && budgetForItem > 24) {
+      // 同口径：二分时就按「这段内容渲染进本片段信封后要花多少 token」来量。
+      const content = clipToRenderedBudget(
+        fragment.content,
+        budgetForItem,
+        (candidate) => renderedTokens({ ...fragment, content: candidate }),
+      );
       const clipped = contextFragment({ ...fragment, content });
       const clippedTokens = renderedTokens(clipped);
-      if (clipped.content && clippedTokens <= remaining) {
+      if (clipped.content && clippedTokens <= budgetForItem) {
         included.push({ fragment: clipped, index: item.index });
-        clippedIds.push(fragment.id);
-        remaining -= clippedTokens;
+        if (clipped.content !== fragment.content) clippedIds.push(fragment.id);
+        remaining -= clippedTokens + separator;
         continue;
       }
     }
@@ -122,7 +156,7 @@ export function assembleContextBundle(
   // Rendering follows semantic input order, not selection priority.
   included.sort((a, b) => a.index - b.index);
   const selected = included.map((item) => item.fragment);
-  const text = selected.map(renderContextFragment).join('\n\n');
+  const text = selected.map(renderContextFragment).join(FRAGMENT_SEPARATOR);
   const usedTokens = countTextTokens(text);
   return {
     fragments: selected,

package/src/harness/stallGuard.js

@@ -5,33 +5,48 @@ function hash(value) {
 /** Run-local deterministic guard. It never persists arguments or result text. */
 export class StallGuard {
     #lastFailure = '';
+    #lastFailureCall = '';
     #failureCount = 0;
     #lastBatch = '';
     #batchCount = 0;
+    shouldReject(call) {
+        return (this.#failureCount >= 2 &&
+            this.#lastFailureCall ===
+                hash(`${call.function.name}\0${call.function.arguments}`));
+    }
     observe(calls, results, assistantText) {
         for (let index = 0; index < calls.length; index += 1) {
             const call = calls[index];
             const result = results[index];
             if (result.ok) {
                 this.#lastFailure = '';
+                this.#lastFailureCall = '';
                 this.#failureCount = 0;
                 continue;
             }
-            const signature = hash(`${call.function.name}\0${call.function.arguments}\0${result.error}`);
+            const callSignature = hash(`${call.function.name}\0${call.function.arguments}`);
+            const signature = hash(`${callSignature}\0${result.error}`);
             if (signature === this.#lastFailure)
                 this.#failureCount += 1;
             else {
                 this.#lastFailure = signature;
+                this.#lastFailureCall = callSignature;
                 this.#failureCount = 1;
             }
             if (this.#failureCount >= 3) {
-                return { stalled: true, reason: 'repeated_failure', signature: signature.slice(0, 16) };
+                return {
+                    stalled: true,
+                    reason: 'repeated_failure',
+                    signature: signature.slice(0, 16),
+                };
             }
         }
-        const batch = hash(calls.map((call, index) => {
+        const batch = hash(calls
+            .map((call, index) => {
             const result = results[index];
             return `${call.function.name}\0${call.function.arguments}\0${result.ok ? result.content : result.error}`;
-        }).join('\x01'));
+        })
+            .join('\x01'));
         if (assistantText.trim().length === 0 && batch === this.#lastBatch)
             this.#batchCount += 1;
         else {

package/src/harness/stallGuard.ts

@@ -14,10 +14,19 @@ export interface StallObservation {
 /** Run-local deterministic guard. It never persists arguments or result text. */
 export class StallGuard {
   #lastFailure = '';
+  #lastFailureCall = '';
   #failureCount = 0;
   #lastBatch = '';
   #batchCount = 0;
 
+  shouldReject(call: ToolCall): boolean {
+    return (
+      this.#failureCount >= 2 &&
+      this.#lastFailureCall ===
+        hash(`${call.function.name}\0${call.function.arguments}`)
+    );
+  }
+
   observe(
     calls: readonly ToolCall[],
     results: readonly ToolResult[],
@@ -28,25 +37,39 @@ export class StallGuard {
       const result = results[index]!;
       if (result.ok) {
         this.#lastFailure = '';
+        this.#lastFailureCall = '';
         this.#failureCount = 0;
         continue;
       }
-      const signature = hash(`${call.function.name}\0${call.function.arguments}\0${result.error}`);
+      const callSignature = hash(
+        `${call.function.name}\0${call.function.arguments}`,
+      );
+      const signature = hash(`${callSignature}\0${result.error}`);
       if (signature === this.#lastFailure) this.#failureCount += 1;
       else {
         this.#lastFailure = signature;
+        this.#lastFailureCall = callSignature;
         this.#failureCount = 1;
       }
       if (this.#failureCount >= 3) {
-        return { stalled: true, reason: 'repeated_failure', signature: signature.slice(0, 16) };
+        return {
+          stalled: true,
+          reason: 'repeated_failure',
+          signature: signature.slice(0, 16),
+        };
       }
     }
 
-    const batch = hash(calls.map((call, index) => {
-      const result = results[index]!;
-      return `${call.function.name}\0${call.function.arguments}\0${result.ok ? result.content : result.error}`;
-    }).join('\x01'));
-    if (assistantText.trim().length === 0 && batch === this.#lastBatch) this.#batchCount += 1;
+    const batch = hash(
+      calls
+        .map((call, index) => {
+          const result = results[index]!;
+          return `${call.function.name}\0${call.function.arguments}\0${result.ok ? result.content : result.error}`;
+        })
+        .join('\x01'),
+    );
+    if (assistantText.trim().length === 0 && batch === this.#lastBatch)
+      this.#batchCount += 1;
     else {
       this.#lastBatch = batch;
       this.#batchCount = assistantText.trim().length === 0 ? 1 : 0;