chore: version packages

[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to master, this PR will be updated.

Releases

agentsmesh@0.41.0

Minor Changes

• 3bbf68b: Lessons are harder to lose and easier to use correctly, after a full manual QA pass.

  Team merges: if git could not start the lessons merge driver, it kept only your side of lessons.json with no conflict markers, and every check still passed. Now lessons validate, check and generate --check fail while the file is missing the other branch's lessons, and tell you to run agentsmesh lessons resolve before git add. The driver is also only saved when git can start it later (npx caches and package-script bin folders on PATH no longer count, and the npx form needs agentsmesh at the repository root or installed globally). lessons resolve now works again after you fix a broken side by hand in the file.

  The recall hook never breaks your tool: a log file it cannot write, or a bad line in a log, no longer makes it fail or lose a lesson for the session. A failed read (for example reading a file that does not exist yet) is no longer counted against a later write. A multi-file patch shows each warned rule once and stays under the size cap. Invisible and look-alike characters can no longer fake the end of the recalled-lessons block. An unreadable graph is now also reported when a session starts with no prompt. The hook does nothing in a folder with no lessons project, such as your home folder.

  The lessons CLI works from any subfolder of a project, like git. It refuses an unquoted multi-word rule (instead of saving only its first word) and a single-value flag given twice. Re-adding a lesson says what changed. A regex the linear engine cannot run is dropped with a DEAD_COMMAND_PATTERN warning when the lesson has another trigger, or refused as UNRECALLABLE_LESSON (exit 2) when it is the only one. Every lessons command, and every MCP lessons tool, now explains an unreadable graph (merge conflict, bad JSON, SCHEMA_INVALID, newer version) instead of printing parser output. A negated glob (!path) is flagged as broad. A lessons write that lost its lock (the process was paused past the 60 s stale window) refuses to save instead of erasing a later write, and a lock dated in the future no longer blocks writers.

  MCP lessons tools never create a lessons graph in your home folder or outside a project: they use the nearest lessons folder, then agentsmesh.yaml, then the git repository root, and otherwise refuse writes with NO_PROJECT. lessons_show accepts a lesson id, lessons_query with always honors session and no_dedup, and refused writes return VALIDATION_FAILED with the finding codes.

  Installed pack hooks are no longer lost: two packs on the same event are combined, a pack adds to an extend's hooks for that event, and a local hooks.yaml event keeps the pack's hooks after its own. Before, agentsmesh init --lessons (which adds the recall hook locally) silently removed installed pack hooks for the same event. Extends layers still override per event, as documented. To drop a pack hook, uninstall the pack or leave hooks out of its features.

• 2cb8b48: The MCP server now tells every client about lessons when it connects, and says something true in both states.

  Where a project has lessons, the server hands over the same recall and capture contract that init --lessons writes into your instruction file. That matters for a plugin: a plugin can ship skills and servers but never your instruction file, so a plugin-only install previously had no standing instruction at all and recall depended on the agent opening the skill first. The obligations are phrased in tool names rather than shell commands, since a client reaching the server this way may have no shell.

  Where a project has no lessons, the server says so plainly and explains how to start one. It does not name a graph file you do not have, does not point at a skill you never installed, and mandates nothing. This matters because the server also carries the configuration tools and is documented on its own: most people who wire it up never opted into lessons, and they should not be told to query a memory that cannot answer.

• cea0c7e: Lessons now hold up for teams, for more tools, and against a graph you did not write yourself.

  Teams: two branches that both capture lessons now merge cleanly. The git merge driver merges each lesson field by field, and every clone gets the driver the next time it runs agentsmesh generate (or init --lessons), so it no longer depends on one person's setup. If a merge still leaves conflict markers, the new agentsmesh lessons resolve combines both sides. agentsmesh check and generate --check now fail when lessons.json cannot be read (a merge conflict, a corrupt file, a newer schema), instead of passing while recall is silently off. The lessons lock no longer loses a lesson when many agents write at once. A fresh clone no longer prunes lessons whose files are merely not created yet: a glob counts as dead only when git history shows its file was deleted or renamed. Generated recall hooks run npx --no --offline agentsmesh when the project depends on agentsmesh, so teammates without a global install still get recall.

  More tools: the recall hook now answers Gemini CLI's BeforeAgent, Cursor's sessionStart and postToolUseFailure, and GitHub Copilot's sessionStart and postToolUseFailure. Codex apply_patch edits get recall for the files they touch, subagents get their own session dedup, and the hook and MCP server find the lessons project from a subdirectory. Target descriptors (including plugins) can declare hookContextEvents so recall is only wired to events whose output reaches the model.

  Capture and effectiveness: failures are read from the field Claude Code actually sends, keyed on the command that really ran, and user interrupts are no longer counted. A --trigger-file path outside the project is rejected instead of being stored where it could never fire. The outcome log is on by default (turn it off with "outcomeLog": false in .agentsmesh/lessons/config.json or AGENTSMESH_LESSONS_OUTCOME_LOG=0), and a lesson counts as missed only when the same action fails again in the same session within 30 minutes, so validate and stats stop flagging lessons that work.

  Safety: recalled rules are delivered one per line inside a <recalled-lessons> block, with ids, and every CLI and MCP answer is size-capped, so a hostile rule cannot pose as a system message or flood the context. file_glob triggers use a linear-time matcher over a safe glob subset; other glob syntax is rejected with UNSAFE_GLOB_PATTERN and never matches, so a crafted glob can no longer stall recall. recallLimit and recallMaxTokens in config.json are capped at 50 and 8000, lessons query --always has the same budget as the hook, and legacy migration refuses index paths outside .agentsmesh/lessons/.

• 98cb625: Upgrade notes. Most projects need to do nothing, but check these after you upgrade:

  • Run agentsmesh generate once. Hooks for Gemini CLI, GitHub Copilot, Cursor and Windsurf, lessons recall hooks, and hooks from installed packs can be generated differently now, so agentsmesh generate --check reports drift until you regenerate.
  • Run agentsmesh lessons validate. Some lesson file triggers no longer match anything, and lessons validate and agentsmesh lint now report them as UNSAFE_GLOB_PATTERN errors: ** inside a name (src/**.ts, write src/**/*.ts), extglobs (+(a|b), write {a,b}) and ranges ({1..3}, write {1,2,3}).
  • Gemini CLI hooks match exact tool names now. Bash, Edit, Write and Read reach Gemini's own tools (before, they matched nothing there), but a partial Gemini name such as shell no longer matches: use the full name (run_shell_command) or the Claude Code name (Bash). A UserPromptSubmit hook now also runs on Gemini, as BeforeAgent.
  • SessionStart and PostToolUseFailure hooks now also reach GitHub Copilot, and PostToolUseFailure hooks reach Cursor.
  • The lessons outcome log (.agentsmesh/lessons/outcome-log.jsonl, local and gitignored) is on by default, and turning telemetry off no longer stops it. Turn it off with "outcomeLog": false in .agentsmesh/lessons/config.json or AGENTSMESH_LESSONS_OUTCOME_LOG=0.
  • agentsmesh lessons exits 2 for a flag with an empty value, so a script that runs --cmd "$CMD" fails when CMD is empty.
  • Installing the same whole source again rebuilds its pack, so files you added by hand inside .agentsmesh/packs/<name>/ are removed, as refresh already did. Keep your own changes in .agentsmesh/, outside packs/.

Patch Changes

• 579a6c5: agentsmesh check now gives the fix that matches the drift it found, instead of always saying "Run 'agentsmesh merge' to resolve, or 'agentsmesh generate --force' to accept current state." For canonical or generated-output drift (including the stale hashes a merge can leave) it points to plain agentsmesh generate; only changed locked features (collaboration.strategy: lock) get the generate --force advice. A .agentsmesh/.lock with git conflict markers is now reported as a lock conflict with the agentsmesh merge fix (and lockConflict: true in --json, in the MCP check tool result, and in the LockSyncReport returned by the programmatic check()), instead of "Not initialized for collaboration".

• 499f7c7: The agentsmesh lessons commands handle bad input clearly instead of quietly doing the wrong thing.

  A flag that needs a value but gets none now fails with exit 2 and "--flag needs a value". Before, it was ignored: deprecate X --superseded-by deprecated without the supersede link, and query --session skipped dedup. A value that starts with -- can be passed as --flag=value (for example --rule="--no-verify is forbidden"), and the error for an unknown "flag" that is really text says so. prune --cap abc no longer falls back to the default cap. lessons help [subcommand] works.

  lessons add uses exit 2 for every input error, with messages that name the flag: --scope other than always, a topic id that is not kebab-case (it suggests one), --new-topic without a non-blank --topic-summary, and an unsafe --trigger-file glob, which is now refused before any trigger id is made. A change the graph validator refuses also exits 2, and the message is plain: Refused to save the lessons graph: … Nothing was written. Internal names such as mutateLessonsGraph: no longer show in messages. Repeated --evidence refs are stored once. The 2000-character rule limit counts characters, so an emoji counts once.

  --trigger-file is stored in one form: surrounding spaces and a leading ./ are dropped, so ./src/a.ts reuses the src/a.ts trigger. A relative path that climbs out of the project (../x.ts), the project root, or an existing folder is rejected with a clear reason; for a folder it suggests folder/**. The same applies to the MCP lessons_add tool. A --trigger-cmd with a \u{…} escape is rejected, as the docs said.

  query says when a legacy index.yaml store could not be migrated, instead of printing only "(no matches)". import-md --migrated-at must be a real date. AGENTSMESH_LESSONS_TELEMETRY and AGENTSMESH_LESSONS_OUTCOME_LOG also accept true/false, yes/no and on/off. lessons query warns when config.json has a switch that is not true/false (such as "outcomeLog": "no") or is not a JSON object.

  Smaller fixes: show <lesson> includes the rationale, journal marks deprecated and superseded lessons, validate --json names the error codes, messages end with a period, the --ids help text says what it does, and the docs keyword example uses one --trigger-kw per keyword.

• b2bfc56: The lessons recall hook, the lessons.json merge driver and the lessons logs handle more edge cases correctly.

  Recall hook: it reads a payload over 1 MB to the end before ignoring it, so the host no longer gets a broken pipe, and it reads JSON that starts with a UTF-8 BOM. An event name it does not know (such as Gemini's BeforeTool) now does nothing instead of answering as PostToolUse; a payload with no event name is still a tool call. Copilot's VS Code compatible SessionStart (with initial_prompt) gets task recall. The project now comes from the touched file first, so in a monorepo a package's own lessons apply to its files, and a cwd outside the project no longer hides it. A decomposed (NFD) path matches its glob. On a prompt, the note about hidden matches is back: at most 5 keyword rules are shown, the always-on lessons keep their own budget, and the hook says what either cap left out. The recurrence warning counts only failures from the last 24 hours (it used to count every failure ever, "failed 833×"). A Cursor permission_denied is not recorded as a failure, a failed command's nudge no longer suggests a file glob, and recalls running at the same time no longer lose each other's session dedup entries.

  Logs and files: an append after a line cut off mid-write starts a new line instead of gluing two records, the logs are capped by size (a single huge line is dropped) and readers read only the newest part, a read-only lessons.json is not written over and a save keeps its file mode, a .lessons.lock that is a file gets a clear message instead of ENOTDIR, a UTF-8 BOM in lessons.json or config.json is ignored, a writer waiting on a busy lock says once who holds it, and old temp files and stale lock folders are cleaned up.

  Merge: a trigger or topic deleted on one branch stays deleted when the other branch did not change it and no lesson uses it. lessons resolve refuses to save a result with errors when it had to rebuild the sides from conflict markers (and keeps the markers), warns when the markers have no merge base, counts lessons after same-id renames, and names the right next step for a merge, rebase, cherry-pick or revert (none outside git). The merge driver falls back to the bare command when npx is missing but agentsmesh is on PATH. init and generate ask you to re-run init --lessons when the recall hook command in hooks.yaml no longer matches the project's agentsmesh dependency, and generate no longer repeats "Kept your own lessons.json merge driver" on every run.

• e2f4219: Installing a changed pack again now updates it in place. Before, if the source dropped a folder (say commands/), running agentsmesh install again with the same --name failed with "Auto-generated pack name … collides", and without --name it added a second pack from the same source that kept the removed files. Now the install finds the pack from the same source, --target and --as (by --name, by feature set, or the one pack that covers the whole source). When both cover the whole source, it replaces the pack's contents like refresh, so files removed at the source go away. A picked subset still merges, and packs split by --path, --as or pick stay separate. A re-install without --name also keeps the pack's name (it used to rename a local pack to an auto-generated name), and --dry-run shows the pack it would update. A --name that belongs to a pack from another source now fails with a clear message that names the pack.

  Skill folders are no longer dropped because of their name. A source with rules/ next to skills/my_skill/ or skills/S1/ used to look like a lone rules folder, so its skills, README and LICENSE were skipped without a word. Any skill folder name now counts, except names that start with . or _.

  mcp.json, hooks.yaml, permissions.yaml and ignore at the root of a source without .agentsmesh/ are still not installed (settings install only from a source's .agentsmesh/ folder), but install now prints one warning that names each skipped file.

• 7857ff1: The MCP check and generate tools now tell the same story as the CLI. check runs the same check as agentsmesh check, so an unreadable .agentsmesh/lessons/lessons.json (a merge conflict, bad JSON, a schema error or a newer version) is now reported in a new lessonsGraphError field, with the same text as the CLI JSON error, instead of looking like a clean result. generate sets lockfileUpdated only when the run really rewrote .agentsmesh/.lock; a run that changes nothing leaves the lock alone and now says false.

• b7ab200: Internal-only: removed dead code and collapsed duplicated generators. Two exported helpers had no caller outside their own tests since the commit that introduced them, and one install helper was fully superseded by a more general sibling. Fifteen targets each carried the same four-line ignore generator differing only in a path constant, and the pack writer and merger each repeated the same copy-into-subdirectory loop three times; both now call one shared helper. No behaviour changes: every generated artifact is byte-identical.

• ab63a79: agentsmesh generate no longer rewrites .agentsmesh/.lock when nothing changed. Before, a run that printed "Nothing changed" still wrote a new generated_at, so the git tree was dirty after every generate and each teammate's run showed a lock diff. Now the lock is rewritten only when its checksums, extends, packs or outputs change; otherwise the file stays byte-for-byte the same. generated_at, generated_by and lib_version now describe the last run that changed the lock. check, generate --check and watch work as before.

• cab70b6: On Windows, agentsmesh now copes when another process is removing or reading the same lock folder or file at that moment. A command waiting for a busy lock (for example generate, install or a lessons write) no longer crashes with EPERM, and lessons recalls running at the same time no longer lose a session dedup entry. Short EPERM, EACCES and EBUSY errors are retried for a moment, as renames already were; an error that does not clear still stops the command.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.