feat(2284): give the Bugbot gate a caller here (advisory)

[LukasWodka]

What this is

One of 19 sibling PRs rolling bugbot-gate.yml out to the rest of the org — the fleet half of tracebloc/backend#2284, following the claude-skills pilot (claude-skills#36).

This is step 2 of the three-step arming order the reusable's own header sets out: (1) the reusable reaches .github's main — done in #305/#312; (2) callers, advisory; (3) the bugbot / review required context — not taken here.

This is advice, not a gate — and that is deliberate

No branch protection is touched by this PR, and bugbot / review is not added to any required-context list. One measurement is why step 3 is not taken:

Bugbot does not review Dependabot PRs. Sampled 2026-08-25 over the last 30 PRs each of .github, cli, release-train, tracebloc-website, averaging-service and backend: every non-draft PR with no Cursor Bugbot check run on its head was authored by dependabot (cli#574, cli#575, tracebloc-website#510). The only other misses were drafts, which this gate passes by design.

Bugbot re-runs only on a push or an explicit bugbot run comment, so a requiredbugbot / review would park every Dependabot PR at a red check with no route to green. Until that is answered the verdict is reported and nothing is required.

Why it is safe to arm here

• Bugbot demonstrably runs in this repo. Derived rather than assumed: the cursor GitHub App is installed on the org with repository_selection: all, and a per-repo sample of the last 15 PRs found a terminal Cursor Bugbot check on every non-draft, non-Dependabot head. A repo where Bugbot never runs would only ever measure the gate's own 900s timeout.
• The pilot ran first.claude-skills#36 put the gate on real PRs on one low-traffic private repo before this went fleet-wide — CLAUDE.md rule 4 (arm while green), in the order that lets a misbehaviour be attributed to one repo rather than twenty.

repo-inventory.yml is not touched here — the flip is the required follow-up

Caller first, inventory entry after..github's caller state is read from its audit branch over the API, so a caller and its required row cannot land in one PR: the row would be checked against a branch the caller is not on yet. Same two-step blocked-gate and backend#2396 were forced into.

Between the two, a caller against an exempt row is the stale-exemption finding and the org audit is red. That window is the accepted cost of this order — the alternative is a PR that can never go green — and flipping every row to required is the required follow-up, one PR against tracebloc/.github covering all 20 repos.

Verification

• actionlint1.7.12 — the version pinned as a required check in tracebloc/.github — clean on this file.
• python3 -c "import yaml; yaml.safe_load(...)" clean.
• The uses: target resolves: bugbot-gate.yml is blob 936771bb on .github's main and develop alike, and really declares all four inputs this caller declines to pass.
• Below the header, this file is byte-identical to the proven pilot caller.
• No paths: filter · ready_for_review present · no secrets: inherit · no workflow_dispatch · no inputs passed (all four defaulted, min-severity: high included — restating the callee's default in 20 files is CLAUDE.md rule 1 in reverse).

What to watch on this PR

This PR is its own first observation: the caller is on the head, so the gate runs against it. Expected — bugbot / review reports, having read Cursor Bugbot's terminal verdict on the head, and passes with no findings open at or above high.

Refs tracebloc/backend#2284

---

Note

Low Risk
Adds a non-blocking CI workflow with minimal read permissions; merge policy and branch protection are untouched.

Overview
Adds bugbot-gate-caller.yml, a thin GitHub Actions caller that invokes the org reusable tracebloc/.githubbugbot-gate.yml@main on pull requests (opened, reopened, synchronize, ready_for_review). The job surfaces check context bugbot / review by polling Bugbot and open review threads (defaults: 900s wait, min-severity: high).

This is step 2 of the backend#2284 rollout: advisory only — no branch protection changes and bugbot / review is not required, partly because Dependabot PRs often never get a Cursor Bugbot run. repo-inventory.yml is unchanged; flipping rows to required is a separate follow-up.

The caller sets concurrency (cancel superseded runs), read-onlycontents / checks / pull-requests permissions, no paths: filter, no secrets: inherit, and no custom inputs (fleet defaults stay on the reusable).

^{Reviewed by Cursor Bugbot for commit 0b94633. Bugbot is set up for automated code reviews on this repo. Configure here.}

[saadqbal]

Checked this against the other four callers in the batch first: all five patches are byte-identical (same sha256), so this is one review, not five guesses.

Verified the claims that would hurt if wrong, rather than the prose:

• bugbot-gate.yml on .github's main really is blob 936771bb, and develop is the same blob — so @main isn't racing a divergent branch.
• It declares all four inputs (min-severity, wait-seconds, poll-seconds, quality-ref), so passing none can't die at startup_failure.
• Callee job id is review, caller job id is bugbot — bugbot / review is the context name, which matches what's already reporting on claude-skills PRs from the pilot.
• The dependabot measurement holds: cli#574, cli#575 and tracebloc-website#510 have noCursor Bugbot check run at all, while human PRs do (pass on e2e-test-agent#248, skipping on client#833). So a required bugbot / review really would park every Dependabot PR at red with no route to green, and deferring step 3 is the right call rather than a missing one.

Advisory-only with no repo-inventory.yml change in the same PR is the correct order for the reason the header gives — a required row checked against a branch the caller isn't on yet can never go green. Worth being explicit that the stale-exemption window is real and the org audit will go red until the follow-up lands; the header says so, so nobody will be surprised.

ready_for_review in the trigger list, no paths: filter, no secrets: inherit, no workflow_dispatch, concurrency keyed on github.ref — all four of those are the ones that bite later, and all four are right here.