Skip to content
Conduit

skill

peer-consult

Multi-agent peer-consultation skill for collecting independent coding advice from multiple LLMs.

AsaqeLee/peer-consultcommunitylistedunreviewed
Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/AsaqeLee/peer-consult itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 26, 2026

Peer-Consult

English | 简体中文

CI License: MIT Python

A Codex skill plus a small, dependency-free Python CLI that asks Claude Code and Gemini CLI for independent advice on a coding / architecture problem, then writes one Markdown summary with their common points, disagreements, and an auto-drafted decision that Codex (or you) can edit.

It is a developer workflow aid, not an automated decision engine.

How it works

flowchart LR
    Q["--question + optional<br/>--context-file (docs/peer_consult/ only)"] --> P[build_request_pack<br/>prompt + JSON output contract]
    P --> C[claude CLI<br/>JSON schema]
    P --> G[gemini CLI<br/>--output-format json<br/>+ repair retry]
    C --> N[compact_peer<br/>truncate / limit]
    G --> N
    N --> D[common_and_diff<br/>root causes & option names]
    D --> R[decision_draft]
    R --> O["docs/peer_consult/YY-MM-DD-HHMM-slug.md"]
    C -. on failure or --save-raw .-> RAW[docs/peer_consult/raw/*.json<br/>auto-gitignored]
    G -.-> RAW

Decision draft rules (src/peer_consult/engine.py):

  1. If both sides failed → no decision; fix inputs / CLI availability and rerun.
  2. If one side failed → use the other side's recommendation.
  3. If both recommend the same option (normalized name match) → pick it.
  4. Otherwise prefer the recommendation with more listed tests (verifiability), then the one with fewer listed risks.
  5. Still tied → "write tests to arbitrate the disagreement first".

Safety boundaries

  • Context is only read from files under docs/peer_consult/ of the project root (path traversal is rejected).
  • --extra, --include-diff, --include-status are intentionally disabled; the tool does not read git diff/status itself.
  • The prompt tells peers they have no tool / file / network access and must answer from the given text only.
  • Raw peer outputs are saved only on failure or with --save-raw, into docs/peer_consult/raw/, which is added to docs/peer_consult/.gitignore.

Requirements

  • Python 3.9+ (standard library only at runtime)
  • claude and gemini CLIs on PATH (a missing CLI is reported as an error in the summary instead of crashing)

Quick start

git clone https://github.com/AsaqeLee/peer-consult.git
cd peer-consult
pip install -e .

# inside the project you want advice on
mkdir -p docs/peer_consult
$EDITOR docs/peer_consult/request-login-bug.md   # redacted logs / snippets
peer-consult --question "Login returns 500 after token refresh" \
             --context-file docs/peer_consult/request-login-bug.md
# -> Wrote: docs/peer_consult/<timestamp>-login-returns-500-after-token.md

Without installing: ./bin/peer-consult ....

CLI options

OptionMeaning
--questionRequired problem summary
--project-root / --cdProject root (default: git top level, else cwd)
--extra-file / --context-fileContext file, must be under docs/peer_consult/
--max-charsTruncate context (default 12000)
--save-rawAlso save raw peer outputs
--strictExit with code 2 if any peer call fails

Use as a Codex skill

The skill definition lives in .codex/skills/peer-consult/SKILL.md (with assets/request_template.md). See README_ZH.md for installing it to ~/.codex/skills/peer-consult.

Project layout

.codex/skills/peer-consult/   Codex skill + request template
bin/peer-consult              runner without installation
src/peer_consult/
  cli.py        argument parsing
  main.py       orchestration, output / raw files
  providers.py  prompt pack, claude / gemini calls, diffing
  engine.py     compaction, decision draft, Markdown rendering
  paths.py      project root and docs/peer_consult boundary
  utils.py      truncation, slugs, formatting helpers
tests/          pytest / unittest tests
reference/      upstream skills (git submodule) and notes

Tests

pip install -e . pytest
pytest -q tests

CI runs the tests on Python 3.9 and 3.12. Tests do not call real LLM CLIs.

Status / limitations

Personal tooling. Synthesis heuristics are simple string matching and counting; they do not replace human judgment.

License

MIT. See LICENSE.

↑↓ navigate openesc close