Skip to content

lidge-jun/codexclaw / plugins/codexclaw/skills/dev-devops

cxc-dev-devops

Verified

Use for infra, CI/CD, releases, repo policy, worktree cleanup and native desktop apps. Triggers: Dockerfile, K8s, IaC, SRE, Tauri, 스택 PR CI, 배포, 인프라, 쿠버네티스, 브랜치 정리, 브랜치 삭제, 워크트리 정리, 저장소 세팅, 브랜치 보호, 에이전트 PR, PR 정책, 데스크톱 앱, 메뉴 막대.

references/agent-pr-intake.md

View on GitHub
# Agent PR Intake — Identity, Draft-First, Supersede, Policy Strengths

Last reviewed: 2026-09-09
Applies to: GitHub repositories receiving pull requests or issues from coding agents — Copilot cloud agent, Codex cloud, Claude Code GitHub Actions, and human-driven agent sessions pushing from a personal account
When to read: Many agent PRs or issues; designing or auditing an intake policy; a superseded or duplicate PR; deciding what to auto-close
Canonical owner: dev-devops §2.9 (`DEVOPS-AGENT-INTAKE-01`)

---

## §0 The scarce resource is review

Three projects that took the flood seriously state the same economics. LLVM: "a
contribution should be worth more to the project than the time it takes to review
it." OpenSSL: "Review, not authorship, is the scarce resource in this project."
dotnet/runtime after ten months with Copilot coding agent: "One person with good
judgment and a phone can generate PRs faster than a team can review them." Ghostty
names the mechanism: agentic programming "eliminated the natural effort-based
backpressure." A policy therefore prices reviewer attention; it does not police who
typed the code.

Volume looks different from what a bot filter expects. On lidge-jun/opencodex on
2026-09-09, 0 of 72 open PRs were authored by a bot account, while 37 carried
machine-style branch names (24 in the `codex/`, `copilot/`, `claude/` family plus 13
`agent/<slug>-<date>`) pushed from human accounts, and 56 were drafts.

## §1 Identity tiers (`DEVOPS-AGENT-IDENTITY-01`, DEFAULT)

Declare in CONTRIBUTING which tier the repository observes; enforce only what it can see.

| Tier | Signal | Reliability | Notes |
|---|---|---|---|
| 0 | Human account, machine-style branch prefix (`codex/`, `agent/`, `claude/`, `copilot/`) | Heuristic only | Codex cloud's prefix is a user setting (`codex/{feature}`); a ruleset keyed on it can be defeated by changing the setting |
| 1 | Commit trailer: `Assisted-by: <agent>:<model>` or `Co-authored-by:` | Contested | OpenSSL mandates `Assisted-by`; Linux kernel, Fedora and LLVM recommend it; Kubernetes bans trailers and wants PR-description prose; Crossplane discourages attribution trailers as statistics pollution. Pick the form your upstreams accept |
| 2 | Bot or App identity (`copilot-swe-agent[bot]`; `claude[bot]`, the Claude action's default `bot_name` input; a custom App) plus `actor_is_agent` in audit logs (enterprise agent control plane, GA 2026-02-26) | Strong | Only tier that a ruleset or a required check can key on safely |

Agent conventions as documented on 2026-09-09:

| | Copilot cloud agent | Codex cloud | Claude Code GitHub Actions |
|---|---|---|---|
| Branch prefix | `copilot/` (fixed) | `codex/{feature}` (user setting; `{date}`, `{time}` tags) | `claude/` (`branch_prefix` input) |
| Opens the PR? | Yes, as **draft** | User clicks "Create Pull Request" (draft default unverified) | No — pushes a branch, returns a prefilled PR link |
| Can mark ready / approve / merge? | No; requester's approval does not count | n/a | n/a |
| Commit author | Copilot; human as co-author; **signed** by default (2026-04-03) | unverified | `claude[bot]` or the token owner; signing opt-in (`use_commit_signing`) |
| Workflow gate | "Approve and run workflows" required by default; admin toggle "Require approval for workflow runs" with a documented secret-exposure warning | n/a | `GITHUB_TOKEN` pushes do not trigger workflows; use the Claude App or a custom App token |
| Instruction files | `.github/copilot-instructions.md`, `.github/instructions/**/*.instructions.md`, `AGENTS.md`, `CLAUDE.md` | `AGENTS.override.md` → `AGENTS.md`, root-down, 32 KiB cap | `CLAUDE.md` from the base branch; `AGENTS.md` unsupported |

Cells marked unverified come from the ledger; do not tighten them without a source.

## §2 Draft-first (`DEVOPS-DRAFT-FIRST-01`, DEFAULT)

Agent PRs open as drafts; a human, or a gate check, marks them ready. Copilot cloud
agent already behaves this way and cannot mark its own PR ready. GitHub's pull request
limits exempt drafts from the per-user cap, so draft-first is also the cheapest way for
an agent to stay under the cap while work is in progress (`repo-bootstrap.md` §5).

Keep the workflow-approval gate on. GitHub's own warning on the admin toggle: allowing
workflows to run without approval "may allow unreviewed code written by Copilot to gain
write access to your repository or access your GitHub Actions secrets."

## §3 Supersede procedure (`DEVOPS-PR-SUPERSEDE-01`, STRICT)

1. Classify each candidate as LIVE, PARTIAL or SUPERSEDED from implementation evidence
   — what the replacement actually lands — never from age or file overlap.
2. Carry credit into the replacement: `Co-authored-by:` trailer or a CREDITS entry
   naming the original author and PR.
3. Comment on the superseded PR: the replacement PR number, what was carried, what was
   dropped and why.
4. Label `superseded`, then close. Keep PARTIAL work open with a `needs-info` or a
   focused replacement request instead of closing it.
5. The closed head branch is now a candidate for the closed-PR cleanup job and is
   protected by keep rules 8-10 in `branch-lifecycle.md` (namespace, tip SHA).
6. A superseded **stack parent** stays until no open child targets it
   (`branch-lifecycle.md` rule 5; `stacked-prs.md` DEV-STACK-04).

## §4 Policy strengths (`DEVOPS-AGENT-INTAKE-01`, DEFAULT)

Pick one column, write it into CONTRIBUTING, and wire only the controls it names. Values
in parentheses are what the cited project used; this file does not choose a number.

| Control | Weak | Medium | Strong |
|---|---|---|---|
| Identity | tier 0 label (`agent`) | tier 1 or 2, disclosed in PR body | tier 2 required |
| Draft rule | none | draft-first (§2) | draft-first; human marks ready |
| Linked issue | optional | recommended | required, and the issue carries an accepted label (ghostty; llama.cpp: "features must begin with an issue, not a PR") |
| PR limits | off | on, drafts exempt (GitHub, 2026-06) | on with a low cap (OpenSSL 3-4 open; llama.cpp 1 for new contributors) |
| Size | none | guidance in template | cap (Kubernetes: "large AI generated PRs … not allowed"; Godot: one self-contained change) |
| Auto-close | never | no CI green or no activity after N days (dotnet/runtime: 30 days; 44% of its closed agent PRs) | plus "cannot explain the change" (Kubernetes, LLVM) and unlinked issue |
| Supersede | manual `superseded` label | §3 | §3 plus a bot comment linking the replacement |
| Review | human | Copilot code review or CodeRabbit first pass, then human (dotnet/runtime, Kubernetes) | required status check acting as a gate (evidence file, allowed paths, required checks — the `agent-pr-gate` pattern) plus human |
| Trust tier | none | merged-PR count relaxes limits (Godot: ≤3 merged = new contributor; GitHub "smarter bypass signals" roadmap) | same, plus vouch list (ghostty) |
| Autonomous submission | allowed | allowed with disclosure | banned — a human must submit and be able to explain (LLVM, Godot, OpenSSF BCP draft) |

Four disagreements the sources do not settle; record which side the repository takes:

- **Disclosure.** ghostty, LLVM, Kubernetes, OpenSSL require it, and Godot's June 2026
  blog post does too (its CONTRIBUTING rule is commented out and not in force); Crossplane
  explicitly does not ("everyone uses AI, the signal is worthless").
- **Trailer.** `Assisted-by:` mandated (OpenSSL), recommended (kernel, Fedora, LLVM),
  accepted (KubeVirt), banned (Kubernetes), discouraged (Crossplane).
- **Is the flood permanent?** Godot, Ladybird and Zig treat it as structural. curl's
  2026-04-22 follow-up says "the slop situation is not a problem anymore" after the bug
  bounty was suspended on 2026-01-31 (HackerOne reporting later resumed) — removing the
  incentive worked where banning the tool would not.
  Do not cite curl as proof that AI reports are worthless.
- **Ban the tool or cap the rate.** Zig, QEMU, Gentoo and Servo judge AI-ness (total
  bans; see the policy summary in `oss-ai-contribution-policies.md` under the unit evidence); Godot's ban is announced but not yet in its CONTRIBUTING. OpenSSL,
  llama.cpp, GitHub and Crossplane cap concurrency regardless of provenance. GitHub's
  implementation is the tell: agent PRs count toward the limit, drafts do not.

The one sentence that recurs across otherwise opposed policies: **if you cannot
explain your change without the AI, it is closed** (ghostty, Kubernetes,
elasticsearch-py, Home Assistant, Fedora, Node.js, Trickster).

## §5 Issues

Issues attract agent PRs. Laravel disabled Issues on most packages (2026-09-03) and asks
for a PR from an agent instead, on the reasoning that a PR at least carries a diff.
GitHub announced per-user issue limits as in development. The medium option is a
`needs-info` label with a stale-close workflow (opencodex `stale-needs-info.yml`) and
removal of bait labels (`repo-bootstrap.md` §6).

## §6 Anti-patterns

| Banned | Why | Fix |
|---|---|---|
| Bot-account filter as the only agent detector | 0 hits on a repository with 37 machine-branch PRs | Tier 0 branch heuristics plus a declared tier |
| Auto-approving workflow runs for agent pushes | Unreviewed code reaches secrets (GitHub's own warning) | Keep "Require approval for workflow runs" |
| `required_signatures` without checking each agent path | Claude Actions unsigned by default; Codex unverified | `repo-bootstrap.md` §2 |
| Requiring a trailer an upstream bans | Contributors cannot satisfy both | Choose per upstream; document it |
| Closing PARTIAL work as superseded | Discards contributor work silently | §3 step 4 |
| Quoting an X post as a project policy | It is one maintainer's report | Cite it as a first-hand report only |

## §7 Sources (read 2026-09-09)

- Copilot cloud agent: https://docs.github.com/en/copilot/concepts/agents/cloud-agent/risks-and-mitigations ; https://docs.github.com/en/copilot/how-tos/use-copilot-agents/cloud-agent/configuring-agent-settings ; https://docs.github.com/en/copilot/how-tos/use-copilot-agents/cloud-agent/use-cloud-agent-on-github
- Codex cloud branch format: live settings UI at chatgpt.com/codex/cloud/settings/general; PR author identity and draft default were searched in the Codex docs and not found (negative result recorded in `devlog/_plan/260909_agent_swarm_repo_hygiene/evidence/research/github-rulesets-and-agent-conventions.md` §2.2)
- Claude Code GitHub Actions: action inputs and docs as summarized in `devlog/_plan/260909_agent_swarm_repo_hygiene/evidence/research/github-rulesets-and-agent-conventions.md` §2.3
- Agent control plane GA (`actor_is_agent`): https://github.blog/changelog/2026-02-26-enterprise-ai-controls-agent-control-plane-now-generally-available ; Agents tab: https://github.blog/changelog/2026-01-26-introducing-the-agents-tab-in-your-repository
- Pull request limits: https://github.blog/open-source/maintainers/how-pull-request-limits-are-cutting-down-the-noise/
- Practitioner policies (LLVM, OpenSSL, dotnet/runtime, ghostty, Godot, Kubernetes, Crossplane, llama.cpp, curl, Zig, Ladybird, OpenSSF): `devlog/_plan/260909_agent_swarm_repo_hygiene/evidence/research/agent-pr-flood-practitioner-writeups.md` and `evidence/research/oss-ai-contribution-policies.md` in the same unit, each entry with its URL
- Laravel Issues decision: https://x.com/taylorotwell/status/2095516796748996843 (first-hand report)
- Live opencodex numbers: `devlog/_plan/260909_agent_swarm_repo_hygiene/evidence/research/lidge-jun-repos-settings-audit.md` §5.1
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 319kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 118kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall