fix(stats): count every worktree of a repo as that repo (#809)

[SaulMoro]

Summary

stats --by-repo, the dashboard, session save and the wiki slug treated each git worktree as its own project, because events recorded only their cwd. Events now record the repo's main checkout, and every consumer attributes sessions to it.

 hook-dispatch handler (and legacy `teamai dashboard-report`)
   resolveConfigForDir(cwd) -> resolveAnchors(cwd)       # now memoized per dir
   event.dataHome = getDataHome(config)
+  event.projectAnchor = eventProjectAnchor(event.cwd)   # unset: no cwd (Copilot), no git, cwd gone
   appendEvent

+repoKeys(events)    session -> last projectAnchor it recorded ?? last cwd
+repoName(anchor)    directory name; a bare git dir (repo/.bare, repo.git) -> repo
+repoLabel(key, all) repoName, or parent/name on a clash between path keys (work/api)

Case	Before	After
stats --by-repo in the repo	my-repo 2 / wt-demo 2 / src 1 sess	my-repo 5 sess
two unrelated repos named api	api 2 sess	work/api 1 / personal/api 1
session save (worktree)	Project: `/…/wt-demo`	Project: `my-repo` + Directory:
import --dir . / codebase --extract in the worktree	slug wt-demo	slug my-repo
bare repo repo/.bare + main, feature	slugs main, feature; rows main, feature	slug and row repo
dashboard after git worktree remove	"Unassigned sessions"	my-repo, listed once in the filter

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature causing existing behavior to change)
• Documentation only
• Refactor / internal cleanup

Test Plan

• npx tsc --noEmit passes
• npx vitest run passes: 307 files, 4737 passed, 1 skipped
• Added/updated tests: e2e repo-attribution-worktrees-809.test.ts; unit cases in anchors, session-analytics, dashboard-workspaces, dashboard-report-scope, dashboard-ui
• npm run test:e2e passes: 48 files (3 skipped), 252 passed, 26 skipped (on 69e97889; later: labels/docs only, [bug] stats, dashboard, session save and wiki slugs split one repo into one project per worktree #809 e2e 6/6)

The e2e test drives the built CLI (sandbox HOME, local team repo, a worktree outside the checkout, hook payloads on stdin):

setup  sessions in repo, repo/src, worktree (both writers), HOME then repo, two unrelated repos named api
1. stats --by-repo in the repo           -> [[my-repo, 5]]
2. stats --by-repo in HOME               -> [work/api, 1], [personal/api, 1]
3. session save --session-id wt-1        -> Project: `my-repo`, Directory: `<worktree>`
4. import --dir . --dry-run; --extract   -> (project: my-repo); evidence/code/my-repo
5. git worktree remove; teamai dashboard -> my-repo holds all 5 sessions; filter options ["my-repo"]

On origin/main all five fail, e.g. expected [ [ 'my-repo', 2 ], …(2) ] to deeply equal [ [ 'my-repo', 5 ] ]; with this PR all pass.

Removing any one of 30 parts (writers, memo, filter, slug, repoName, remote-key rules) and rebuilding fails a test.

Reproduced with the real CLI, git mode: worktree

$ teamai pull --force        # in my-repo: legacy .teamai/ moves into its partition
$ git worktree add ../wt-demo -b wt-demo
$ teamai stats --by-repo     # in my-repo
  my-repo  5 sess, 5 turns, 5 tools
$ teamai session save --session-id wt-1
- Project: `my-repo`
- Directory: `…/wt-demo`
$ teamai import --dir . --dry-run   # in wt-demo
ℹ Scanning local directory: …/wt-demo (project: my-repo)
$ git worktree remove ../wt-demo; teamai dashboard   # /api/workspaces, /api/sessions
my-repo (project): cd-1, main-1, main-2, main-3, wt-1
  Repository filter options: ["my-repo"]
Unassigned sessions: -

Reproduced with the real CLI: bare repo (repo/.bare + repo/main + repo/feature)

$ teamai stats --by-repo     # user scope, one session per worktree
  repo  2 sess, 2 turns, 2 tools
$ teamai import --dir . --dry-run   # in repo/main, then repo/feature
ℹ Scanning local directory: …/repo/main (project: repo)
ℹ Scanning local directory: …/repo/feature (project: repo)
$ teamai codebase --extract .       # in repo/feature; ls teamwiki/evidence/code
repo
$ teamai session save --session-id s-feature
- Project: `repo`

From reading the code, not reproduced with the real CLI:

• Copilot events get no anchor (unit test through the real dispatcher).
• filterEventsByScope's legacy branch (src/team-push.ts:415) is unchanged.
• Remote-form keys: unit-tested only.

Not run: gitlab/github providers and other agents; attribution depends on neither (left to CI).

Related Issues

Closes #809

Notes for Reviewers

• Two writers. Current installs write through the hook-dispatch handler (src/hook-handlers.ts:189), not only dashboardReport. Both call eventProjectAnchor, which skips a missing cwd without git, so [bug] hook events after a worktree is deleted fall back to the user scope #810 changes one place. process_exit copies the anchor, like dataHome.
• Memo. resolveAnchors keeps non-null results for the process's life, so the hook's second lookup runs no git inside a repo. null is not cached, per the issue. A long-lived dashboard keeps a removed worktree's answer, which attribution wants.
• Wiki slug, narrower than the issue's rule. Only a linked worktree's root takes the repo's name; subdirectories keep theirs (--extract packages/api stays api). repoName detects a bare git dir on disk (HEAD and objects/, no .git). Gap: a main checkout reached through a differently named symlink keeps that name.
• Labels. Rows group by label, so no_repo dirs still share a row; a repo keeps its name even in a dir like ~/workspace. Only path keys get a parent prefix, then the whole path on a second collision; remote keys keep owner/repo, one row across hosts. session save labels against the whole log.
• Dashboard discovery reads event anchors; git runs only for events lacking one or whose anchor has no config.
• Reversible. No migration; the field is optional and events without it fall back to cwd.
• Follow-ups. Remote key github.com/acme/api and local /x/acme/api both label acme/api (src/utils/repo-attribution.ts:138-141); hooks never write remote keys. Unrelated flake: local-agent.test.ts > emits hint only once per sessionId shares an os.tmpdir() marker keyed by a fixed session id (src/local-agent.ts:1191).
• Not in this PR, per the issue: backfill of old events, the legacy filterEventsByScope branch, already-pushed wiki entries, [bug] hook events after a worktree is deleted fall back to the user scope #810, a per-worktree breakdown.
• Docs: data-directory-layout.md (projectAnchor, plus the fix(report): each scope keeps its own reported dashboard snapshots (#786) #795 scope note), dashboard-unified{,.zh-CN}.md and usage-guide{,.zh-CN}.md (repository filter, slug), --project help and regenerated commands.md.

[github-actions]

• [P2 non-blocking] src/codebase-extract.ts:548 — Bare-repository worktrees are explicitly excluded from anchor-based slugging. For a bare repo with worktrees main and feature, running codebase --extract or import --dir from each still writes separate main and feature evidence trees, so the wiki consumer does not satisfy the PR’s worktree-unification goal.
• [P3 nit] src/utils/repo-attribution.ts:95 — Collision handling breaks the existing remote canonicalization contract. Events keyed as github.com/Tencent/teamai-cli and cnb.cool/Tencent/teamai-cli both canonicalize to Tencent/teamai-cli, but repoLabel() detects that collision and returns the raw host-qualified keys, producing two stats rows for the same repository.

The PR description includes sufficient real-CLI end-to-end verification for this runtime change.

[SaulMoro]

Re: #813 (comment)

Thanks, both confirmed and fixed in 69e9788.

1. [P2] bare-repo worktrees and the slug. Fixed. A new repoName(anchor) (src/utils/repo-attribution.ts) is shared by the slug and repoLabel. It detects a bare git dir on disk (HEAD and objects/, no .git) and names it after its repo: repo.git → repo, and repo/.bare or repo/.git → repo. In repo/.bare + repo/main + repo/feature, both worktrees now write teamwiki/evidence/code/repo/, and stats shows a single repo row. That also fixes the .bare label noted as a follow-up, and a session in the bare layout's own directory (repo/) counts as the same repo. Only a worktree's root is renamed; subdirectories keep their own name. Covered by unit tests (anchors.test.ts, session-analytics.test.ts) and a real-CLI run in the PR body.

2. [P3] remote-form keys split by the collision handling. Fixed. Only absolute-path keys take part in collision handling now. github.com/Tencent/teamai-cli and cnb.cool/Tencent/teamai-cli both keep Tencent/teamai-cli and merge into one row again. Two local repos that share a basename still split into work/api and personal/api. Covered by a unit test on repoLabel and attributeByRepo.

[github-actions]

• [P1 blocking] src/dashboard/client.ts:37 — The filter now groups by repository rather than working directory, but docs/usage-guide.md:1447 and docs/usage-guide.zh-CN.md:1386 still document a working-directory filter. This violates the repository rule requiring every affected bilingual document to be updated for behavior changes.
• [P2 non-blocking] src/utils/repo-attribution.ts:128 — Absolute projectAnchor paths are passed through attributeRepo() before repoName(). A valid Git repository named workspace, home, data, etc. is therefore labeled no_repo; all its worktrees disappear into the unassigned stats row and session save records Project: no_repo.
• [P3 nit] src/utils/repo-attribution.ts:132 — Remote keys bypass collision handling. With keys such as github.com/acme/api, /x/acme/api, and /y/other/api, the first local path and remote key both receive acme/api, causing attributeByRepo() to merge unrelated sessions.

The two earlier findings are resolved in the current diff: bare-repository worktrees now use the repository name, and equivalent remote-host keys canonicalize to one row. The PR description includes sufficient real-CLI end-to-end verification.

[SaulMoro]

Re: #813 (comment)

Fixed in 8c72294:

1. [P1] docs for the Repository filter. Fixed. docs/usage-guide.md and .zh-CN.md now say the Execution page filters by repository, with every worktree of a repo as one entry. docs/designs/dashboard-unified{,.zh-CN}.md had the same stale "working-directory filter" wording and is fixed too.
2. [P2] repos in a directory named like workspace. Fixed. A key that is itself a repository (a checkout with .git, or a bare git dir) keeps its name and skips attributeRepo's non-project words. So a repo at ~/workspace and its worktrees are one workspace row, and session save prints Project: workspace. A plain directory with that name is still no_repo. Covered by a unit test for workspace, home and data.
3. [P3] remote key vs a colliding local path. Left as a follow-up in the PR notes (src/utils/repo-attribution.ts:138-141). It needs remote-form keys, and hooks only record filesystem paths.

[github-actions]

• [P3 nit] src/codebase-extract.ts:548 — A main checkout accessed through a differently named symlink keeps the symlink basename, while its linked worktrees use the real main-checkout name. For example, extracting /tmp/alias -> /repos/my-repo creates alias, but extracting a linked worktree creates my-repo, splitting one repository’s evidence.
• [P3 nit] src/utils/repo-attribution.ts:138 — Remote keys still bypass collision handling. With github.com/acme/api, /x/acme/api, and /y/other/api, the remote and first local repository both receive acme/api, so stats --by-repo merges unrelated sessions.

The earlier bare-repository, reserved-name, remote-host canonicalization, and bilingual documentation findings are resolved. The PR description includes sufficient real-CLI end-to-end verification.

[SaulMoro]

Follow-ups from this PR's Notes and its last review pass are tracked in #823 (Cleanup: the symlinked main-checkout slug at codebase-extract.ts:548, the remote key vs local path label collision at repo-attribution.ts:139-145, and the flaky local-agent.test.ts).