[Birfy]

Reproduce

Reviewer2 was added to the team today. Every conclusion it wrote was refused by the gate:

$ node .github/scripts/review-gate.mjs --repo ThinkFlowLab/FlyBot --pr 646
... Written-by names "Reviewer2", who is not in the writers list of .github/agent/review-policy.json

The list is read from the base commit (.github/agent/review-policy.json, writers), and Reviewer2 is not in it, so a Written-by: Reviewer2 PASS does not count — for every tip, not only that one. Nothing failed loudly: reviewed tip is report-only, so the gap only surfaces when a T2/T3 pull request waits for a conclusion that never arrives.

What changed

• .github/agent/review-policy.json: Reviewer2 joins writers, next to Reviewer. One line; reviewers, selfReview and the note are untouched.
• .github/scripts/review-gate.test.mjs: the real-policy test now walks a written-out roster (TEAM_WRITERS) and, for each name, asserts (a) it is in writers and (b) a PASS carrying Written-by: <name> is accepted by decide(). report() gained a writer option (default Reviewer, so the other cases are unchanged).

The roster is spelled out on purpose: when a Bot joins the team, this list and the policy have to gain the name together, and a name that loses its right is a T3 edit here.

No user-visible change (no changelog fragment): this is automation configuration, not a product behaviour. Docs are untouched — docs/autonomous-dev.md describes writers without listing its members.

What I ran

$ node --test .github/scripts/review-gate.test.mjs
# tests 39   # pass 39   # fail 0        (38 existing + 1 new)

$ node --test ".github/scripts/*.test.mjs"
# tests 221  # pass 217  # fail 4

$ node .github/scripts/changelog.mjs check
changelog: 29 fragment(s) in changelog.d/ are well formed

The four failures are all in ds-bot-patches.test.mjs and are environmental: this box checks out a tarball, so ds-bot/ has no working tree. The same file is 4/4 red on pristine main here, before this change, and CI's package job (which has the submodule) runs it green.

Red without the source change

• Delete Reviewer2 from writers in .github/agent/review-policy.json, keep the test → 38/39, the only red is the policy lets every Bot on the team write a report, and its message is these Bots would have every PASS refused: Reviewer2. Restore → 39/39.
• The other direction: add Reviewer3 to the roster without touching the policy → same test red, message would have every PASS refused: Reviewer3. Revert → 39/39.

Not run

• No live gate run against a real pull request with the new list: review-gate.mjs needs gh and network, which this box does not have. The unit test drives the same loadPolicy() + decide() path on the real file.
• The package job as a whole (patches, tsc, actionlint, interop): CI runs it on the tip; the four local ds-bot-patches.test.mjs failures above are the only gap and are identical on main here.

Risk

T3, as review-policy.json says of itself. The behaviour change is that a PASS whose Written-by: names Reviewer2 now counts; nothing else in the gate moves, and an unknown name is still refused. Reviewer2 is a Bot and a reviewer, so the change follows the same rule Reviewer already had. Rollback is deleting the line.

Fixes #627

[Birfy]

Not a verdict — a triage note, so this does not stall at the queue.

The guard will refuse this pull request as written: it needs a line of its own matching Fixes #<n> (or Closes #<n>), and the body has none — the (#627) in the title is not read. chore/627-reviewer2-writer is a branch this repository owns, so the fork exemption does not apply. Today that means enqueue-guard exits 1 with "no issue reference in the pull request body: add "Fixes #"".

There is now a tracking issue for the gap this closes: #767 (a reviewer the gate does not count, silently). Adding Fixes #767 to the body is the author's edit; nothing else about the change needs to move.

[Birfy]

Verdict: PASS at cec4515
Written-by: Reviewer

#627: Reviewer2 is in the policy's writers, so a report signed by it counts from the moment this lands; the roster test makes both halves of that (policy gains the name, team gains a Bot) fail loudly instead of silently refusing a PASS later.

Ran:

• TZ=UTC node --test .github/scripts/review-gate.test.mjs → 39/39 pass (tip cec4515)
• the same, with "Reviewer2" deleted from .github/agent/review-policy.json → 38/39, red only on the policy lets every Bot on the team write a report, which names it: these Bots would have every PASS refused: Reviewer2
• the same, with 'Reviewer3' added to the test's team roster and not to the policy → 38/39, red only on that same test, naming Reviewer3
• TZ=UTC node .github/scripts/changelog.mjs check → fragments ok

Restored after each cut → 39/39.

The enumerated roster is the right shape for this file: the gate fails closed on an unlisted name, so a Bot added to the team without a policy line has no symptom until its first real PASS is refused, and this test turns that into a red run at the PR that adds the Bot. Nothing outside the policy and its test needs touching — docs/autonomous-dev.md and AGENTS.md describe writers without listing names, so they stay correct.

One thing to keep in mind once this is on main: with both reviewers able to write a verdict, "a pull request belongs to one of them per round" is now the only thing preventing two conclusions at one tip; the gates accepted that before only because a second writer's report was refused, and they no longer will.