feat(webhook): a review verdict on the box's own PR spawns a session (#255)

[defangdevs]

Closes #255.

The hole

Your changes-requested review on #306 this morning reached the box and started nothing:

00:14:13Z  pull_request_review / submitted  ->  HTTP 200
00:14:13Z  not spawning for pull_request_review on defangdevs/agent-box
           — the subscribed watch declined it (when/drop rules or ignoreSenders)

The decline was correct — the watch had no clause a review could match. action is submitted, so the opened/reopened, assigned and mention clauses miss, and the verdict sits at review.state, which the bare {path="state"} CI clause never reads.

So for a PR the box wrote, the last event that could start a session was a CI failure. Success ended the box's involvement. #255 filed this as the approval case; this PR takes the widened scope you asked for — a denial is an instruction too.

The clause

{ path = "action";                  "in"  = [ "submitted" ]; }
{ path = "review.state";            "in"  = [ "approved" "changes_requested" "commented" ]; }
{ path = "pull_request.user.login"; "in"  = [ "defangdevs" ]; }
{ path = "sender.login";            notIn = [ "defangdevs" ]; }

pull_request.user.login keeps it narrow: a review on somebody else's PR is that person's business. Every leaf is in/notIn on a path GitHub already sends, so unlike the mention clause (#296) this needs no new operator and no pin bump — it works on the 0.13.0 the box runs today.

commented is in, because such a review is a batch of notes — the same request in a softer voice — and it is the only form that reaches the box at all: pull_request_review_comment is not among the events the repo hook is registered for, so a lone inline comment is invisible either way.

action = "dismissed" is deliberately out. GitHub rewrites review.state to dismissed on that action, so it cannot ride the clause anyway, and a withdrawn verdict is the removal of news. The test asserts the exclusion instead of leaving it to be assumed.

Why green CI is still not a trigger

It does not need to be, which is what changed since #255 was written. The spawned session is subscribed to the event's own repo at spawn, and #260 landed the --channels flag that had made every session delivery a no-op box-wide. So the session spawned by an approval now receives the green run as a live delivery and can merge on it. Verified here: --channels plugin:local-webhook@local-channels is on the live claude process, and a signed synthetic pull_request_review landed in a subscribed session as a <channel> message. Before #260 this clause would have spawned a session that then went deaf.

Verification

On this aarch64 box:

• nix build .#checks.aarch64-linux.{module-generated-up-to-date,module-single-file,golden-snapshot,webhook-route,multi-user} — all pass. golden-snapshot failed first on exactly the intended diff; the fixture is regenerated with nix run .#update-golden and committed, so that check asserts the shipped default carries the clause.
• bin/check-testscript.sh tests/webhook.nix — ty and ruff clean.
• nix-instantiate --parse on the generated module and the test.
• nix eval .#checks.x86_64-linux.webhook.drvPath — evaluates.
• Replayed the real declined delivery (3838135792247504896) through match_predicate against the shipped golden policy:

MATCH    REAL #306 changes_requested by lionello on a defangdevs PR
MATCH    approved,    box PR,   human
MATCH    commented,   box PR,   human
decline  changes_req, HUMAN PR, human
decline  changes_req, box PR,   BOX (self-review)
decline  dismissed,   box PR,   human
decline  CHANGES_REQUESTED uppercase (REST spelling)

New VM coverage in tests/webhook.nix on its own governed topic (the spawn name derives from the key, and the cases need three declines and one spawn): a review on a human-authored PR is declined, the box's self-review is declined, a dismissal is declined, and a changes-requested review on a box-authored PR spawns. The VM test is x86-only and runs in CI.

Not in this PR

pull_request_review_comment is not registered on the repo hook, so a single inline comment added without submitting a review never arrives. That is a hook-settings change, not a policy one — say the word and I file it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PsqGhkpjsb4kKAdVz25xKN