mirror of
https://github.com/Tencent/teamai-cli.git
synced 2026-10-02 03:14:40 +08:00
Saul Moro
5fbaa377c7
* refactor(usage): extract the locked JSONL store from the usage file (#884, ticket 01) Move the usage file's lock, side-record, fold and rewrite protocol into src/utils/jsonl-store.ts so the recall log can reuse it. The usage file keeps its behavior: same waits, mode, side-file names and gitignore healing. The store adds what the recall log needs: reads that include unfolded side records, owner-only files by default, and a newline before an append when a killed writer left a torn last line. * feat(recall): attribute a Claude direct recall to the doc it opens (#884, ticket 02) recall prints run=<id> on the region start line and appends a run record (session from the environment, each doc's vote key, scope, printed path and eligibility; never the query) to the active scope's recall log. A new PostToolUse handler appends a claim for a Bash call that ran teamai recall, and evidence for a Read under the knowledge roots. votes-sync now credits through the recall-log reducer instead of the transcript parser and no longer needs transcript_path. * feat(recall): exclude the recall subagent's own reads, mark its runs with --caller (#884, ticket 03) Claims and evidence carry the actor (session, agent_id, agent_type). A run is marked as the teamai-recall subagent's by its hidden --caller or its claim's agent_type; reads by that actor never count for it, while reads by the main agent or any other subagent do. The recall agent passes --caller teamai-recall on every recall and drops the unread referenced-doc-ids instruction. * feat(recall): settle each run by a valid claim or an unambiguous env session (#884, ticket 04) A run now records its agent family, via (env or none) and whether the environment held a single session id. The hook records a claim for every run id a shell call printed, noting whether its command ran teamai recall itself (parsed, so a quoted mention does not count); it also reads Codex's string tool_response. The reducer applies only the earliest direct claim for a run in the log, else the env session when unambiguous; unsettled runs never vote. agentSessionIdFromEnv callers are unchanged. * feat(recall): count a shell reader's read of a recalled doc, the Codex path (#884, ticket 05) Add the tool-call classifier (src/utils/tool-call.ts): category, paths read and status per PostToolUse, replacing recordToolCall's Bash and Read branches. The shell tokenizer moves to src/utils/shell-command.ts and now returns each simple command's operator and redirects; invokesRecall and the reader check share it. A read is one reader command (cat, bat, batcat, less, more, head, tail, nl, sed -n printing lines) alone or at the head of a pipeline; any ;, &&, || or & makes the call no read. A failed read is not recorded; a read of unknown status (Codex's string tool_response) counts only when simple. * feat(recall): count a search that shows a recalled doc's lines, never a listing or a count (#884, ticket 06) A search tool in content mode, or a shell grep/rg/ag/ack/git grep, records evidence for each file under the knowledge roots whose lines its output shows: a line that starts with the file's path and ':', under a searched root, or the lone operand when the output is non-empty. The hook only scans the output it already holds; it logs resolved paths, never the command or the output. Listings (Glob, ls, find, fd, tree, rg --files, git ls-files, grep -l/-L, a Grep file list) and counts (-c, --count, count mode) record nothing. OMP's and Cursor's search tools produce no evidence yet. * feat(recall): count Windows reads: PowerShell readers, drive paths, Git Bash /c/ (#884, ticket 07) PowerShell (Claude, CodeBuddy, Copilot) is a shell tool. Get-Content, gc, type and cat read their -Path, -LiteralPath and positional files. Paths are resolved and compared by the new agent-path module, independent of the host: drive-letter case, either separator and Git Bash /c/ name one file, and a search line's drive colon is part of its path. * fix(recall): compare Windows paths without case, look up shell verbs by own key (#884, ticket 07 follow-up) Windows paths ignore case, so a drive-lettered path's comparison key is now lowercased whole, as isUnderRoots did through path.win32.relative before ticket 07. A verb such as `constructor` no longer resolves to an Object.prototype member in the reader and search tables. * feat(recall): count Cursor, Copilot, CodeBuddy, WorkBuddy, Qoder and ZCode reads (#884, ticket 08) Session derivation reads Cursor's conversation_id after session_id and sessionId. The classifier reads Cursor's tool_output JSON string and Copilot's tool_result (text_result_for_llm, result_type), a read's path from file_path, filePath or path, and the names view, ReadFile, bash, Shell and run_in_terminal. Unverified payload shapes are marked in the row names. * feat(recall): OpenCode bridge sends its session, output and status, and links task children to their parent (#884, ticket 09) * feat(recall): Pi and OMP bridges send their session, output and status; OMP subagents their agent id and type (#884, ticket 10) * feat(recall): credit late reads at SubagentStop and pull, and prune the recall log at pull (#884, ticket 11) - Register SubagentStop for Claude Code, Codex, CodeBuddy and Qoder (and their internal builds); votes-sync runs the reducer there, without the end-of-turn summary. - `teamai pull` credits sessions whose evidence is still pending, then prunes the log under its lock: 30 days, then 5000 lines, never dropping pending evidence younger than 24 h or the run, claim, credited reads and links it needs to vote. Evidence and its consumed lines go together. - A run's doc is credited once, so a late read drained after the ledger's 24 h window cannot vote twice. - The store gains appendJsonlBatch (one lock, one write, one side record); a hook call and a reducer pass each make at most one write. Folds now work per line. * refactor(recall): the upvote judge skips docs in the session's ledger, and the transcript parser loses its adoption logic (#884, ticket 12) The opt-in judge (TEAMAI_UPVOTE_JUDGE) excluded the parser's adoptedDocIds on top of the session's upvote ledger. Since votes-sync credits from the recall log, a doc the parser counted but the hook path did not (a Glob listing, say) was neither credited nor judged. The judge now excludes only the ledger, and the parser keeps its recalled-doc detection, region parsing and recalled-doc-ids comment for the judge and older-format sessions. * fix(recall): the upvote judge credits the turn before reading the ledger (#884, ticket 12 follow-up) The dispatcher starts the detached judge before the foreground votes-sync credits the turn, so a doc opened in that turn was sent to the judge CLI once. The judge now runs the same reducer pass first; the ledger and the consumed marks make a second pass a no-op. * feat(stats): a recall section lists the last 10 sessions' runs, recalled and adopted docs (#884, ticket 13) * fix(stats): name the agent whose hook claimed a recall run (#884, ticket 13 follow-up) A claim now records the dispatch tool id of the hook that sent it, and the recall section prefers it over the run's environment family. A Codex run started from a Claude shell, or an OMP run, showed '-' before. * docs(recall): one adoption section with the per-agent table and known limits (#884, ticket 14) The usage guides gather adoption, the recall log, run ids, run ownership, subagents, vote timing and the stats section under "Recall adoption and upvotes", with a direct-recall / subagent-path table per agent and the known limits. The core skill's troubleshooting reference says where a recalled doc gets its upvote, and the recall subagent no longer says the Stop hook parses its doc-id comment for adoption. * docs(skill): CodeBuddy and WorkBuddy hooks are installed, not skipped (#884) The core skill's hooks table said CodeBuddy and WorkBuddy hooks are skipped by design, which contradicts the settings.json injection in src/hooks.ts and the per-agent adoption list this PR adds to the same file. * test(hooks): Claude settings carry the SubagentStop hook (#884, ticket 11 follow-up) Ticket 11 registers votes-sync on SubagentStop for Claude; the hooks e2e still expected four events. Cursor keeps four. * fix(recall): a search line counts only from a file-name prefix, and the log keeps only .md paths (#884, review finding 1) A one-file search's lines carry no path, so text before a colon (Cause:, a -n false Grep line, Pi's basename) no longer becomes <doc>/<text> in the recall log. Other lines need a prefix ending in a file name, then :<line>:, or : in grep/rg path:text and OpenCode's header. The recorder drops any evidence path that is not .md. * fix(recall): a one-file search counts only when it shows a line, not a tool's no-match text (#884, review finding 2) OpenCode's No files found and Pi's and CodeBuddy's No matches found made a grep of the doc that found nothing credit it. The target now needs a line that is neither blank nor a search tool's status line. Shell searches print none, so their lines are never taken for one. * fix(recall): the upvote judge keys each doc by the key its run recorded, not the parser's basename (#884, review finding 3) The transcript parser names learnings/setup and a skill's SKILL.md by their basename, so a doc the hook path had credited was judged again and upvoted under another doc's key. recalledKeyOf maps the printed path to the run's key for the ledger check and the upvote; a path no run printed keeps the parser's id. * fix(recall): SubagentStop credits locally and leaves the vote push to the next Stop (#884, review finding 4) The main agent waits on SubagentStop, so the git round trip of the reports push (and the ledger prune) blocked it mid-turn, up to the handler budget, after every subagent while deltas were pending. Stop and pull push what it credited. Stop is unchanged. * docs(hooks): state only what is known about Codex and an unknown hook event key (#884, review finding 6) Codex 0.159 knows SubagentStop, and Codex main's HookEventsToml has no deny_unknown_fields. How older hook-capable builds parse an unknown key is unverified. * docs(skill): restore the Claude hooks row to its main wording (#884, review finding 7) The CodeBuddy/WorkBuddy row keeps its fix, which the adoption section below needs; the Claude row's rewording was unrelated to #884. * refactor(recall): the adoption window and retention constants are module-private (#884, review finding 8) Nothing imports them, tests included. * fix(recall): an unquoted # that starts a word begins a shell comment (#884, PR review finding 1) `cat other.md # /kb/doc.md` reads only other.md, yet the doc counted as an operand. The comment runs to the end of the line; `a#b` and a quoted `#` stay literal. A claim with a trailing comment still claims. * fix(recall): an unquoted backslash escapes a space or shell character, and keeps a Windows path whole (#884, PR review finding 2) `cat /kb/my\ doc.md` was split into two operands. Outside quotes a backslash now escapes whitespace and ' " \ $ ` ( ) & ; | < > * ? # !; before any other character it stays literal, so an unquoted `C:\kb\learnings\x.md` is unchanged. Double quotes already followed POSIX for \" and \\, and single quotes stay literal. * fix(recall): a claim skips the root program's options before recall (#884, PR review finding 3) `teamai -v recall "q"` is a valid recall, but its claim was marked indirect, so a run with no session variable (OMP, via: none) never settled. The root options now live in one table that index.ts registers on the program and the claim parser reads through Commander's Option, so it also skips the value of an option that takes one. `npx teamai-cli[@v]` is handled as before. * fix(recall): with no status, a shell call that printed only its command's errors failed (#884, PR review finding 4) Codex sends no status, so `grep needle /kb/doc.md` answering "grep: /kb/doc.md: Permission denied" counted as the doc's content and upvoted it; `cat doc` answering "cat: doc: No such file or directory" did the same. When the status is unknown, output lines that start with the command word (as written or by name) and ": " are errors, not content. A read or search that printed nothing else is a failure and records no evidence. Both usage guides say so. * fix(recall): Copilot's SessionEnd credits and pushes the session's votes (#884, PR review round 2 item 1) Copilot's last turn can end with SessionEnd and no Stop, so votes-sync now runs on session-end too: foreground, git only, with Stop's timeout. It pushes as Stop does and prints no adopted summary. The Copilot rows end with the real SessionEnd payload instead of a synthetic stop. * fix(recall): a search with no hits prints its run id, so its shell call can claim it (#884, PR review round 2 item 2) The no-hit line now ends in run=<id>, and the claim pattern reads it after the logger's glyph. An OMP recall with no hits and no session variable settles through its bash call's claim and shows in teamai stats. --check still prints no run id and records no run. * fix(recall): type and gc read only in PowerShell or for a Windows path, and a shell's own diagnostic is an error line (#884, PR review round 2 item 3) In bash, type is a builtin that prints what a name is, so Codex's type /posix/doc.md under Linux gave a false upvote with no status. type and gc now count under a PowerShell tool, or when every file is a Windows path. Get-Content counts everywhere. With no status, bash:, sh: and zsh: diagnostics are error lines too. * fix(recall): with no status, a read drops a file its command's error names (#884, PR review round 2 item 4) In cat a.md doc.md where only doc.md errors, doc.md still counted. An error line <verb>: <file>: … now removes that file from the read's paths, matched as written and as resolved against the cwd; the other files still count.
TeamAI — Make Every Team AI Native
English | 中文 | 日本語 | 한국어 | ไทย
The shared foundation for how your team works, learns, and improves with AI.
TeamAI turns individual AI capabilities into shared team capabilities — across agents, machines, and team members.
Why TeamAI
Quick Start
Send this one line to your AI tool:
Install the teamai skill: https://github.com/Tencent/teamai-cli/tree/main/skills/teamai , load the teamai skill, then set up TeamAI for my team from scratch.
Once TeamAI is set up, just talk to the /teamai skill in your AI tool:
Set up a team from scratch
/teamai Help me set up TeamAI for my team from scratch
Join a team
/teamai Help me join my team's TeamAI, repo URL is https://github.com/your-org/your-repo
Share with the team
Skills, rules, MCP servers, and other agent resources can all be shared:
/teamai Share my xxx skill with the team
Open the dashboard
/teamai Open the TeamAI dashboard
Once a teammate is set up, they just open their agent and already have the team's full set of AI assets.
Command-line install
Install
npm install -g teamai-cli
Team admin / solo user
Create a shared-experience repo on your git host (GitHub, GitLab, GitCode, CNB, TGit, or a private Git service), grant write access to team members, then run teamai init https://github.com/your-org/your-repo.
No team repo yet? Start from a template pre-loaded with production-ready skills, rules, and review agents. Browse the teamai-hub org, click Fork, then
teamai initagainst your new repo.
Team members
# Choose one, depending on where you want resources installed
# Project-scope init (default, resources installed under the project directory)
cd /path/to/my-project
teamai init https://github.com/your-org/your-repo
# Or, user-scope init (resources installed under ~/)
teamai init https://github.com/your-org/your-repo --scope user
Once initialized, every AI session automatically pulls the latest skills / rules and other Harness updates published by admins — no manual sync needed.
Product Overview
Three layers of capability, built on Git:
- Team Execution — make every agent work the team's way: skills, rules, docs, env, agents, hooks, MCP, models.
- Team Context (beta) — make every agent understand the team: learnings, codebase graph, teamwiki.
- Team Improvement (beta) — make every execution improve the team: usage, sessions, dashboard.
| Agent | Team Execution | Team Context (beta) | Team Improvement (beta) | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| skills | rules | docs | env | agents | hooks | mcp | models | learnings | codebase | teamwiki | usage | sessions | dashboard | |
| Claude Code | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Codex | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Cursor | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| GitHub Copilot CLI | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| CodeBuddy | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| WorkBuddy | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| OpenCode | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | — | — | — |
| Pi Coding Agent | ✓ | ✓ | ✓ | — | — | ✓ | — | ✓ | — | — | — | — | — | — |
| OpenClaw | ✓ | ✓ | ✓ | ✓ | — | — | — | — | ✓ | ✓ | ✓ | — | — | — |
| Hermes | ✓ | — | ✓ | ✓ | — | — | — | — | ✓ | ✓ | ✓ | — | — | — |
| DeepSeek Harness | ✓ | — | ✓ | — | — | — | — | — | ✓ | ✓ | ✓ | — | — | — |
| Qoder | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Qoder CN | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Kiro | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| ZCode | ✓ | — | ✓ | — | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Oh My Pi | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | — | — | — |
Learn More
- Usage Guide (中文版) — setup, onboarding, daily workflows, and commands
- Product Overview (中文版) — architecture, distribution controls, and capability details
- Git Providers — supported repository providers
- Windows Setup (中文版) — hooks and shell configuration
- Technical Designs — design documents and proposals
Contributors
Thanks to everyone who has contributed to TeamAI!
Made with contrib.rocks.
Contributing
Join the conversation, or open an issue or PR. See CONTRIBUTING.md for how to contribute.