Multi-agent cooperation for AI developers who want more than a chat transcript.
Parley Deck is a protocol for making several AI agents think independently, challenge each other, converge on a decision, implement it, review it, and leave behind a durable audit trail. This repository contains the portable skill layer: a vendor-neutral SKILL.md that teaches a capable agent runtime how to facilitate that workflow with local CLI agents, files, GitHub PRs, or GitLab MRs.
It is intentionally not tied to one model, one vendor, or one orchestration platform.
Fastest path:
npx -y parley-deck-skill@latest install --target allThen restart your agent runtime and verify:
npx -y parley-deck-skill@latest doctor --target allHomebrew:
brew install feci/parley/parley-deck-skill
parley-deck-skill install --target allAfter installing, ask your agent to use the skill by name. In Codex, use $parley-deck.
Use $parley-deck to design this feature.
Discover available CLI agents, use the default participants, and write the Parley Deck artifacts.
Use $parley-deck to implement the accepted plan for <idea-slug>.
Follow FINAL.md, record IMPLEMENTATION.md, run review rounds, and do not merge until consensus is ready.
Use $parley-deck to review this branch against the idea's FINAL.md.
Ask every non-implementer agent for its own review file, draft review consensus, and list agreed fixes.
Use $parley-deck to continue the current Parley Deck workflow.
Start with the session-start checklist, read inbox and open ideas, then tell me the next required action.
Useful copy/paste prompts:
Use $parley-deck for a quick architecture decision on:
<task>
Use participants codex, claude, gemini, hermes if available.
Keep the scope small and stop after FINAL.md.
Use $parley-deck to compare two implementation approaches:
<approach A>
<approach B>
Make each agent argue independently in round 1, then converge on a recommendation.
Use $parley-deck to ship this change end to end:
<task>
Run design, implementation, code review, fix-up, tests, and merge only after all signoffs are ready.
Use $parley-deck with GitHub PR transport.
Create or continue the idea for <task>, keep canonical files under parley-deck/ideas/, and mirror the lifecycle in PRs.
If your runtime does not support skills directly, attach SKILL.md and references/COOPERATION.md as instruction context. The skill is plain Markdown by design, so any capable tier-1 model can follow it.
Most multi-agent workflows fail in predictable ways:
- one agent anchors the rest before they form their own view
- disagreements vanish inside a long chat history
- implementation starts before there is real consensus
- reviews are informal, unowned, and hard to resume
- vendor-specific assumptions leak into the workflow
Parley Deck turns the conversation into structured project artifacts. Every participant writes its own files. Every round is explicit. Consensus is gated. Implementation and review are separate phases. Recovery is possible because the state lives in the repository.
When invoked by an AI agent, this skill guides it to:
- read the live
COOPERATION.mdprotocol, or the bundled fallback - discover available local CLI agents without assuming vendor names
- ask only for the required task statement, then continue with clear defaults unless the user overrides them
- start a new idea with independent Round 1 analysis
- run cross-review rounds and consensus signoff
- continue into implementation, code review, review consensus, and fix-up cycles
- enforce file ownership so headless agents create their own protocol artifacts
- handle partial completion, timeouts, unreachable agents, and user escalations
The result is a reusable cooperation loop for serious AI-assisted engineering work.
parley-deck-skill/
|-- bin/
| `-- parley-deck-skill.js
|-- lib/
| `-- installer.js
|-- SKILL.md
|-- agents/
| |-- manifest.yaml
| `-- openai.yaml
|-- gemini-extension.json
`-- references/
|-- COOPERATION.md
|-- compatibility.json
`-- WORKED_EXAMPLES.md
bin/andlib/contain the dependency-free Node installer.SKILL.mdis the canonical entrypoint for agents.references/COOPERATION.mdis a portability snapshot of the protocol.references/compatibility.jsondescribes packaged protocol and project metadata schema compatibility.references/WORKED_EXAMPLES.mdcontains non-authoritative examples and config shapes.agents/manifest.yamlis vendor-neutral metadata.agents/openai.yamlis only UI metadata for Codex/OpenAI skill tooling.gemini-extension.jsonlets Gemini CLI load the repository as an extension.
Recommended:
npx -y parley-deck-skill@latest install --target allThe installer uses an AionUI-style local runtime registry: it checks known user-level agent directories and CLI commands, then installs into the runtimes it can detect. For broad AionUI-derived targets, a marker-only directory created by this installer is not treated as a real runtime. Current native targets are Codex, Claude Code, Gemini CLI extension mode, Hermes, Qwen, CodeBuddy, Goose, Kimi, Factory Droid, Vibe, Cursor, OpenCode, and AionRS.
Install into every detected runtime:
npx -y parley-deck-skill@latest install --target allSeed every supported target path even when the runtime is not detected:
npx -y parley-deck-skill@latest install --target all --include-undetectedInstall a single target:
npx -y parley-deck-skill@latest install --target codex
npx -y parley-deck-skill@latest install --target claude
npx -y parley-deck-skill@latest install --target gemini
npx -y parley-deck-skill@latest install --target hermesInstall into a project instead of your personal skill directories:
npx -y parley-deck-skill@latest install --scope project --target all --project .Install into an explicit directory:
npx -y parley-deck-skill@latest install --target generic --dest /path/to/skills/parley-deckGlobal npm install:
npm install -g parley-deck-skill
parley-deck-skill installHomebrew:
brew install feci/parley/parley-deck-skill
parley-deck-skill installStandalone Windows binaries are attached to GitHub releases. They do not require Node:
.\parley-deck-skill-v1.1.1-windows-x64.exe install --target all --forceThis is the packaging shape intended for WinGet. Until the WinGet manifest is accepted, download the .exe from the latest GitHub release.
Gemini CLI can also install the repository as an extension:
gemini extensions install https://github.com/feci/parley-deck-skillUse either the Gemini extension command or parley-deck-skill install --target gemini, not both. The npm installer writes to ~/.gemini/extensions/parley-deck; Gemini's own GitHub installer may name the copy from the repository URL.
Manual paths:
Codex: ${CODEX_HOME:-~/.codex}/skills/parley-deck
Claude: ~/.claude/skills/parley-deck
Gemini: ~/.gemini/extensions/parley-deck
Hermes: ~/.hermes/skills/parley-deck
Qwen: ~/.qwen/skills/parley-deck
Goose: ~/.goose/skills/parley-deck
OpenCode: ~/.opencode/skills/parley-deck
Codex users can also use the built-in $skill-installer with the GitHub repository URL, then restart Codex.
parley-deck-skill install
parley-deck-skill paths
parley-deck-skill doctor
parley-deck-skill status --target all --project . --json
parley-deck-skill sync-project --project . --dry-run
parley-deck-skill uninstall
parley-deck-skill --versionUseful flags:
--target auto|all|codex|claude|gemini|hermes|qwen|codebuddy|goose|kimi|droid|vibe|cursor|opencode|aionrs|generic
--scope user|project
--project <path>
--dest <path>
--force
--dry-run
--yes
--json
--include-undetected
The installer writes .parley-deck-skill-install.json into every managed destination. Updates replace marked installs safely. Unmarked directories are never overwritten or removed unless you pass --force.
Recommended update:
npx -y parley-deck-skill@latest install --target all --forceThat downloads the latest npm release and replaces existing managed installs for every detected target. Restart the affected agent runtime after updating so it reloads SKILL.md.
Update only one runtime:
npx -y parley-deck-skill@latest install --target codex --force
npx -y parley-deck-skill@latest install --target claude --force
npx -y parley-deck-skill@latest install --target gemini --force
npx -y parley-deck-skill@latest install --target hermes --forcePreview an update without writing files:
npx -y parley-deck-skill@latest install --target all --dry-runIf you installed the package globally, update the global installer first:
npm install -g parley-deck-skill@latest
parley-deck-skill install --target all --forceIf you installed via Gemini's native extension manager instead of the npm installer:
gemini extensions update parley-deckIf you installed via Homebrew after the tap is published:
brew update
brew upgrade parley-deck-skill
parley-deck-skill install --target all --forceCheck an install:
parley-deck-skill doctor --target allUninstall managed copies:
parley-deck-skill uninstall --target allParley Deck does not require hardcoded agent names. Any CLI agent can participate if it can:
- run headlessly or semi-headlessly
- receive a prompt through stdin or a configured prompt argument
- read the project workspace
- write exactly the requested protocol file
- report enough failure information for recovery
The facilitator builds a capability matrix before each workflow. By default it uses all discovered installed CLI agents, the current agent as facilitator, the strongest discovered model and thinking mode per agent, balanced smart-fast speed, a 30 minute timeout, and YES for sending the task brief plus necessary repository/code context to external CLI backends. Obvious secrets and clearly sensitive private/customer data still require explicit handling.
The skill supports the three protocol transports:
local-dir: canonical files in the repositorygithub-pr: canonical files plus GitHub PR ergonomicsgitlab-mr: canonical files plus GitLab MR ergonomics
Canonical files remain the source of truth. PR/MR comments are mirrors and workflow aids, not the protocol authority.
This repository is only the skill layer.
parley-deck: server app, protocol deck, A2A facilitator, UI, database, auth, and spikesparley-deck-cli: standalone CLI for local and server-backed workflowsparley-deck-skill: this portable AI skill
The skill implements manual facilitation. Deterministic automated orchestration belongs in the server and CLI repositories.
This is an early, practical skill for developers experimenting with multi-agent engineering workflows. The protocol is deliberately file-first, auditable, and resumable. That makes it useful today with ordinary CLI agents, while leaving room for deeper A2A automation later.
Start with the skill, run one real discussion, and inspect the files it leaves behind. The protocol's value should be obvious from the artifact trail.
Apache-2.0. See LICENSE.