cadence

Cadence

AI coding lifecycle skills for Claude Code and Codex. Cover the path from issue intake to merged PR:

triage-issue → investigate-issue → self-review → pr-review → pr-babysit

The Problem

You: "fixed it, ready to push"
Agent: *pushes, opens PR with author-bias review, ignores half of reviewer feedback,
        PR sits open for days while CI flakes*

…and upstream:

You: "triage the backlog"
Agent: *labels everything as medium, comments are noise, two reviewers re-do the work*

With Cadence

You: /cadence:triage-issue #234
Cadence: *reads body, decides type + priority, comments only when there's meta worth recording*

You: /cadence:investigate-issue #234
Cadence: *verifies file:line at HEAD, probes blast radius, posts verdict + fix direction*

You: /cadence:self-review
Cadence: *codex cross-model loop — finds 3 issues before you push*
         → fix → re-review → green

You: gh pr create ... && /cadence:pr-review
Cadence: *dispatches 4 fresh role subagents in parallel*
         → security-reviewer / staff-engineer / sdet / spec-auditor
         → posts sticky summary + inline comments

You: /cadence:pr-babysit
Cadence: *watches CI, triages reviewer feedback into Valid/Discuss/Out-of-scope,
          addresses valid items with atomic commits, replies to threads,
          waits until green — reports ready-to-merge*

Skills

| Command | Stage | What it does | |---|---|---| | /cadence:triage-issue | intake | Assigns type/ + priority/ labels, drops status/needs-triage, leaves a meta comment only when there's judgment worth recording. One issue per invocation. | | /cadence:investigate-issue | intake (deep) | Verifies the issue's claims against HEAD (file:line still valid? still reproducing?), probes blast radius, and posts a Confirmed / Dismissed / Needs-info verdict + concrete fix direction. | | /cadence:self-review | pre-push | Codex (OpenAI) cross-model review of your branch in an auto-loop. Per-finding fix + re-review until converged. Cross-model so author bias doesn't compound. | | /cadence:pr-review | on PR open | Multi-role subagent dispatch — security, staff-engineer, sdet, spec-auditor — runs against the PR diff in isolated fresh contexts. Posts sticky summary + inline comments. Has a HARD-GATE preventing main-session self-review (author bias). | | /cadence:pr-babysit | until merge | Watches PR/MR until CI is green and every valid reviewer feedback is addressed. Triages comments, replies inline, escalates 3-round bot deadlocks. Reports ready-to-merge (never auto-merges). |

Why these skills, separately

Each skill targets a distinct stage with its own bias profile and output target:

• triage-issue is single-issue and minimal — backlog walkthroughs are the caller's job, not a skill responsibility. Output is gh labels + an optional meta comment.
• investigate-issue burns 30 min – hours per issue to verify claims against actual code before they get picked up. Output is a verdict comment with fix direction.
• self-review is cross-model (codex grades Claude's diff) because same-model self-review collapses. Output is back-to-you.
• pr-review is multi-role and main-session-isolated because PR review needs perspective diversity and a clean context. Output is GitHub/GitLab sticky + inline.
• pr-babysit is sequential and CI-driven because it's a wait-loop, not a review. Output is patches + thread replies.

Trying to fold these into one skill loses either the bias isolation or the workflow fit.

Installation

Claude Code

# Direct (dev):
claude --plugin-dir /path/to/cadence

# Or via marketplace.json in ~/.claude/:
# point ~/.claude/plugins/marketplace.json at this repo

Skills become /cadence:triage-issue, /cadence:investigate-issue, /cadence:self-review, /cadence:pr-review, /cadence:pr-babysit.

Codex CLI

The repo includes .agents/plugins/marketplace.json and plugins/cadence/ overlay — point your Codex marketplace at this repo.

Requirements

• gh (GitHub) or glab (GitLab) CLI installed
• codex CLI installed (only for self-review)
• Project with package.json or Makefile (for test-command detection in self-review)

Repository Layout

cadence/
├── .claude-plugin/plugin.json         # Claude Code plugin manifest
├── .claude-plugin/marketplace.json    # Claude Code marketplace manifest
├── .codex-plugin/plugin.json          # Codex plugin manifest
├── .agents/plugins/marketplace.json   # Codex marketplace manifest
├── plugins/cadence/                   # Codex plugin overlay (symlinks to root)
├── skills/
│   ├── triage-issue/SKILL.md
│   ├── investigate-issue/SKILL.md
│   ├── self-review/SKILL.md
│   ├── pr-review/SKILL.md + 4 role prompts (sdet, security, staff-engineer, spec-auditor)
│   └── pr-babysit/SKILL.md
├── CLAUDE.md                          # AGENTS.md → CLAUDE.md symlink
└── LICENSE                            # MIT

Design Notes

• Author bias is the core constraint for the review skills. self-review / pr-review / pr-babysit are designed around the research finding that framing a diff as "bug-free" produces the strongest detection drop among framing conditions tested across 6 LLMs (Mitropoulos et al., arXiv:2603.18740). The skills are structured so the entity finding the issue is never the entity that wrote the code.
• Cross-model + multi-role + main-session-isolation are three independent mitigations applied where each is the cheapest fix.
• pr-review has a mode: local for callers (supervisor sessions doing pre-PR critique) that want findings JSON instead of GitHub posts. Same dispatch, different output target.
• The intake skills (triage-issue / investigate-issue) are single-issue by design. Batch / backlog work is the caller's responsibility — the skill stays a pure unit, the caller loops.

License

MIT