merge-dependabot-prs
Skill320 starsChanged 9 days agoMITView on GitHub
Triages and lands a batch of open Dependabot PRs in the Onyx repo, where main is gated exclusively by GitHub's merge queue: approves and enqueues green PRs, closes superseded duplicates, fixes mechanical CI failures (stale backend/requirements exports, stale bun.lock), tells real regressions apart from pre-existing breakage and flakes, and tracks every PR through to merge. Use when the user asks to merge, clean up, clear out, or land Dependabot (or similar bot-authored) PRs.
4 other projects use a version of this.
Adopted by
Most recently taken up by
- glslang/windbg-mcptheir own version, first seen Nov 6, 2026
- Addono/skillstheir own version, first seen Oct 13, 2026
- onyx-dot-app/onyxword for word, first seen Oct 8, 2026
- hussainweb/skillstheir own version, first seen Sep 18, 2026
Other versions4 projects use a version of this.
Word for word · 1
- onyx-dot-app/onyx★ 33k100% same
Edited · 0
None found.
Wrote their own version · 3
- glslang/windbg-mcp★ 21same name
- hussainweb/skills★ 3same name
- Addono/skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- backend/AGENTS.mdAGENTS.md
- documentation/AGENTS.mdAGENTS.md
- mobile/AGENTS.mdAGENTS.md
- web/AGENTS.mdAGENTS.md
- docs-writingSkill
- feature-mapSkill
- onyx-craft-devSkill
- onyx-dev-stackSkill
- playwright-e2e-testsSkill
The file
We split this skill into the rules it gives an agent
5 rules. 0 are common, 2 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 21
Treat each blocked PR as its own small diagnosis, not one "just merge it" action.
- Very specific: line 93
Run ods audit on the PR's code before you enqueue it.
- Distinctive: line 109
If it exits non-zero, STOP: Don't enqueue and don't try to work out whether the finding came from this bump or was already on main.
- Very specific: line 120
If a PR falls out of the queue (e.g.
- Very specific: line 172
Prefer one polling loop (e.g.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Merge Dependabot PRs
Treat each blocked PR as its own small diagnosis, not one "just merge it" action.
Confirm before anything consequential. Approving, closing, pushing to a PR
branch, merging main into someone's branch, and rerunning CI all touch shared
state. Triage everything into buckets first, then confirm the plan per bucket
with AskUserQuestion — not per PR, and not after the fact. Re-confirm if a
genuinely new kind of problem appears mid-flight (e.g. a real regression where
a mechanical fix was expected).
How Onyx gates merges
mainmerges exclusively through the merge queue ("Main Protection" ruleset). Enqueue withgh pr merge <pr> --auto— no strategy flag; the queue picks the method.--squash/--merge/--rebasewill be rejected.- Required checks:
quality-checks,playwright-required,database-tests,backend-check,mypy-check,Jest Tests,required. required(inpr-integration-tests.yml) andplaywright-required(inpr-playwright-tests.yml) are aggregateneeds: [...]+if: always()jobs over the slow integration/playwright matrices. If either is absent fromgh pr checks, its matrix is still running — the PR isn't blocked. (merge-group.ymlprovides fast same-named jobs inside the queue itself; they only check that the PR run's verdict used a base that landed, and skip that check for Dependabot PRs, which cannot record it.)mergeStateStatus/autoMergeRequestare unreliable for "is it actually queued" — query the queue directly: references/graphql-queries.md.
1. Discover the batch
gh pr list --search "is:open author:app/dependabot assignee:<user>" \
--json number,title,headRefName,mergeable,mergeStateStatus,reviewDecision,statusCheckRollupPRs are labeled by ecosystem (dependabot:python, dependabot:javascript,
dependabot:docker, dependabot:actions, plus dependabot:sandbox) and
assigned per .github/dependabot.yml. Flag duplicates: two PRs bumping the
same package, or a manual bump PR overlapping a bot one.
2. Triage into buckets, confirm the plan
These PRs aren't in a rush. Wait for every check to complete — required and advisory alike — and treat any red check as a failure to classify, never as noise to skip. One advisory check matters here even though the queue ignores it:
audit(audit.yml) — runs on every lockfile,pyproject.toml,.github/workflows/**, andtools/ods/**change, so it runs on almost every Dependabot PR. Treat it as required for this skill: a Dependabot PR does not get enqueued whileauditis red. A redauditmeans the bump either pulled in a vulnerable version or landed next to one, and merging it ships the finding.- Green & ready — every check completed and passing, advisory included.
- Failing — mechanical — only a generated/lock file wasn't regenerated after the bump (see step 5 for the Onyx cases).
- Failing — needs diagnosis — anything else; requires reading the failing job's log first, never just the check name.
- Conflicting — real merge conflict against main.
- Superseded — a newer PR in the batch covers it.
Independent of bucket, assess compatibility: note each PR's semver jump and whether the package is production or dev-only. Major bumps (and 0.x minors, which semver allows to break) are never covered by a blanket "enqueue the green ones" — green CI proves the build, not the behavior. Surface each one individually in the confirmation with a one-line breaking-changes summary from the release notes Dependabot embeds in the PR body, and let the user opt in per PR.
Summarize buckets and proposed actions, confirm with AskUserQuestion, then act.
3. Green: audit, approve, enqueue
Run ods audit on the PR's code before you enqueue it. CI's audit job covers
this, but it is skipped when the PR touches no audited path, and it is advisory
in the queue — so confirm it yourself for every Dependabot PR.
gh pr checkout <pr>
source .venv/bin/activate # ods ships in the repo venv
ods audit --fail-on=criticalods audit scans bun.lock and uv.lock plus open Dependabot alerts, and
exits non-zero on an unignored finding at or above --fail-on
(tools/ods/README.md). It reads the working
tree, so the PR branch must be checked out — a run on main says nothing
about the bump.
If it exits non-zero, STOP. Don't enqueue and don't try to work out whether
the finding came from this bump or was already on main. Report the advisory ids
and let the user say how to proceed.
Then, once clean:
gh pr review <pr> --approve
gh pr merge <pr> --autoIf a PR falls out of the queue (e.g. head_ref_force_pushed after a Dependabot
rebase), re-running gh pr merge <pr> --auto is routine, not a failure.
4. Superseded: comment which PR supersedes it, then gh pr close
5. Mechanical: fix in an isolated worktree
Once the bucket is approved, individual stale-generated-file fixes don't need per-PR confirmation. Known Onyx cases:
- uv bumps (
dependabot:python): the exportedbackend/requirements/*.txtfiles go stale when onlypyproject.toml/uv.lockwere bumped. Regenerate via the same pre-commit hooks CI uses:pre-commit run --files pyproject.toml uv.lock backend/requirements/*.txt - bun bumps (
dependabot:javascript): stalebun.lock— runbun installin the bumped directory (repo root orweb/).
Getting the branch: prefer the harness's isolated-worktree feature if it has
one (Claude Code: EnterWorktree). Otherwise, if the current checkout is
clean, work in place — gh pr checkout <pr>, fix, push, and return to the
previous branch. If neither applies (e.g. dirty checkout), ask the user where
to resolve (AskUserQuestion) rather than picking a spot — and steer away
from tmpfs paths like /tmp, which can hit "Disk quota exceeded"
mid-post-checkout hook (uv sync/bun install) even when they look roomy.
6. Needs diagnosis: read the failing job's log, then classify
- Pre-existing — the same job also fails on main's current HEAD (references/graphql-queries.md). Not this PR's problem; leave it and say so.
- External flake (rate limit, transient network, shared infra) — confirm
from the log, rerun once (
gh run rerun <runId> --failed). If the same failure recurs, stop and ask — a persisting "flake" may not be one. - Real regression from the bump — diagnose the root cause, then ask whether to fix or leave it. Never patch unrelated source just to force CI green.
7. Conflicts
- Untouched Dependabot branch: comment
@dependabot rebase(orrecreate) — Dependabot owns the branch and will redo it properly. - Branch we already pushed fixes to:
@dependabot rebasediscards non-Dependabot commits. Instead rebase ontoorigin/mainyourself, rerun the step 5 regeneration if lockfiles conflicted, push with--force-with-lease, and verifygit diff origin/main...HEAD --statstill shows only the intended bump. If a pre-push hook trips on a stale local cache (e.g. dev type-gen referencing a file deleted upstream), clear the cache and retry — don't skip the hook.
8. Track to completion
Prefer one polling loop (e.g. Monitor) over repeated manual checks: watch
both queue state and PR state (MERGED/CLOSED), and surface new failures as
they appear rather than staying silent until success.
9. Report
Merged (N): #A, #B, #C
Closed as superseded (N): #D (superseded by #E)
Left for you (N): #F — <specific unresolved reason>Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

