ci(triage): stop over-labelling well-specified issues

[Hmbown]

Both auto-labellers classify issues from full body text, so the better-specified an issue is, the more wrong labels it collects. #4787 and #4790 landed tagged bug, question because their acceptance criteria said "fails when a pack drifts" and a scope heading said "How to add a locale".

Changes

triage.yml — type labels (bug, enhancement, documentation, question) now match the title only. A title states what an issue is; a body just discusses it. error and fail(ed|ure) are dropped from the bug rule outright — they describe what nearly every issue mentions.

agent-task-labels.yml — tui required only a bare \btui\b, so any passing reference to the terminal UI tagged the issue as touching it. Now anchored on crates/tui, ratatui, codewhale-tui.

Those anchors are deliberately the ones that survive the single-binary refactor (#4747): the crate path and the rendering framework stay when codewhale-tui is lib-ized, while the binary name goes away. A comment in the file says not to reintroduce a bare word match.

Verification

Replayed both regex sets over the real title+body of the four issues that misfired:

Issue	Before	After
#4787 foundation	bug, question, tui	tui
#4790 Hindi	bug, question, tui	tui
#4786 schema defect	bug, tui	tui
#4791 Ukrainian	tui	tui

Regression guards: a synthetic Crash on startup: panic in ratatui render loop still classifies bug, tui; Docs: explain how the tui differs from the web UI no longer fires tui.

Both workflows re-validated: YAML parses and the embedded github-script bodies pass node --check.

Known tradeoff

Title-only typing trades recall for precision. #4786 is a real bug whose title doesn't contain a type keyword, so it would no longer be auto-tagged bug — it was labelled by hand. Under-labelling is the cheaper error here; a wrong bug label on planned work actively misleads triage.

Separate finding — not fixed here, needs your call

Every area/OS rule in triage.yml (lines 32–44) is inert. It emits area:tui, area:core, area:mcp, area:state, area:execpolicy, area:tools, area:install, os:windows, os:macos, os:linux, lang:zh — and none of those 11 labels exist on the repo, so the existing-label filter drops all of them. That classification has never worked. Two options, both out of scope for this PR:

1. Create the 11 labels — but area:tui would duplicate the tui label this PR just tightened, and the area taxonomy needs rethinking against the single-binary layout anyway.
2. Delete the dead rules.

I'd lean (2), then reintroduce a deliberate area taxonomy once #4747 settles what the areas actually are.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[claude]

Claude finished @Hmbown's task in 1m 54s —— View job

---

Review: PR #4792 — stop over-labelling well-specified issues

• Read changed workflow files (triage.yml, agent-task-labels.yml)
• Check diff against origin/main
• Validate regex logic correctness
• Check for correctness bugs, provider/model safety, reuse, tests, security
• Post review summary

Summary

Small, well-scoped fix. I traced both regex changes against origin/main and the logic checks out — no bugs found.

Correctness

• triage.yml:34-37 — type rules now run .test(title) instead of .test(text). title is already lowercased at line 20, and text (still title\nbody) remains correctly used by the (separately, already-dead) area/OS rules at lines 43-52, so nothing else regressed.
• agent-task-labels.yml:45 — crates\/tui\b|ratatui|\bcodewhale-tui\b correctly drops the bare \btui\b alternative that over-fired on prose mentions, while keeping path- and framework-anchored signals. Traced the two regression cases from the PR body by hand: "Crash on startup: panic in ratatui render loop" still matches bug (title) and tui (ratatui), and "Docs: explain how the tui differs from the web UI" no longer matches tui (bare word gone, no crates/tui/codewhale-tui present). Both match the PR's stated expectations.
• Nothing here touches provider/model routing, so the EPIC v0.8.65 EPIC: Separate provider facts, model facts, offerings, and route resolution #2608 invariant doesn't apply to this diff.

Minor / non-blocking

• PR description says the old tui rule "required only a bare \btui\b" — checking origin/main, the prior regex was actually crates\/tui|\btui\b|ratatui, so it already had the path and framework anchors; only the bare-word alternative was the over-fire source. Doesn't affect the code change (which is correct), just a slightly overstated description.
• No automated test covers the regex behavior itself — .github/scripts/agent-task-metadata.test.sh only asserts structural invariants (trigger type, agent-ready gating, no milestone coupling), not label-matching behavior, and none of the existing scripts have a regex-behavior test to extend. Given verification here was manual replay against real issue text, a small grep/node-based fixture test (a few title/body strings → expected label set) would catch future regressions in either file. Not a blocker, since the existing test suite doesn't have this pattern to follow yet.

Agreed, out of scope

The area/OS labels at triage.yml:43-52 (area:tui, area:core, os:windows, etc.) are confirmed dead — none of those labels exist on the repo, so the existing-label filter (line 63) silently drops all of them. Agree with deferring this and leaning toward deleting the rules rather than creating an area:* taxonomy that will need rethinking once #4747 lands.

No changes requested.

[claude]

Claude finished @Hmbown's task in 1m 32s —— View job

---

Review: PR #4792 — stop over-labelling well-specified issues (re-check)

• Fetch latest branch state and diff against origin/main
• Read changed workflow files
• Validate regex/logic correctness
• Check for correctness bugs, provider/model safety, reuse, tests, security
• Post review summary

Summary

No new commits since the prior automated pass (still 1 commit, 2fb9202, sha 677859d/81e81c8 on the two files) — this re-check confirms the same conclusion: the diff is small, correct, and CI is green (CI, Web Frontend, DCO, Sync to CNB all passed; Claude PR Review still running).

Correctness

• triage.yml:34-37 — type rules now test title (already lowercased at line 20) instead of text. The still-text-driven locale/area/OS rules at 40/43-52 are untouched by this hunk, so nothing else regressed.
• triage.yml:32-34 — err(?:or)?|fail(?:ed|ure)? fully removed from the bug rule (not partially), matching the PR's stated intent — titles like "App fails on startup" no longer auto-tag bug on that term alone.
• agent-task-labels.yml:45 — crates\/tui\b|ratatui|\bcodewhale-tui\b correctly drops the bare \btui\b alternative. Hand-traced both regression fixtures from the PR body: "Crash on startup: panic in ratatui render loop" → bug (title) + tui (ratatui); "Docs: explain how the tui differs from the web UI" → no tui match (bare word gone, no crates/tui/codewhale-tui present). Both match spec.
• No provider/model routing touched — the EPIC v0.8.65 EPIC: Separate provider facts, model facts, offerings, and route resolution #2608 invariant doesn't apply here.

Minor / non-blocking (unchanged from prior pass)

• PR description says the old tui rule "required only a bare \btui\b"; origin/main actually had crates\/tui|\btui\b|ratatui — the path/framework anchors already existed, only the bare-word alternative was the over-fire source. Doesn't affect the correctness of the change itself.
• Still no automated fixture test for the regex behavior (a small title/body → expected-labels table would catch future regressions in either file). Not a blocker — no existing script in this repo follows that pattern yet to extend.

Agreed, out of scope

triage.yml:43-52's area/OS rules (area:tui, area:core, os:windows, …) remain confirmed dead — none of those 11 labels exist on the repo, so the existing-label filter at line 63 silently drops them all. This PR correctly leaves that alone; agree with deferring the taxonomy decision until #4747 (single-binary refactor) settles.

No changes requested.