fix(recall): search nothing in a project whose config cannot be read (#796)

[SaulMoro]

Fixes #796
Part of #752

#792 is merged (1fd400e); rebased onto origin/main 1fd400e, head 22824a1.

Summary

 teamai recall [--check] "<query>"   (project whose config exists but cannot be read)
+  no query and no --check → "Usage: teamai recall <query>"   (before any detection, as on main)
   detectProjectConfig(cwd, onUnreadable)
-    broken partition skipped → legacy .teamai/ (maybe another team) or user scope
-    search → print results / RELEVANT|NOT_RELEVANT → record recalled counts + quality
+    unreadable → "Nothing was searched: <file>: <reason>. Fix the file, or move it
+                  aside and run `teamai init` to write a new one."   exit 1
+  empty --check → NOT_RELEVANT   (only after detection, so it cannot answer there)

 agents/teamai-recall.md  Step 0 / Step 3
+  output contains "Nothing was searched:" → return that line verbatim, stop,
+    and tell the main conversation to act on it only with the user's consent
 src/builtin-rules.ts  Bash method (tools that run `teamai recall` directly)
+  same instruction

No silent mode: recall has no --silent flag and no hook runs it. A hook that runs recall later should follow pull --silent (#784): print nothing, write the line to ~/.teamai/debug.log, exit 1. skill-data troubleshooting gets one line next to #792's Nothing was synced entry.

Evidence

Real CLI at 58195ab vs the stack base b119ec5 (#792's head). Per provider: team repos are local bare remotes behind an https:// URL, with provider: in teamai.yaml set to git / github / gitlab; teamai init --agent claude,codex,codebuddy,opencode + pull; then the partition config.yaml is broken and another team's legacy .teamai/ put behind it. Agent axis: the recall agent and rule each tool gets deployed, and teamai hook-dispatch … --tool <agent> payloads with the project as cwd.

Case (each of git, github, gitlab)	b119ec5	58195ab
recall deployment timeout	exit 0, [1/1] [docs] other-team deployment timeout fix	exit 1, ✖ Nothing was searched: <partition>/config.yaml: Flow sequence … at line 2, column 1. Fix the file, or move it aside and run `teamai init` to write a new one.
recall --check deployment timeout	exit 0, RELEVANT … title="other-team …"	exit 1, same line
recall --check ''	exit 0, NOT_RELEVANT score=0.0	exit 1, same line
recall ''	✖ Usage: teamai recall <query>	same, no detection
recorded after the above	votes/alice.yaml, sessions/…-recall-cache.json	nothing
deployed teamai-recall agent: .claude/…md, .codex/…toml, .codebuddy/…md, .opencode/…md	no refusal relay	relays Nothing was searched:
teamai-recall rule: .claude/rules/, .codebuddy/rules/	no refusal relay	relays it
hook-dispatch post-tool-use --matcher TodoWrite and stop (transcript with recalled doc-ids), --tool claude/codex/codebuddy/opencode	no nudge, no output (#748 gate)	same
readable config: recall --check; TodoWrite nudge, four agents	RELEVANT title="team-a …"; nudge shown	same

Tests (src/__tests__/recall-unreadable-config.test.ts, real configs and index in a sandbox HOME):

broken partition + legacy other-team  → nothing searched/recorded, exit 1, message   ✗ → ✓
broken partition + user scope only     → not searched, exit 1                          ✗ → ✓
--check (query / empty query)          → message, no verdict, exit 1                   ✗ → ✓
missing query                          → usage error, not the refusal                  ✗ → ✓ (CI review)
readable, inheritUserScope, user only, no config → as before                         ✓ → ✓

Earlier, on 5bcb8ee: a live claude -p with the deployed subagent returned the other team's doc before, and the refusal line verbatim after.

Test plan

Rebase to 22824a1 (onto origin/main 1fd400e): own diff unchanged except src/recall.ts. #787 (dc233e4) conflicted there: #796's refusal kept, plus #787's guard that a detection which throws (not a deleted cwd) records no votes; one test commit follows #787's per-scope votes directories and turns #787's broken-project recall test into "nothing searched, nothing recorded". On 22824a1: npx tsc --noEmit, npx vitest run (4674 passed, 1 skipped), npm run build green. No real-CLI rerun: a rebase.

All below on 58195ab:

• npx tsc --noEmit
• npx vitest run: 4526 passed, 1 skipped
• npm run test:e2e: 229 passed, 26 skipped
• npm run build + the real-CLI matrix above: Claude, Codex, CodeBuddy, OpenCode through their deployed recall agent/rule and hook-dispatch --tool; git, github, gitlab through the team's provider: (local bare remotes; the refusal happens at detection, before any provider API call).

CI review (5812473529)

Finding	Outcome
P1 real-CLI matrix incomplete	Fixed: matrix above, three providers × four agents
P2 recall.ts:404 detection before the missing-query check	Fixed: a missing query is rejected first; only an empty --check still resolves the project, so it cannot print NOT_RELEVANT there
P2 recall.ts:419 silent branch no caller can invoke	Fixed: removed with its test; the rule for a future hook is stated in Summary
P2 docs call session-start pull machine-level work	Already true after the rebase: #792's b119ec5 still pulls at session start where teamai is not set up and stops only for an unreadable config (pullHandler, src/hook-handlers.ts), which is what docs/usage-guide*.md and CHANGELOG say. Those lines are #792's

Merge Danger

Door: two-way

Blast Radius: narrow

Projects whose config exists but cannot be read: recall <query> and --check fail loudly there instead of answering from another scope. Everywhere else output is unchanged; recall --check '' now runs project detection (including self-mode bootstrap on a fresh single-repo clone), as any recall <query> already did. The recall subcommands (enable/disable/status/feedback/maintenance/promote) still resolve their scope as before; that is a follow-up under #752.

[github-actions]

• [P1 blocking] Required real-CLI validation is incomplete. The PR test plan explicitly says Codex, OpenCode, CodeBuddy, GitLab, and GitHub were not successfully tested. AGENTS.md requires real-CLI verification to actually pass for Claude, Codex, CodeBuddy, OpenCode and all three providers before the PR.
• [P2 non-blocking] src/recall.ts:404 resolves project configuration before rejecting a missing query. Consequently, plain teamai recall can run self-mode bootstrap and perform filesystem/network side effects even though the invocation is invalid. Keep the ordinary empty-query check before detection, while retaining early detection specifically for --check.
• [P2 non-blocking] src/recall.ts:419 adds a silent-mode contract that production cannot invoke: the recall command defines no --silent option and no non-test caller passes silent. The PR body confirms there is currently no such path. Remove this speculative branch and its dedicated test per the repository’s anti-overengineering rule.
• [P2 non-blocking] docs/usage-guide.md:536, docs/usage-guide.zh-CN.md:501, and CHANGELOG.md:34 still state that session-start pull runs as machine-level work without a resolved config, but src/hook-handlers.ts:108 now returns before pulling whenever config is null. Update all affected documentation to match the new behavior.

[github-actions]

• [P2 non-blocking] The PR description still promises an options.silent path and lists its test among “5 red → green” cases, but the final commit removes both (src/recall.ts:404, src/__tests__/recall-unreadable-config.test.ts:179). Update the Summary/Evidence so the documented behavior and test count match the current diff.

Previously reported empty-query ordering, speculative silent implementation, and documentation issues are resolved. The earlier full provider × agent requirement does not apply under the checked-out review rules; the documented representative real-CLI run is sufficient.

[github-actions]

No findings.

• Previously reported empty-query ordering, speculative silent mode, documentation wording, and PR-description mismatch are resolved.
• The PR description documents successful build, tests, and representative real-CLI validation, satisfying the review rules.
• Reviewed the diff only; no PR code was executed.

[github-actions]

No findings.

• Previously reported query ordering, speculative silent mode, documentation wording, and PR-description mismatch remain resolved.
• The PR description includes a successful representative real-CLI matrix; the older commit reference is non-blocking because the rebase was conflict-only.
• Reviewed only 8cee7ab...b5b5588; no PR code was executed.

[github-actions]

No findings.

• Previously reported query ordering, speculative silent mode, documentation wording, and PR-description mismatch remain resolved.
• The PR description includes sufficient tests and representative real-CLI validation. The validation predates merge commit 9a936bb, which is non-blocking under the review rules.
• Reviewed only the specified diff; no PR code was executed.

[github-actions]

No findings.

• Previously reported query ordering, speculative silent mode, documentation wording, and PR-description mismatch remain resolved.
• The PR description includes sufficient representative real-CLI validation; testing on an older runtime commit is non-blocking under the review rules.
• Reviewed only 1fd400e...22824a1; no PR code was executed.