feat(webhook): add optional action allow-list filter for routes

[cwest]

Problem

A webhook route subscribed to a whole event type fires on every action that event emits. A single human operation on a pull request produces several deliveries — opened, then later synchronize, reopened, closed, ready_for_review — and each one currently triggers a run. A route that only cares about new and updated PRs has no way to narrow that down, so it spawns redundant handlers for actions it doesn't care about.

Change

Add an optional actions allow-list to route config, mirroring the existing events filter. After the event-type check, if actions is set and the payload's action is not in it, the delivery is ignored before any prompt render or dispatch:

allowed_actions = route_config.get("actions", [])
action = payload.get("action", "")
if allowed_actions and action and action not in allowed_actions:
    return web.json_response({"status": "ignored", "action": action})

Empty or absent actions accepts every action, so the change is additive and fully backward-compatible — existing routes behave exactly as before.

Tests

Added to tests/gateway/test_webhook_integration.py:

• a disallowed action (closed) is ignored with no dispatch
• allowed actions (opened, synchronize) pass through to dispatch
• the no-filter case accepts all actions (backward-compat)

Full webhook test suite passes (105 tests), ruff clean.

Docs

Documented the actions property in the webhook route-properties table and the full config example.

[github-actions]

🔎 Lint report: topic/webhook-action-filter vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11012 on HEAD, 11011 on base (🆕 +1)

🆕 New issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`

✅ Fixed issues (2):

Rule	Count
unresolved-attribute	2

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`
run_agent.py:2941: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`

Unchanged: 5768 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[cwest]

The action allow-list lands in the right spot in _handle_webhook: after the signature check and the event-type filter, before prompt render and dispatch. An ignored action never reaches dispatch and never takes an idempotency slot. Empty or absent actions accepts everything, so existing routes keep their current behavior.

I ran the webhook integration suite at the PR head — the three new TestActionFilter cases pass along with the rest of the file, and ruff is clean on the changed files. I confirmed the test route secret matches the secret used to sign the request, so signatures validate and the new tests hit the real accept/ignore paths rather than bailing out on a 401.

Two notes inline: the fail-open comment in webhook.py reads narrower than the code's actual reach (the filter is generic, not pull_request-only), and the contributor-mapping line in scripts/release.py is unrelated to this feature — worth its own commit so it doesn't get tangled into a later revert of the webhook change. Neither blocks.

[cwest]

The action filter is correct and the new tests exercise it. I ran the suite in a throwaway clone: the four TestActionFilter cases and the rest of the webhook tests pass, and ruff is clean.

The filter mirrors the existing events check and sits in the right spot. Signature validation runs before it, so an unauthenticated payload never reaches the action check, and an ignored action never lands in the dedup cache, which is what you want. The fail-open on a missing action field is deliberate and pinned by test_missing_action_fails_open, so a later tightening can't silently start dropping action-less deliveries without a red test.

Docs match the behavior, including the fail-open note. One unrelated change flagged inline.

[cwest]

The filter does what the description says and the tests cover the paths that matter: disallowed action dropped, allowed actions through, no-filter backward-compat, and the missing-action fail-open. I checked out the head in a throwaway clone and ran the suite — the four new tests pass and ruff is clean, so the 105-passing claim holds. The fail-open on a missing action is the right call and the inline rationale makes the intent clear to the next reader. A couple of small things noted inline; none of them block.

[cwest]

Closing this. It was our internal dogfood PR for the webhook action-filter, used to exercise the team's build/review/merge loop and the live GitHub webhook, not meant to merge here. The same feature already has an open PR upstream (NousResearch#21744), so rather than duplicate it I left the one refinement worth adding (fail-open on a missing action) as a comment there.