thincoder 0.12.38 → 0.12.40

package/CHANGELOG.md

@@ -0,0 +1,311 @@
+# Changelog
+
+本文件记录 ThinCoder CLI 的发布历史。格式基于 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.0.0/)，版本遵循[语义化版本](https://semver.org/lang/zh-CN/)。
+
+## [0.12.40] — 2026-08-23
+
+### Fixed
+
+- **`shrinkOversized` 数据丢失**：原地改共享消息对象会污染持久化的人读线（巨型粘贴内容被永久截断）——改为复制-on-write（机器线替换、人读线不动）
+- 轮末探索蒸馏边界只在压缩**真重建**时重置（shrink 路径不再误重置）
+- `createAgent` 补 `_emptyRetries` 初始化；删 advisor 分支无用 JSON.parse
+
+## [0.12.39] — 2026-08-23
+
+### Added
+
+- **主 agent 委托策略**：`main.md` 把「广度探索 → explore 子代理（隔离上下文，逐步读/搜不进主历史）」「仅当即将立刻编辑时才自己 read」「coder 验证 = 读改动文件 + 跑测试」从建议升级为明确规则
+- **历史卫生**：轮末 `summarizeRunExplorations` 把一轮内连续探索结果（read/grep/glob/ls/code_search/doc_search/repo_outline）LLM 蒸馏为 `[Exploration summary]`（机器线收缩、人读线全量不变）；压缩 `SUMMARIZE_PROMPT` 加「已改动文件清单 + 未决点/待办」两清单
+- **编码纪律**：`discipline.md` 工作流程 + 调试策略要求用 `task` 跟踪；「改码前读文档」「中/小改后更新文档」嵌入 Workflow 各 tier 箭头序列
+
+### Fixed
+
+- 轮末探索摘要边界 `_runStartHistoryLen` 在压缩重建机器线后变 stale → 重置到 tail 起点
+
+## [0.12.38] — 2026-08-23
+
+### Added
+
+- **对齐 thinworker 编程工具集**：新增 `file_ops`（move/copy/rename）、`process`（列进程）、`get_current_time`、`sleep` 四个内置工具，及 `tree`（递归目录树 + `depth`，对齐 thinworker `repomap`）；各工具描述带「Route to X instead of bash」反向路由
+- **内部能力对齐 thinworker**：`grep` 加 `literal`/`ignoreCase`；`ls` 加 `filter`（通配符）；`bash` 加 `filter`（正则行过滤）；`git` 加 `show`/`rm`/`commit`/`push` 子命令 + `filter`；`verify` 加 `filter`/`workdir`
+- **`execute` 工具重做**：由同步 vm 沙箱改为 `node --input-type=module --eval` 子进程——支持顶层 `await`、动态 `import()`（直接加载项目 `.mjs`）、原生 `console`/`fetch`，新增 `workdir`/`filter` 参数，保留 killable 超时（无限循环可被终止）；`exec-prelude.mjs` 提供 readFile/writeFile/glob/grep/log/require（路径隔离 workspace）
+- **consult 指定子集模型**：`consult_start` 新增 `models` 选择器（`provider:model`/裸 provider/裸 model，大小写不敏感、去重保序）
+- **模型规格更新**：新增 glm-5.3、deepseek-v4-flash-vision-exp；gpt-5.6/claude-5 收敛 `thinking=false`
+
+### Fixed
+
+- **apply_patch 多 hunk 错位**：同一文件多个 hunks 因前一段改变行数而错位应用、静默损坏——改为重扫上下文 + splice
+- **git 写操作吞错**：`commit`/`push`/`rm` 改用 `runGitStrict` 返回 stderr + exit code，失败不再伪装成功；`show` 补 ref 校验（防 `-` 开头注入选项）
+- **tool 补齐项加固**：`get_current_time` 补 weekday；`sleep` 防 NaN；`tree` 深度/计数校验、省略号单次、根目录报错；`log` 非法 count 回退默认
+
+### Docs
+
+- `TOOLS.md` 注册表计数同步（含 ops/tree）
+- Changelog backfill 0.12.36 / 0.12.33
+
+## [0.12.37] — 2026-08-22
+
+### Fixed
+
+- **advisor 标签还原**：状态栏徽标与 /advisor 菜单项由 "GUARD" 显示名还原为 "ADVISOR"（内部配置字段 `advisor.guard` 不变，仅显示名）
+- **SSE 流式 tool_calls 防御性合并**（PROVIDER.md §10）：跳过 null/畸形元素并计数，按 index/id/name/tail 合并、补齐缺失 id、计数丢弃
+- **Windows 剪贴板 BOM 防御**：UTF-8 输出后剥离前导 `\uFEFF`（IK9UWM 后续）
+- **上下文压缩空安全**：`t.function?.name` 防 tool name 缺失
+
+### Refactor
+
+- **session.mjs 拆分**：`migrateHashLength` 抽到 `session-migrate.mjs`，文件回到 ≤500 行硬限内
+- **/config embedding 保存重构**：`embeddingPatch()` 抽取 + `DEFAULTS` 导出（供单测）
+
+### Prompt
+
+- **确认纪律 carve-out**：system.md 补 doc/code 一致性例外——"改动前确认"门禁不适用于已确认任务的既有义务（文档跟码一致、记录刚做的决策、关闭 advisor 标记的文档缺口）
+
+## [0.12.36] — 2026-08-22
+
+### Docs & advisor
+
+- **文档归属纪律 + advisor 设计评审增强**（规格 AGENT-LOOP.md §12）：新建 `docs/design/README.md` 文档地图（板块→文档映射表 + 存量碎片"待合并（TODO）"标注 + 归属规则）；system.md 补文档归属纪律条款（写文档前先查地图定位所属板块——找到就改、不得为既有板块新建文件；确无归属才新建并登记；同一机制只在一处详述权威源、其余引用不复制）；advisor-design.md 加第 7 维 **Document ownership**（与现有文档矛盾 🔴、该并入却新建/重复描述 🟡）与引用纪律（引用原文用精确 file:line、未核实标注 unverified）；design 提示词 fallback 删除转硬加载（`loadPrompt` 同 round1/2/3 待遇，缺失即抛错——静默降级会丢 Approval Signal 规则致评审无法批准）；messages.mjs design 分支 Instructions 补 Methodology compliance 维度、存在文档地图时注入 Document Map 段供归属检查对照。两端 prompts 保持 byte-identical、测试同步覆盖
+
+## [0.12.35] — 2026-08-21
+
+### Changed
+
+- **advisor 开关语义重构**：评审能力恒启用——`advisor` 工具任何模式都可调用（删除 `advisor.enabled` gate，不再返回 "not enabled"）；开关语义收敛为 guard——收尾推回仅当 `advisor.guard === true`（默认 OFF，评审自愿调用，打开才强制）。工程模式行为不变（评审恒可用、guard 豁免）
+- **`advisor.enabled` 废弃**：字段不再读写，存量配置不迁移——旧 `{ enabled: true }` 用户升级后不再强制评审（pre-release 约定，CHANGELOG 说明即可）；/advisor 菜单删除 "Advisor ON/OFF" toggle，Guard 成为唯一开关
+
+### Prompt system
+
+- **提示词借鉴增量（kimi-code 对照）**：explore.md 新增 Thoroughness levels 三档（quick 单点定向 / medium 默认适度并行 / thorough 全面分析且报告须列出搜索过什么与没找到什么）；main.md Delegate well 补委派 explore 时在 task 描述中指定彻底度（未指定走默认）；system.md 确认理解句补 "including the most important acceptance criteria"；subagent 工具 description 同步补彻底度说明。两端 15 个 prompt 文件保持 byte-identical（新增 CLI 侧比对测试防漂移）
+- **开工前计划确认纪律**：system.md 追加无豁免纪律——任何写文件动作（write/edit/apply_patch/insert_after/delete/hashline_edit 及一切写文件的 bash）前必须纯文字复述理解+计划要点并等待用户明确确认（未确认/沉默/用户回复新问题或新要求 → 一律不动手；"这太明显了不用问"不是跳过理由；用户的新问题不是确认；需求变化后重新复述重新确认）；engineering.md 澄清完成后、写需求/设计文档前同样须把理解+计划文字化并等待确认。两端 15 个 prompt 文件保持 byte-identical（两端测试断言关键句）
+
+## [0.12.34] — 2026-08-18
+
+### Added
+
+- **/rename 命令** — 改会话标题（renameSlot 双写，与 VS Code 共享）
+
+### Fixed
+
+- **/config 候选池 effort picker 显示真实枚举** — 从固定 min/low/medium/high/max 改为动态读 specForModel(model).reasoningEffortEnum；无枚举的模型跳过 effort 步
+- **question 工具 options 防御** — LLM 误传对象时取 label 字段，避免渲染 [object Object]
+
+## [0.12.33] — 2026-08-17
+
+### Changed
+
+- **撞轮数墙可无限继续**：subagent/飞刀/会诊统一经 continue 面板无限续（resume 保留 history，会诊继续重置墙钟；去掉 MAX_RESUMES）
+- **MiMo 预置 provider**：按量付费(api.xiaomimimo.com/v1) + Token Plan(token-plan-cn.xiaomimimo.com/v1)，模型规格 mimo-v2.5-pro/mimo-v2.5（1M 上下文 / 128K 输出 / 深度思考）
+- **环境变量配置源彻底移除**：THINCODER_* 回退全部删除，config.json 唯一配置源；空配置不再合成 deepseek 默认 provider
+- **effort 枚举钳制 + qwen3.8-max spec 修正**（consult/escalate）
+
+## [0.12.32] — 2026-08-16
+
+### Fixed
+
+- **会诊触发条款重构**（两轮会诊驱动的修复）：触发规则从飞刀段移入会诊段且自包含——功能请求语义（"会诊一下"触发、"consult the docs"不触发）+ 用户请求覆盖自主判断；consult_start 描述补对称触发句；飞刀段补 fly-in/口语变体
+- **飞刀三个真实断链**（会诊发现，此前"代码在但真实跑必翻车"）：
+  - 删墙钟看门狗——固定墙钟误杀正常但慢的手术（实测两个 max-effort 顾问读 5 个文件即撞 10min 墙）；完全依赖 turns + FETCH_TIMEOUT + 用户 Stop 直传
+  - effort 枚举钳制——池 effort 越界不再让候选"起飞即死"，回退预设并标注
+  - AUTO 传导对齐 subagent——headless 嵌入下父 autoApprove 正确放行子 agent 写操作
+- **config 加载校验**：consultModels 池 provider 名必须存在于 providers[]，条目形状校验——静默运行时失败改为启动即报错
+- **撞墙可继续（kimi-k3 飞刀）**：escalate 子 agent 撞 turn 上限后弹"继续?"（复用 onPermissionRequest，TUI 同款 y/n 面板），resume:true 续跑不重复注入任务、预算重置，上限 2 次；顺带修复 ContinueError e.turns → e.turn（原来打印 "undefined turns"）
+- consult 死代码补 precheck（无 key 时明确失败回复而非原始 401）
+
+### Docs
+
+- CLI CONSULTATION.md / ESCALATE.md 文档地图收录、FEATURES.md 功能全览补齐（7→11 个）、README 会诊别名、checklist T2-T6 验收
+
+## [0.12.31] — 2026-08-16
+
+### Fixed
+
+- **/config consult model picker**: adding a consult model now uses pickModelForSlot (provider AND model are both OPTION pickers, reusing /model's async-fetched model list) — was free-text for the model name
+- **Prompt adaptation**: CLI main.md was missing the consult (会诊) + escalate (飞刀) sections — the CLI main agent did not know "会诊" meant consult_start. Ported both + added the 会诊 alias to consult_start's tool description
+
+### Prompt system
+
+- Attention optimization + cross-end consistency: split over-long sentences, fixed an escalate-timing contradiction (up-front ability judgment), unified Review discipline + advisor rounds — all 15 prompt files byte-identical with the plugin
+
+### Docs
+
+- CONSULTATION.md + ESCALATE.md design docs (CLI implementation differences vs the plugin)
+
+## [0.12.30] — 2026-08-16
+
+### New: 会诊 (consult) + 飞刀 (escalate) — full VS Code plugin parity
+
+- `consult_start` / `consult_check` / `consult_stop`: several configured models run as parallel independent read-only consultants — each with its own TUI activity card, `main_history` access to the failure trail, arrival-order reply queue, stopped/terminated settle states, wall-clock watchdog
+- `escalate`: fly in a stronger model for one expert implementation run — coder role, full write path, permission gate, mutations merge into the parent's verify/advisor guards, turn-cap reads as partial work, timeout reads as timeout
+- Config: `agent.consultModels` ([{ provider, model, effort? }], up to 5, validated), `agent.consultTurns` (40), `agent.consultTimeoutMs` (600000)
+- **/config now manages the consult pool**: list / add / remove models; per-model reasoning-effort is an OPTION picker (none/min/low/medium/high/max); consultTurns + consultTimeoutMs (entered in minutes)
+
+### Discipline
+
+- UI rule added: fixed-choice values must be OPTIONS (picker/menu), never free-text — free-text only for genuinely open-ended input
+
+## [0.12.29] — 2026-08-16
+
+### Fixed
+
+- **Coder sub-agents (subagent role=coder) get verify + advisor** — CLI parity with the plugin's escalate diagnosis: the system prompt names both tools but the tool table only gave them to eng-coder; a coder sub-agent hit unknown-tool and self-verified via bash
+- **Cache-audit follow-ups**: OS/cwd reminder injected once per process (was every run); interrupt-resume now re-grounds the time (was stuck on the pre-interrupt time); skills scan sorted deterministically (filesystem-dependent readdir order could byte-change the system prompt with zero content change)
+
+## [0.12.28] — 2026-08-16
+
+### Cache-hit-rate fix (user-reported low hit rate on session start)
+
+- **Machine line (contextHistory) now keeps transient messages on persist** — every CLI invocation is a new process; the previous reals-only reload plus fresh re-injections (git/OS/outline/doc/memory/time) diverged at index ~1 → whole-prefix cache miss on the first request of every session. Resume now rebuilds a byte-identical machine line; new injections append at the tail
+- Time reminder moved to the END of the message sequence (after the user input) — aligned with the plugin fix, robust against any future machine-line disk reload
+
+### Fixed
+
+- normalizeToolPairing early-return hole: toolById empty must not skip placeholder filling when assistant tool_calls are declared (dangling tool_calls 400 otherwise)
+
+## [0.12.27] — 2026-08-15
+
+- Time injection moved OUT of the system prompt into a transient per-run user reminder — system prompts fully static again (prefix caches hit across hours, not minutes); local time + IANA timezone at second precision; now covers ALL agent depths (subagents previously had no time grounding at all)
+
+## [0.12.26] — 2026-08-15
+
+### ACP extensions for thincoder-desktop (proposals ①②③④, all implemented)
+
+- **① Session persistence**: every ACP turn end (success/cancel/failure — finally semantics) writes the session archive via saveSession; session/list / load / resume now have a real data source. Save is injectable and failures never break the queue
+- **② Checkpoints**: checkpoint/create / checkpoint/list / checkpoint/restore ACP handlers; NON-git cwds now snapshot by full-directory copy (v2 layout, nongit meta) instead of silently returning null
+- **③ Memory**: memory/list / memory/remove ACP handlers over the shared ~/.thincoder store
+- **④ Custom provider headers**: provider.headers object in config.json merges into every LLM request (chat + /models); Authorization cannot be overridden; non-string values sanitized out
+
+### Fixed
+
+- Time injection vs prefix-cache conflict: system-prompt "Current time" is now MINUTE precision — byte-identical within the same minute so DeepSeek prefix caches still hit (was: seconds precision broke the cache every run)
+
+## [0.12.25] — 2026-08-15
+
+- Local time + timezone injected into every system prompt (main agent, subagents, advisor) — prepareRun appends `Current time: <local> (<IANA zone>)`; sessionStart was ISO/UTC and session-scoped, subagents had nothing
+- bash.md Windows guidance corrected: the shell is cmd.exe (NOT Git Bash) — &&/|| work, cmd built-ins, NUL not /dev/null, prefer node -e for complex logic
+
+## [0.12.24] — 2026-08-14
+
+### Added
+
+- **glm-code provider preset** — the Zhipu GLM Coding Plan endpoint (`https://open.bigmodel.cn/api/coding/paas/v4`, glm-5.2, same key as GLM; server-side forced thinking).
+
+### Fixed
+
+- **Model specs synced with official vendor docs (verified 2026-08)** — DeepSeek v4 duals effort enum +low and cacheMode→auto; qwen3.x/max/plus maxOutput→131072 (qwen-plus was 32K).
+- **Retired models dropped** — deepseek-chat/reasoner, kimi-k2, moonshot v1 (vendor shutdowns; unknown IDs fall back to the 128K default spec).
+
+### Changed
+
+- Repository URL → github.com/xinbo-tech/thincoder.
+
+## [0.12.23] - 2026-08-13
+
+- **修复** svg 图片毒化会话——read_image 读 svg 后以 image_url 进历史，Kimi 等视觉 API（全部仅支持位图）此后每轮请求 400 "unsupported image format"，会话永久卡死；现在发送时按格式净化：非 png/jpeg/gif/webp 的 image part 替换为占位文本，净化上移至 format dispatch 之前覆盖 openai/anthropic/google 全部通路，历史本身不改写（切回支持的模型/格式可恢复）
+- **改进** read_image 对 svg 返回文本源码（svg 本是文本标记，任何模型可读，绕过 vision gate）；bmp 拒绝并提示转 PNG（无主流视觉 API 支持）
+
+## [0.12.22] - 2026-08-13
+
+- **修复** 缓存命中率对 Kimi 显示——usage 缓存字段归一化：Kimi/OpenAI 风格 `prompt_tokens_details.cached_tokens` 映射为 DeepSeek 风格 `prompt_cache_hit_tokens`，miss 由 prompt_tokens − hit 推导（此前 Kimi 的命中率永不显示）
+- **安全** fetch 重定向目标做 SSRF 检查——3xx 可把公网 URL 跳进内网（重定向绕过）；相对 URL 正确解析、仅 http/https、私网/元数据拦截
+- **新增** Tavily 结构化搜索（可选）——config 配 `websearch.apiKey` 后 `websearch` 走 Tavily API（稳定 JSON，不再依赖 Bing 页面结构）；无 key 回退 Bing 抓取，零门槛不破坏
+
+## [0.12.21] - 2026-08-13
+
+- **修复** 恢复会话大量重复 "❯ ThinCoder:" 标签——history 按每次 LLM 调用存一条 assistant 消息（一个 turn 多段），恢复时每段都渲染了标签；现在只在 turn 开始渲染一次，跨页懒加载边界状态正确保留
+- **改进** 恢复保真度：完整工具结果（不再一行摘要）+ reasoning 思考流以 dim 行恢复（超长自动折叠）——恢复后的会话与退出前基本一致；首帧渲染实测约 50ms
+
+## [0.12.20] - 2026-08-13
+
+- **修复** TUI 恢复旧 display 快照导致"看不到最新消息"——display 字段彻底废弃（saveSession 不再写、loadSession 不再读），恢复永远从 history 重建；配合 VS Code 端 0.1.5 的清空，跨端会话漂移根治
+- **新增** TUI 懒加载历史恢复：启动只物化最近 200 条消息（8000+ 条会话不再冻结启动），PgUp 到顶按 50 条/页加载更早历史，scroll 补偿保持视觉位置
+- **新增** question 选项列表末尾追加"✍ Custom answer…"——选中切自由输入，用户可补充/修正 AI 的预设选项
+- **重构** execute 工具移除假沙箱：require()/process 全可用（bash 本就能触达任意 Node API，拦 require 只会误导模型）；移除动态 import 拦截与 SSRF 私网拒绝；保留 timeout / cwd 约束 / 输出上限等工程保护
+- **改进** 工程模式 prompt：新增提问风格指引（默认开放式自由文本，选项仅用于有限枚举）；审查修复 5 处（需求优先步骤、designToken 仅走参数、用户审批呈现 advisor 发现、澄清完成判据、advisor 重试 3 轮上限）
+
+## [0.12.19] - 2026-08-11
+
+- **重构** bash 工具安全模型：移除全部破坏性命令文本拦截（rm -rf / DROP TABLE 等）——文本匹配是安全剧场（恶意模型可绕过、误伤正常操作），真实防线 = 审批层（autoApprove）+ 快照（gitGuardSnapshot / checkpoint），与 env 透传、git"快照后放行永不拦截"统一
+- **新增** 危险命令标注（detectDanger，参考 kimi-code）：recursive-delete / sudo / pipe-to-shell / dd / mkfs / 裸设备 / chmod 777 / fork bomb 在 TUI 与 ACP 审批面板红色标注——只提示不拦截，帮人做审批决策；引号感知（commit message 等纯文本不误标，反引号内容保留）
+- **文档** TOOLS.md 安全边界同步（零文本拦截 + 危险标注）
+
+## [0.12.18] - 2026-08-10
+
+- **修复** Qwen 路由等代理的模型 ID（如 `kimi/kimi-k3`）跳过 `reasoning_effort` 参数——路由可能误处理该参数导致空回复
+
+## [0.12.17] - 2026-08-10
+
+- **修复** `kimi/kimi-k3`（Qwen 路由前缀 ID）模型规格匹配 — 正确继承 kimi-k3 的 1M 上下文 / 131K 输出 / 多模态
+
+## [0.12.16] - 2026-08-10
+
+- **修复** 非 SSE JSON 响应被误判为错误（HTTP 200 + tool_calls 的合法 JSON 响应现在正确解析，而非报 "Response is not SSE"）
+- **改进** API 错误信息可读性：非 SSE 错误响应包含 HTTP 状态码 + 具体错误原因
+
+## [0.12.15] - 2026-08-10
+
+- **改进** API 错误信息可读性：非 SSE 响应预拦截，提取 HTTP 状态码 + 具体错误原因（替代原来模糊的 "Response is not SSE"）
+
+## [0.12.14] - 2026-08-10
+
+- **修复** 小终端 permission 面板挤出输入框（layout 溢出补偿新增 permission 压缩）
+- **修复** iTerm2 Ctrl+C 键盘协议序列泄漏（stdin 剥离未处理的 CSI u 序列）
+- **修复** eng 模式 advisor token 正则错配（改用完整 token 构建正则，与 prompt 格式一致）
+- **修复** 输入框 ↑ 键历史导航草稿丢失（进入/编辑历史模式时扩容草稿保护）
+- **重构** key-handler 拆分搜索模块（key-handler-search.mjs）
+- **文档** 架构文档计数/模块/状态同步更新
+
+## [0.12.13] - 2026-08-08
+
+评审机制全面重构（用户驱动的三轮决策）：
+
+- **prior 硬解析移除**：收敛轮注入上一轮评审的完整原文（模型直接理解），删除表头匹配与 all-clear 短语两类"字符串解析 LLM 输出"的脆弱机制
+- **评审触发范围收缩**：评审只跟代码修改绑定——bash/git 等副作用工具不再触发多余评审轮（评审后读日志/清理临时文件不再要求重复评审）
+- **AGENTS.md 文档地图**：需求基线声明（REQUIREMENTS.md + 设计文档 + 对话背景）+ docs/design/ 27 份文档分组清单，评审者按地图定位需求文档
+- **项目根发现**：多项目工作区从评审范围定位子项目 AGENTS.md（工作区元地图不遮蔽）；修复混合路径分隔符误判
+- **收敛体共享模块**：round 2+ 消息构建单一来源；空回复/纯工具输出不再冒充评审记录
+
+## [0.12.12] - 2026-08-07
+
+- advisor 记录按真实时序落盘（timeline）、markdown 表格 render-before-measure 对齐修复（含 heading 多行/双重粗体）、requirements 兜底、评审结论可用性提示
+- 双线消息历史（人读线 + 机读线）、压缩只作用于机读线、机读消息不进人读线
+- 临时文件（tmp-*）不触发 advisor guard；config.mjs 加固（spec 预排序、providers 守卫、saveConfig 写副本）
+- VS Code 扩展发布准备（marketplace 元数据、.vscodeignore、vscode-mock 依赖修复）
+
+## [0.12.11] - 2026-08-05
+
+- subagent 按类型配置模型（`/submodel` + `subagentModels`）
+- 可配置 bash shell（`/shell platform` 切换）
+- 其他稳定性与体验改进
+
+## [0.12.10] - 2026-08-05
+
+- 代码质量梳理：清理未使用的导出、advisor 计时器与静态导入修复、复评不再因旧会话数据误报
+
+## [0.12.9] - 2026-08-04
+
+- 提示词体系质量梳理：移除工程模式与 advisor 的冲突、交付评审语义修正
+
+## [0.12.8] - 2026-08-04
+
+- pending-task 推回最多触发一次（消除无界完成循环）
+
+## [0.12.7] - 2026-08-03
+
+- 折叠可读性修复（主输出/思考永不折叠）、窄终端宽表格裁剪
+
+## [0.12.6] - 2026-08-02
+
+- checkpoint v2、git 破坏性命令保护、鼠标支持、长消息折叠、bash 行为约束
+
+## [0.12.5] - 2026-08-01
+
+- 行内代码下划线样式
+
+## [0.12.4] - 2026-08-01
+
+- 压缩统一规范、Kimi For Coding、Ctrl+C 双重确认、空响应重试、markdown 渲染修复
+
+## [0.12.3] 及更早
+
+v0.12.x 早期版本、v0.11.x、v0.8.x、v0.7.x 与 v0.2–v0.6 系列——完整历史见 [git 提交记录](https://gitee.com/shanghai-xinbo/thincoder/commits/main)。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "thincoder",
-  "version": "0.12.38",
+  "version": "0.12.40",
   "description": "Thin coding agent - zero dependencies, no build step, Node.js native. Sharp code, zero bloat.",
   "keywords": [
     "ai",
@@ -23,11 +23,12 @@
     "bin/",
     "src/",
     "README.md",
+    "CHANGELOG.md",
     "LICENSE"
   ],
   "repository": {
     "type": "git",
-    "url": "https://github.com/xinbo-tech/thincoder.git"
+    "url": "git+https://github.com/xinbo-tech/thincoder.git"
   },
   "scripts": {
     "test": "node --test \"test/*.mjs\"",

package/src/agent.mjs

@@ -4,7 +4,7 @@
  */
 import { chat } from "./provider/index.mjs"
 import { estimateText } from "./provider/rate.mjs"
-import { compressIfNeeded, compressFallback, COMPRESS_FAILURE_LIMIT, pushReal } from "./context.mjs"
+import { compressIfNeeded, compressFallback, COMPRESS_FAILURE_LIMIT, pushReal, summarizeRunExplorations } from "./context.mjs"
 import { specForModel } from "./config.mjs"
 import { readFileSync } from "node:fs"
 import { join, dirname } from "node:path"
@@ -95,6 +95,8 @@ export function createAgent({
     _sessionStart: sessionStart,
     _lastPromptTokens: null, _usageAtLen: null,
     _compressFailures: 0,
+    _emptyRetries: 0, // empty-response retry budget (per-run; reset on a fresh user turn)
+    _runStartHistoryLen: 0, // machine-line length at the start of the current run — end-of-run exploration distillation slices from here
     _currentTurn: 0, _maxTurns: 100, // turn counter for status bar display
   }
 }
@@ -106,6 +108,10 @@ export async function runAgent(agent, input, callbacks = {}, { depth = 0, signal
     { depth, signal, overrideTurns, resume, systemPrompt: SYSTEM_PROMPT, disciplineRules: DISCIPLINE_RULES, mainOverlay: MAIN_OVERLAY },
   )
 
+  // End-of-run exploration distillation boundary (CONTEXT-COMPACTION §5): prepareRun has already
+  // pushed the user input + injections, so everything appended from here is "this run's" work.
+  agent._runStartHistoryLen = agent.history.length
+
   // Per-run bookkeeping reset. On `resume` (ContinueError continuation) these are
   // PRESERVED: the resumed run must keep mutation tracking so the advisor/verify
   // guards stay active (a guard pushback on the last turn must not silently vanish),
@@ -301,6 +307,12 @@ export async function runAgent(agent, input, callbacks = {}, { depth = 0, signal
       honestReminderInjected = cr.honestReminderInjected
       advisorPushbacks = cr.advisorPushbacks
       if (cr.action === "continue") continue
+      if (depth === 0) {
+        // End-of-run exploration distillation (CONTEXT-COMPACTION §5): this run's inline
+        // exploration results become one semantic note before the final return. Silent (N3):
+        // distillation failure must never block the return or lose history.
+        try { await summarizeRunExplorations(agent, callbacks, signal) } catch { /* silent (N3) */ }
+      }
       return cr.content
     }
 
@@ -407,11 +419,6 @@ export async function runAgent(agent, input, callbacks = {}, { depth = 0, signal
           // how many reviews have run (round 1→2→3→4→5), not how many succeeded.
           // A failed/interrupted review is still a review attempt and should use
           // the next round's prompt on retry.
-          try {
-            JSON.parse(toolCall.arguments || "{}")
-          } catch {
-            /* arguments unparseable — still counts as a review attempt */
-          }
           agent._advisorRound++
         }
         if (FILE_MUTATORS.has(toolCall.name)) {

package/src/context.mjs

@@ -48,13 +48,14 @@ function keepTailSize(provider, historyLen) {
   return Math.min(Math.max(10, Math.floor((ctxWindow / 100_000) * 30)), Math.floor(historyLen * 0.4))
 }
 
-const SUMMARIZE_PROMPT = `You are a conversation compressor. Summarize the following agent work log into a compact summary for use as context in the ongoing conversation.
+export const SUMMARIZE_PROMPT = `You are a conversation compressor. Summarize the following agent work log into a compact summary for use as context in the ongoing conversation.
 Requirements:
 - Write in first person, present tense — these are "my" handover notes, continuing my own train of thought
 - Most important: preserve design decisions and their reasons — architecture choices, API contracts, naming conventions, trade-off rationale. These are the anchors the subsequent code must not deviate from
 - Distinguish COMPLETED vs IN-PROGRESS work: completed tasks get a ONE-LINE recap each (what was done, key outcome); spend the detail budget on unresolved issues, next steps, and the CURRENT task
 - The user's most recent request defines the current task — anchor on it. Earlier requests are likely already completed and only need the one-line recap; do NOT preserve them at full fidelity
-- Keep: files modified and why, unresolved issues, next steps
+- Explicitly list FILES CHANGED: every modified file path plus a one-line "why" — so post-compaction work can re-locate what was edited and where
+- Explicitly list UNRESOLVED ISSUES / TODOs: anything still open plus the next steps — so post-compaction recovery knows where to resume
 - Drop: pleasantries, repetition, fine-grained tool output details
 - Honestly mark uncertain items: anything not actually verified must say "unverified"; do not present guesses as facts
 - Use bullet-point output; aim for information completeness, not a hard word limit (old 500-char cap is deprecated; in a 1M-context era, err on the long side)
@@ -152,6 +153,15 @@ function applyCompression(agent, headEnd, tailStart, note) {
     { role: "assistant", content: "Understood. I'll continue from these notes, re-verifying anything transient." },
     ...tail,
   ]
+  // Compaction REBUILDS the machine line (head + note + "Understood" + tail), so the pre-compaction
+  // _runStartHistoryLen index is stale — a longer array shrank beneath it, and end-of-run exploration
+  // distillation would then silently skip or slice from the wrong offset. Reset the boundary to the
+  // verbatim tail start (head.length + 2: the note and the "Understood" placeholder sit between head
+  // and tail). Exploration before the tail was already covered by the compaction summary, so only the
+  // still-raw tail needs distilling. `head` is empty today (KEEP_HEAD = 0) — the formula stays
+  // correct if KEEP_HEAD ever grows. (shrinkOversized only truncates message bodies in place and
+  // leaves the array length unchanged, so this boundary stays valid there — no reset needed.)
+  agent._runStartHistoryLen = head.length + 2
   // Measured token baseline is invalidated along with old history (prompt_tokens were for pre-compaction context), fall back to estimation until next response
   agent._lastPromptTokens = null
   agent._usageAtLen = null
@@ -262,22 +272,167 @@ const OVERSIZE_CONTENT_LIMIT = 8_000
  */
 function shrinkOversized(agent, limit = OVERSIZE_CONTENT_LIMIT) {
   let shrunk = false
-  for (const m of agent.history) {
-    if ((m.role !== "user" && m.role !== "tool") || typeof m.content !== "string") continue
-    if (m.content.length <= limit) continue
+  // Copy-on-write: build a NEW array and replace only truncated entries. pushReal stores the SAME
+  // message object in both `agent.history` (machine line) and `agent._fullHistory` (human/persistence
+  // line), so in-place `m.content = ...` would ALSO truncate the never-compacted human line and lose
+  // the original pasted content on session persist (session.mjs persists _fullHistory). VS Code port
+  // already copies (`history.map(m => ({ ...m }))`); this brings CLI to parity.
+  const next = agent.history.map((m) => {
+    if ((m.role !== "user" && m.role !== "tool") || typeof m.content !== "string") return m
+    if (m.content.length <= limit) return m
     // Truncate keeping head + tail, insert stub in between; keepHead/keepTail proportional but not exceeding 50%/25% of limit
     const keepHead = Math.min(Math.floor(limit * 0.5), 4000)
     const keepTail = Math.min(Math.floor(limit * 0.25), 2000)
-    m.content =
-      m.content.slice(0, keepHead) +
-      `\n[... ${m.content.length - keepHead - keepTail} chars truncated — single message too large for context window ...]\n` +
-      m.content.slice(-keepTail)
     shrunk = true
-  }
+    return {
+      ...m,
+      content:
+        m.content.slice(0, keepHead) +
+        `\n[... ${m.content.length - keepHead - keepTail} chars truncated — single message too large for context window ...]\n` +
+        m.content.slice(-keepTail),
+    }
+  })
   if (shrunk) {
+    agent.history = next
     // Same as compaction: measured token baseline is invalidated by the changed history, fall back to estimation until next response
     agent._lastPromptTokens = null
     agent._usageAtLen = null
   }
   return shrunk
 }
+
+// ─── End-of-run exploration distillation (AGENT-LOOP §13 + CONTEXT-COMPACTION §5, 2026-08-23) ───
+// The main agent's machine line is flooded by inline step-by-step exploration (read/grep/...).
+// At run end we distill THIS run's exploration tool-results into one semantic summary note that
+// replaces them in the machine line, while agent._fullHistory (the human line) stays untouched.
+
+/** Read-only knowledge tools counted as "exploration" (execute writes files → never exploration). */
+export const EXPLORE_TOOLS = new Set([
+  "read", "grep", "glob", "ls", "code_search", "doc_search", "repo_outline",
+])
+
+/** Summary prompt for turning a burst of exploration results into a semantic summary. */
+export const EXPLORE_SUMMARY_PROMPT = `You are distilling exploration tool results. Summarize the following read-only codebase exploration into a compact semantic summary for the main agent's own context.
+
+Requirements:
+- Capture WHAT was discovered, WHERE (which files / directories / symbols), and the KEY CONCLUSIONS — do not list tool calls mechanically
+- Keep actionable facts the main agent needs to continue: code locations, function names, file paths, structure, and open questions the exploration raised
+- Drop raw tool-output noise, repeated lines, and verbatim file dumps — keep only what must be remembered
+- Be honest: mark anything not actually verified as "unverified"; do not present guesses as facts
+- Use bullet points; aim for information completeness, not a hard word limit
+
+Exploration log:
+`
+
+/** tool_calls name across both stored shapes ({function:{name}} and flat {name}). */
+function toolCallName(tc) {
+  return tc?.function?.name ?? tc?.name ?? ""
+}
+
+/** Tool that produced a tool-result message (falls back to its owner assistant's tool_call). */
+function toolResultName(msg, ownerToolCalls) {
+  if (typeof msg?.name === "string" && msg.name) return msg.name
+  const owner = (ownerToolCalls ?? []).find((tc) => tc.id === msg?.tool_call_id)
+  return owner ? toolCallName(owner) : ""
+}
+
+/**
+ * Find the pure-exploration "assistant(tool_calls)→tool…" pair blocks added since `start`.
+ * A block is explorable only when EVERY tool call AND every tool result in it is an exploration
+ * tool — mixed blocks (read + edit in one turn) stay untouched, or we'd orphan the edit pairing.
+ */
+function findExplorationBlocks(history, start) {
+  const blocks = []
+  let i = start
+  while (i < history.length) {
+    const m = history[i]
+    if (m?.role === "assistant" && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
+      let j = i + 1
+      while (j < history.length && history[j]?.role === "tool") j++
+      const toolMsgs = history.slice(i + 1, j)
+      const allCallsExplore = m.tool_calls.every((tc) => EXPLORE_TOOLS.has(toolCallName(tc)))
+      const allResultsExplore = toolMsgs.length > 0 && toolMsgs.every((t) => EXPLORE_TOOLS.has(toolResultName(t, m.tool_calls)))
+      if (allCallsExplore && allResultsExplore) {
+        blocks.push({ start: i, end: j, messages: history.slice(i, j), toolCount: toolMsgs.length })
+      }
+      i = j
+    } else {
+      i++
+    }
+  }
+  return blocks
+}
+
+/** Serialize a batch of exploration messages for the summary LLM (same shape as compaction serialization). */
+function serializeExplorationMessages(messages) {
+  const cap = 8000 // exploration results ARE the signal to distill — generous cap (quality-first, N1)
+  return messages
+    .map((m) => {
+      const toolNote = m.tool_calls ? ` [called tools: ${m.tool_calls.map(toolCallName).join(", ")}]` : ""
+      let text = ""
+      if (typeof m.content === "string") text = m.content
+      else if (Array.isArray(m.content)) text = m.content.filter((p) => p?.type === "text").map((p) => p.text ?? "").join(" ")
+      return `[${m.role}]${toolNote} ${text.slice(0, cap)}`
+    })
+    .join("\n")
+}
+
+/**
+ * Core (shared) distillation: replace this run's pure-exploration pair blocks with a single
+ * "[Exploration summary]" note placed where the first block was. Returns a NEW history array,
+ * or null when there is nothing to shrink (<3 exploration results / LLM failure). Pairing-safe:
+ * whole assistant→tool blocks are removed, so no orphan tool_calls/tool can survive.
+ */
+async function distillExplorations(history, start, provider, signal) {
+  if (!Array.isArray(history) || history.length - start < 2) return null
+  const blocks = findExplorationBlocks(history, start)
+  const resultCount = blocks.reduce((n, b) => n + b.toolCount, 0)
+  if (resultCount < 3) return null
+
+  const serialized = blocks.map((b) => serializeExplorationMessages(b.messages)).join("\n")
+
+  let summary
+  try {
+    // Silent by design (D11): thinking:null and no onToken/onReasoning — this internal
+    // distillation must not stream to the frontend. signal propagates user cancellation.
+    const resp = await chat({ ...provider, thinking: null, reasoningEffort: null }, {
+      messages: [{ role: "user", content: EXPLORE_SUMMARY_PROMPT + serialized }],
+      signal,
+    })
+    summary = resp?.content
+  } catch {
+    return null // N3: never block the run's return or lose history — original results stay
+  }
+  if (!summary) return null
+
+  const drop = new Set()
+  for (const b of blocks) for (let k = b.start; k < b.end; k++) drop.add(k)
+  const note = { role: "user", content: "[Exploration summary]\n" + summary }
+  const next = []
+  let inserted = false
+  for (let k = 0; k < history.length; k++) {
+    if (drop.has(k)) {
+      if (!inserted) { next.push(note); inserted = true }
+      continue
+    }
+    next.push(history[k])
+  }
+  return next
+}
+
+/**
+ * End-of-run exploration distillation (runAgent's final return). Shrinks the MACHINE line
+ * (agent.history) only; agent._fullHistory is never touched. Triggers when this run added ≥3
+ * exploration tool results; on LLM failure it silently keeps the original history (N3).
+ * `callbacks` is accepted for call-site parity with the other lifecycle hooks — the distillation
+ * is silent by design and never streams (D11).
+ */
+export async function summarizeRunExplorations(agent, callbacks, signal) {
+  const next = await distillExplorations(agent.history, agent._runStartHistoryLen ?? 0, agent.provider, signal)
+  if (!next) return
+  agent.history = next
+  // The machine line changed shape — the measured token baseline was for the pre-shrink context.
+  // Invalidate so the next compaction check re-estimates instead of over-counting stale history.
+  agent._lastPromptTokens = null
+  agent._usageAtLen = null
+}

package/src/prompts/discipline.md

@@ -1,10 +1,14 @@
 Workflow — match the process to the task:
-- Complex (3+ steps, new features): Requirements → Design → Development → Testing. Write a design doc. Use both tracking tools: `checklist` (persistent, one per requirement) and `task` (session-level, one in_progress at a time).
-- Medium (2-3 steps, refactoring): plan briefly, no design doc needed. Use `task` tool.
-- Small (typo, one-line fix): confirm understanding, change, verify. No design doc.
+- Read the relevant docs before changing code — at ANY tier: doc_search the topic, then locate the owning design doc via docs/design/README.md (the document map) and read it — plus AGENTS.md if present.
+- Use `task` to track work for EVERY tier — one item in_progress at a time.
+- Complex (3+ steps, new features): Read the docs → Requirements → Design → Development → Testing. Write a design doc. Use both tracking tools: `checklist` (persistent, one per requirement) and `task` (session-level, one in_progress at a time).
+- Medium (2-3 steps, refactoring): Read the docs → Plan → Change → update the owning doc if you spotted a gap — a decision not yet recorded, or a doc now contradicting the code. No design doc needed. Use `task` tool.
+- Small (typo, one-line fix): Read the docs → Change → Verify → update the owning doc if you spotted a gap — a decision not yet recorded, or a doc now contradicting the code. Use `task` tool. No design doc.
 - If unsure which tier, treat as complex. Under-planning costs more than over-planning.
+- Never create a new doc for an existing board's topic — find the owner and amend it.
 
 Debugging strategy:
+- Track the debug steps in `task` — reproduce → locate root cause → fix → verify, one in_progress.
 - Read the full error output — root cause is often at the end.
 - Verify against official docs before guessing.
 - Binary search: cut the problem in half, test which half has the fault.
@@ -24,4 +28,4 @@ Review discipline (standard mode only — engineering mode has its own review ti
 - **No "pre-existing" cop-out.** You own the whole code. "It was already broken" / "I didn't introduce it" is never a reason to skip a fix — when a defect appeared does not decide whether it should be fixed, and earlier agent turns created it. Rebut only on technical grounds, otherwise fix it.
 - **Do not bury 🔴.** A 🔴 you neither fix nor rebut blocks convergence. `Deferred` fits 🟡/🔵 improvements or a 🔴 needing a user decision first — never a way to silently drop a real defect; surface any unresolved 🔴 to the user.
 - Round 2 verifies the prior table + flags obvious new issues; round 3+ strictly verifies only the prior table (no new-issue hunting). Max 5 rounds total.
-- When the advisor reports all clear (no 🔴 remaining), run `verify`.
+- When the advisor reports all clear (no 🔴 remaining), run `verify`.

package/src/prompts/main.md

@@ -9,11 +9,13 @@ Explore the codebase read-only, design the architecture, present the plan. When
 For tasks that match the Coding discipline's "complex" tier, plan mode is your design step; for "medium" tasks it's optional but recommended.
 
 Delegate well — spawn subagents for independent subtasks.
+- Subagents run in an isolated context: their step-by-step read/grep never enters your history — only their final report comes back. Doing the same broad exploration inline floods your own window with noise and degrades your attention across turns.
 - Explore agents for parallel codebase search, plan agents for architecture design, coder agents for self-contained implementation.
 - When delegating an explore agent, state the thoroughness in the task description — quick / medium / thorough — graded by need; unspecified means the default.
-- Delegate breadth-first exploration; do precision edits yourself.
+- Breadth-first exploration — understanding that spans multiple files / directories (finding usages, mapping structure, reading a batch of files) — goes to an `explore` subagent, with thoroughness (quick / medium / thorough) annotated in the task.
+- Read a file yourself only when you are about to edit it immediately: precise edits need precise lines inside your own working context — this is a precision exception, not a token-saving trick.
 - Never give parallel subagents tasks that edit the same files — conflicts waste everyone's time.
-- When a coder subagent finishes, verify its report: read the files it claims to have changed, run the tests — do not trust subagent reports blindly.
+- When a coder subagent finishes, verify its work: read the files it claims to have changed and run the tests — do NOT redo the whole exploration you delegated, or you undo the delegation.
 - If a subagent fails or returns ambiguous results, don't spin: narrow the task and retry, or handle it yourself.
 - Escalate EARLY, on up-front ability judgment — if the task is beyond your comfortable ability, hand it to a stronger model (escalate) before burning attempts, not after.
 - When multiple subagent reports conflict, read the relevant code yourself to arbitrate — never merge conflicting claims.