| docs(site): 更新 /undo 文档为对话记忆回退语义
v4.24.2 起 /undo 改为回退对话记忆(支持 /undo N)并把该轮 prompt
还原回输入框,不再回滚磁盘文件。同步修正中英 slash-commands /
sessions 两页正文与示例,并重建搜索索引。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| 14 小时前 |
| docs(site): 补全 v4.24.1 新功能文档（WebUI/同步/$ 菜单/后台与破坏性命令）
- 新增 docs/{zh,en}/webui.html：WebUI 启动与关闭、界面功能、/sync 多端实时同步、远程访问指引；
注册到进阶分组、docs.js 侧栏 i18n，所有页面侧栏加链接并重建搜索索引。
- slash-commands：新增「WebUI 与同步」分区（/webui、/webui --host、/webui stop、/sync、/sync off），
并在「配置与帮助」补 /guide。
- skills：新增「通过 $ 菜单」小节（行首 $ 列出/过滤技能，Tab 补全，$name 参数 调用）。
- tools：bash 增补「后台运行长任务」(run_in_background) 与「破坏性命令拦截」两个提示框。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| 1 天前 |
| feat: add graceful fallback for missing auth, promote CodingPlan in docs
| 1 个月前 |
| feat: major architecture overhaul — context management, smart truncation, agent discipline
Phase 1: Tool result externalization
- ToolResultRef + ToolResultStore: large tool outputs cached to disk
- Hot zone inflated from disk, cold zone uses summaries
Phase 2: Turn-aware context windowing
- TurnTracker: tracks conversation turns (user message → completion)
- Budgeted windowing operates at turn granularity
- Cold turns: only user question + assistant conclusion retained
- Synthesized outcomes for turns without assistant text
Phase 3: Smart per-tool truncation
- bash: preserve error lines, skip verbose build logs
- read_file: threshold raised to 2000 lines (prevents multi-read cycles)
- First read of any file forces full content (ignores offset/limit)
- edit_file returns compact diff + file outline after edit
Agent discipline:
- Sleep loop detection (warn at 2, block at 3+)
- Repeated command detection (normalized, warn at 2, block at 3+)
- Same-file multi-edit blocking (threshold 4)
- Re-read soft reminder (not blocking)
- Over-verification detection (3+ consecutive check commands)
- HTTP 200 detection → stop verifying
- 429 rate limit retry with exponential backoff (up to 5 retries)
- Loop detection allows retry after edit (fix-then-retry is not a loop)
- Intercepted/blocked calls don't count toward step limit
- Bash pipe/cat detection with redirect to proper tools
System prompt:
- Restored v1.3.1 proven rules + new additions
- ACT DON'T INSTRUCT / BE CONCISE / ONE SIGNAL IS ENOUGH
- COMMAND DISCIPLINE / ERROR HANDLING sections
- Glob wildcard guidance
- Tech-stack neutral (zero specific tool/language names)
Project context:
- Added pom.xml, build.gradle, requirements.txt, composer.json, CMakeLists.txt to descriptors
- Cross-session context: previous session turns injected into system prompt
- Conversation history loaded on startup (not fresh each session)
- Active services detection: port scanning on startup + URL extraction from bash output
- Frontend hot-reload hint when dev server detected
UX improvements:
- TTFT per LLM call (not cumulative), removed from status bar
- Spinner shows last tool name: "After read_file, thinking..."
- Tool call streaming: shows file_path/command as args stream in
- Turn completion summary: "✻ Completed in N steps, Xs"
- Streaming markdown: trim incomplete backtick/bold tokens
- Pangu spacing (CJK-ASCII)
- Code block styling refresh
- Append input during streaming (Enter sends additional context)
- UTF-8 safe string truncation (char boundary check)
Tooling:
- edit_file: atomic_write retry on rename failure (dev server file locks)
- edit_file: suggested fix on old_string mismatch
- glob: rewritten pattern parser (handles absolute paths, **/ splits)
- LLM request logging to ~/.atomcode/logs/
- Datalog analyzer script (scripts/analyze_datalogs.py) with --deep mode (Claude Code)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
| 2 个月前 |
| chore(release): bump version to v4.24.2
Update workspace crates, install scripts, READMEs, docker docs, and site
docs/badges from v4.24.1 to v4.24.2. latest.json intentionally unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| 14 小时前 |
| feat: major architecture overhaul — context management, smart truncation, agent discipline
Phase 1: Tool result externalization
- ToolResultRef + ToolResultStore: large tool outputs cached to disk
- Hot zone inflated from disk, cold zone uses summaries
Phase 2: Turn-aware context windowing
- TurnTracker: tracks conversation turns (user message → completion)
- Budgeted windowing operates at turn granularity
- Cold turns: only user question + assistant conclusion retained
- Synthesized outcomes for turns without assistant text
Phase 3: Smart per-tool truncation
- bash: preserve error lines, skip verbose build logs
- read_file: threshold raised to 2000 lines (prevents multi-read cycles)
- First read of any file forces full content (ignores offset/limit)
- edit_file returns compact diff + file outline after edit
Agent discipline:
- Sleep loop detection (warn at 2, block at 3+)
- Repeated command detection (normalized, warn at 2, block at 3+)
- Same-file multi-edit blocking (threshold 4)
- Re-read soft reminder (not blocking)
- Over-verification detection (3+ consecutive check commands)
- HTTP 200 detection → stop verifying
- 429 rate limit retry with exponential backoff (up to 5 retries)
- Loop detection allows retry after edit (fix-then-retry is not a loop)
- Intercepted/blocked calls don't count toward step limit
- Bash pipe/cat detection with redirect to proper tools
System prompt:
- Restored v1.3.1 proven rules + new additions
- ACT DON'T INSTRUCT / BE CONCISE / ONE SIGNAL IS ENOUGH
- COMMAND DISCIPLINE / ERROR HANDLING sections
- Glob wildcard guidance
- Tech-stack neutral (zero specific tool/language names)
Project context:
- Added pom.xml, build.gradle, requirements.txt, composer.json, CMakeLists.txt to descriptors
- Cross-session context: previous session turns injected into system prompt
- Conversation history loaded on startup (not fresh each session)
- Active services detection: port scanning on startup + URL extraction from bash output
- Frontend hot-reload hint when dev server detected
UX improvements:
- TTFT per LLM call (not cumulative), removed from status bar
- Spinner shows last tool name: "After read_file, thinking..."
- Tool call streaming: shows file_path/command as args stream in
- Turn completion summary: "✻ Completed in N steps, Xs"
- Streaming markdown: trim incomplete backtick/bold tokens
- Pangu spacing (CJK-ASCII)
- Code block styling refresh
- Append input during streaming (Enter sends additional context)
- UTF-8 safe string truncation (char boundary check)
Tooling:
- edit_file: atomic_write retry on rename failure (dev server file locks)
- edit_file: suggested fix on old_string mismatch
- glob: rewritten pattern parser (handles absolute paths, **/ splits)
- LLM request logging to ~/.atomcode/logs/
- Datalog analyzer script (scripts/analyze_datalogs.py) with --deep mode (Claude Code)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
| 2 个月前 |
| fix: shift+enter handling and optimize website performance
- Fix shift+enter key combination handling in tuix
- Optimize website performance: reduce dependencies, minify assets
| 1 个月前 |
| docs(site): refresh main + docs site with i18n, search and unified design
- rewrite main page as self-contained warm-dark + orange theme (no Tailwind)
- restyle docs to match main: counters-numbered sidebar, callouts, prose tokens
- add chrome i18n (zh/en) and Cmd+K full-text search via build-time index
- split docs into docs/zh and docs/en; English translation for all 16 pages
- hide headless-daemon from sidebar (still reachable via cross-references)
- rename sidebar group 运维/Operations -> 问题/Help
- make sidebar flow inline on narrow viewports and drop internal scrollbar
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
| 17 天前 |
| feat: major architecture overhaul — context management, smart truncation, agent discipline
Phase 1: Tool result externalization
- ToolResultRef + ToolResultStore: large tool outputs cached to disk
- Hot zone inflated from disk, cold zone uses summaries
Phase 2: Turn-aware context windowing
- TurnTracker: tracks conversation turns (user message → completion)
- Budgeted windowing operates at turn granularity
- Cold turns: only user question + assistant conclusion retained
- Synthesized outcomes for turns without assistant text
Phase 3: Smart per-tool truncation
- bash: preserve error lines, skip verbose build logs
- read_file: threshold raised to 2000 lines (prevents multi-read cycles)
- First read of any file forces full content (ignores offset/limit)
- edit_file returns compact diff + file outline after edit
Agent discipline:
- Sleep loop detection (warn at 2, block at 3+)
- Repeated command detection (normalized, warn at 2, block at 3+)
- Same-file multi-edit blocking (threshold 4)
- Re-read soft reminder (not blocking)
- Over-verification detection (3+ consecutive check commands)
- HTTP 200 detection → stop verifying
- 429 rate limit retry with exponential backoff (up to 5 retries)
- Loop detection allows retry after edit (fix-then-retry is not a loop)
- Intercepted/blocked calls don't count toward step limit
- Bash pipe/cat detection with redirect to proper tools
System prompt:
- Restored v1.3.1 proven rules + new additions
- ACT DON'T INSTRUCT / BE CONCISE / ONE SIGNAL IS ENOUGH
- COMMAND DISCIPLINE / ERROR HANDLING sections
- Glob wildcard guidance
- Tech-stack neutral (zero specific tool/language names)
Project context:
- Added pom.xml, build.gradle, requirements.txt, composer.json, CMakeLists.txt to descriptors
- Cross-session context: previous session turns injected into system prompt
- Conversation history loaded on startup (not fresh each session)
- Active services detection: port scanning on startup + URL extraction from bash output
- Frontend hot-reload hint when dev server detected
UX improvements:
- TTFT per LLM call (not cumulative), removed from status bar
- Spinner shows last tool name: "After read_file, thinking..."
- Tool call streaming: shows file_path/command as args stream in
- Turn completion summary: "✻ Completed in N steps, Xs"
- Streaming markdown: trim incomplete backtick/bold tokens
- Pangu spacing (CJK-ASCII)
- Code block styling refresh
- Append input during streaming (Enter sends additional context)
- UTF-8 safe string truncation (char boundary check)
Tooling:
- edit_file: atomic_write retry on rename failure (dev server file locks)
- edit_file: suggested fix on old_string mismatch
- glob: rewritten pattern parser (handles absolute paths, **/ splits)
- LLM request logging to ~/.atomcode/logs/
- Datalog analyzer script (scripts/analyze_datalogs.py) with --deep mode (Claude Code)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
| 2 个月前 |
| chore(release): bump version to v4.24.2
Update workspace crates, install scripts, READMEs, docker docs, and site
docs/badges from v4.24.1 to v4.24.2. latest.json intentionally unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
| 14 小时前 |
| fix: shift+enter handling and optimize website performance
- Fix shift+enter key combination handling in tuix
- Optimize website performance: reduce dependencies, minify assets
| 1 个月前 |
