tree: 41ad1b8ee6f3e58f254155e45c3734183539553f
  1. README.md
docs/pairing/README.md

Table of Contents generated with DocToc

Agentic Pairing skill family

Scope. Works on any project, ASF or not — no Apache-Software-Foundation-specific assumptions baked in.

Developer-side pre-flight review skills that run in the maintainer‘s or contributor’s own dev loop — after local changes are ready but before opening a PR. The family absorbs mechanical implementation-detail review so the eventual human-to-human conversation between contributor and maintainer stays on design, reasoning, and the trade-offs the project cares about.

Mentorship is intrinsic. Agentic Pairing skills are not a replacement for human code review; they are a pre-flight filter that separates implementation-detail nits (formatting, convention violations, obvious logical gaps) from the design-level conversation the project's contributor-to-committer path is built on.

No state changes. Every skill in this family reads local git state and returns a structured report. No PR is opened, no GitHub write happens, no comment is posted, and the working tree is never mutated.


[!TIP] Why this family

  • Review your own diff before a maintainer spends their time on it
  • Findings split into blocking and non-blocking, so the nits do not drown the real problems
  • Nothing is sent, posted, or merged — the report is the output

Install & first runs

Install just this family — one plugin, 2 skills. Review your own change before anyone else has to.

Once you have added the marketplace:

/plugin install magpie-pairing@apache-magpie

New to Magpie? The quick start walks the whole path in one place — install, the first /magpie-setup run, and a recording of it happening — plus the other agents and the secure-isolation setup to run next.

Before the first run

Nothing here has to be configured. These skills read the file below when it exists — yours in .apache-magpie-local/ or the project's in .apache-magpie-overrides/ — and fall back to a documented default when it does not.

Optional. Each has a documented fallback; absent, the skill still runs.

FileWhat it carriesRead by
project.mdProject manifest. Identity, repositories, mailing lists, tools enabled, CVE tooling, GitHub project-board + issue-template field declarations. The single file every skill reads to resolve project-scoped references.multi-agent-review, self-review

Works well with

Third-party packages, none of them required: this family works with none of them installed, and Magpie neither bundles nor depends on any. They are named because they are what a maintainer goes looking for next, and because the answer differs by agent.

Code Review — Anthropic

Automated pull-request review through several specialised agents, scored by confidence.

With this family: Runs the same kind of pass over a PR that self-review runs over your local diff.

Available on Claude Code only.

Superpowers — obra (community)

A skills library and development methodology: brainstorming, plan writing, subagent-driven execution, systematic debugging.

With this family: self-review and multi-agent-review are pre-flight reads of your own diff; Superpowers adds the test-first and systematic-debugging habits upstream of them.

Available on Claude Code, OpenAI Codex CLI, VS Code / GitHub Copilot, Google Gemini CLI, Cursor, OpenCode.

Installing it means first adding a marketplace Magpie does not publish — obra/superpowers-marketplace. The install flow asks before it does, and names whose it is.

Ships a plugin manifest per harness and one shared skills/ tree, the same shape Magpie uses, so it is not a Claude Code-only package.

Install commands per agent are in Companion skill packages.

Try these first

Illustrative shapes, not real transcripts — your output will differ. Both skills are read-only: nothing is sent, merged, or posted.

Self-review before you push.

/magpie-pairing:self-review

A self-review run over four changed files: one blocking finding and two non-blocking ones, with nothing sent, posted or merged

Send it through an adversarial panel.

/magpie-pairing:multi-agent-review

A multi-agent-review run showing what two independent reads of the same diff each found, and what only one of them found

Skills

SkillPurposeStatus
pairing-self-reviewStructured pre-flight self-review of local changes against a configurable base. Single-pass: correctness, security, and conventions in one report.experimental
pairing-multi-agent-reviewFan the same diff through three independent axis-focused passes (correctness, security, conventions); merge findings with deduplication and severity ranking. Higher confidence than a single pass; each axis is isolated so findings cannot cross-contaminate.experimental

When to use which

  • pairing-self-review — the default quick check. Fast, single context, covers correctness + security + conventions.
  • pairing-multi-agent-review — when you want adversarial independence between review axes. Each pass cannot see the others' findings, so a security issue buried in a conventions sweep cannot get lost. Use for changes touching security-sensitive paths or for the final check before a high-stakes PR.

Both skills are read-only / hand-back and produce the same structured report format, so you can switch between them without changing your workflow.


Relationship to pr-management-code-review

pairing-self-review and pairing-multi-agent-review run before a PR is open. Once a PR is open and you want a maintainer-side deep code review of an incoming contribution, use pr-management-code-review instead.


Adopter contract

The Agentic Pairing skills read local git state and write nothing: no PR is opened, no comment posted, and the working tree is never mutated. They resolve <upstream> and <project-config> placeholders when the project defines them, and run without either.

Status

Experimental. Both pairing-self-review and pairing-multi-agent-review are shipped and validate under skill-and-tool-validate. No adopter-pilot evaluation has run yet; shape may change between framework versions.

To provide pilot feedback, copy docs/pilot-report-template.md into your project notes, fill in each section, and optionally validate the filled-in report with:

uv run --project tools/pilot-report-validator pilot-report-validate <your-report.md>

Cross-references