Agent 对话可配置最大工具调用次数 + Token 优化 + 循环检测升级

[cyfung1031]

Summary

Implements the proposal for #1545 (Agent 内置 AI Agent 高频工具调用的痛点：迭代次数硬限制、Token 消耗过高、缺少死循环检测)。

• 可配置最大工具调用次数：新增 AgentConfigRepo（chatMaxIterations，默认 50，可调 1–1000），Settings 页新增"对话"分类；达到上限时消息持久化 errorCode，聊天界面渲染"继续对话"按钮一键复用已持久化的完整历史继续执行（不再需要新建对话、不会丢失已探索的上下文）。
• Token 优化：
  • Anthropic provider 为最后一条消息追加 cache_control 断点，使长 tool loop 中已产生的历史前缀被缓存，后续每轮仅新增部分计费（现有 system/tools 断点 + 此断点共 3 个，未超过 4 个上限）。
  • execute_script 返回值超过 30000 字符时截断为首尾各 15000 字符并标注 truncated/original_length，避免超大返回值（如 DOM dump）被完整重复发送。
  • 新增滑动窗口裁剪：上下文占用达到 40%/60% 两个阈值时各触发一次，保留最近 5 轮 assistant/tool 消息原文，更早轮次的 tool 结果替换为占位文本；仅裁剪内存中传给 LLM 的消息，不影响持久化历史与 UI 展示；按阈值分批触发而非逐轮触发，避免频繁重写消息前缀导致 prompt cache 断点失效。
• 循环检测升级：现有 tool_call_guard 检测到重复/循环模式时只向 LLM 注入提醒；新增 askUserForGuard 回调，连续命中达到 2 次时暂停循环、询问用户"继续"或"停止"，停止则优雅收尾（done 而非 error）。仅 UI 对话（含后台会话，复用既有 ask_user 机制，5 分钟无人应答默认继续）传入该回调；定时任务与子代理保持原有的仅告警不暂停行为，避免无人值守场景被阻塞。

调查过程与技术方案细节见随本次改动附带的 issue-1545-proposal.md。

Test plan

• TDD：所有改动均先写失败测试（BDD 风格中文 describe/it），再实现
• npm run test:ci — 全量 3033 个测试通过（289 个测试文件）
• npm run lint — prettier / tsc / eslint 全部通过
• 全部 8 个 locale（zh-CN、en-US、zh-TW、ja-JP、ru-RU、de-DE、vi-VN、tr-TR）补齐新增文案

🤖 Generated with Claude Code

[cyfung1031]

Thanks for the review — all 4 issues confirmed and fixed in de83fd3, each with a regression test written to fail first, then pass after the fix.

1. Blocker — empty error message replayed to LLM on Continue. Confirmed: buildAndPersistUserMessage copied every non-system persisted message verbatim, including the content: "" / error / errorCode: "max_iterations" message. Fixed by skipping any historical message with error set when rebuilding the LLM-facing context (it stays in the repo/UI for display, just isn't replayed). Regression test seeds history with a max-iterations error message, sends "continue", and asserts the outgoing request contains no empty-content assistant message.

2. Stale pendingAskUser after guard-timeout. Confirmed — the 5-minute timeout in askUserForGuard deleted the resolver but never cleared rc.pendingAskUser, so a later attach would sync a dead prompt. Fixed by clearing rc.pendingAskUser in the timeout callback when it still matches the timed-out askId. The regression test needed care: naturally letting the conversation finish masked the bug because the existing done handler also clears pendingAskUser, so the test now holds the 5th round's fetch pending forever and asserts pendingAskUser is already cleared while status is still "running" — i.e. checks the fix fires at timeout, not incidentally at completion.

3. Bounds enforced only in Settings UI. Confirmed — AgentConfigRepo stored/returned raw values, and chat_service.ts used params.maxIterations || agentConfig.chatMaxIterations (which also let a negative params.maxIterations slip through unclamped, since only 0/NaN/"" are falsy). Added normalizeChatMaxIterations (clamp to [1, 1000], round, fall back to default on non-finite/non-number) as the single source of truth, applied in both getConfig/saveConfig and exported for chat_service.ts to apply to the final resolved value (also switched to ?? so an explicit 0 clamps to 1 instead of silently falling back to the configured default). Tests cover corrupted storage (0, negative, oversized, NaN, string, null, non-integer) and a directly-supplied negative maxIterations.

4. Strike count never resets. Confirmed as described — agreed it was more aggressive than intended. Now resets to 0 after a "Continue" answer, so escalation requires 2 fresh hits again rather than pausing on every subsequent warning for the rest of the run. Test drives 8 rounds of repeated calls (4 hits total) and asserts the pause only fires on hits 2 and 4, not on hit 3.

Full verification after the fixes: npm run test:ci — 3041/3041 tests pass across 289 files (up from 3033, the 8 new regression tests); npm run lint (prettier + tsc + eslint) clean.