Primary Purpose:
Parley Deck Skill is a vendor-neutral AI cooperation tool designed to enable structured multi-agent collaboration. It guides AI agents through a protocol that fosters independent analysis, cross-review processes, and consensus-building, resulting in a durable audit trail.
Key Features:
Multi-Agent Workflow: Facilitates collaboration among multiple AI agents, ensuring each contributes independently before converging on decisions.
Structured Phases: Enforces clear phases for design, implementation, review, and fix-up, with explicit rounds for analysis and sign-off.
Audit Trail: Maintains a comprehensive record of the collaboration process, enhancing transparency and accountability.
Vendor-Neutral Protocol: Works across various platforms without dependency on specific vendors or models.
Installation:
The skill can be installed via winget, making it accessible across different environments including Codex, Claude Code, Antigravity CLI, and others. This ensures seamless integration into diverse AI ecosystems.
Audience & Benefit:
Ideal for AI developers seeking structured collaboration beyond chat transcripts. It provides a clear framework for decision-making, ensuring each phase is meticulously documented and reviewed, leading to more robust outcomes with accountability embedded at every step.
README
Parley Deck Skill
> One model playing four reviewers is still one model.
Parley Deck requires separate participants to write their own files, write round
one before reading the others, and cross-review what the others wrote. Disagreements
stay on disk, and recorded signoffs gate what becomes final.
The working state lives in files in your repository that you can read, diff, and
resume — not a chat log you have to trust.
This package includes six skills: the core cooperation protocol and five add-ons
for design, design enforcement, tracker-ready tickets, parallel worktrees, and
software-procurement bidding.
What's in the box
The installer places six skills into each detected runtime. The first is the protocol; the
other five build on it. All six install by default — --no-addons takes just the core skill,
--only [,] picks specific add-ons.
parley-deck — the multi-agent cooperation protocol.
parley-bidding — evidence-backed bidding on software tenders, with the irreversible steps held by a human.
parley-design — collaborative design that refuses to read as machine-made.
parley-design-check — that doctrine's rules, enforced against files on disk.
parley-tracker — tickets a stakeholder, a reviewer, and an agent can all read.
parley-worktrees — parallel agents over one repo, without silent corruption.
> [!IMPORTANT]
> parley-bidding installs by default, including on upgrade. A routine
> install --force places a procurement-portal skill into every runtime this installer
> covers — including runtimes belonging to people who never asked for a bidding tool. What
> expands is availability, not permission: every gate in the skill still binds, it performs
> no portal action without an action-specific human approval, and it never handles
> credentials. Use --no-addons, or --only without it, to leave it out.
Use the core skill when a design, plan, implementation, or review deserves
independent analysis. Every participant owns its canonical artifact; one agent
does not proxy-write another agent's round, review, or signoff.
The protocol records kickoff, independent round one, cross-review, consensus,
FINAL.md, IMPLEMENTATION.md, code review, and fix-up. The fast, standard,
and deliberation tracks scale the route to the risk. Canonical files remain
authoritative whether the working surface is a local directory, GitHub pull
requests, or GitLab merge requests. Reach for it when the work is worth more than
one model's first answer.
parley-bidding — bid on software tenders without ever guessing that it worked
Public software procurement punishes the two failure modes agents are worst at: acting on
content that reads like an instruction, and reporting success it cannot evidence. This add-on
treats tender and portal text as untrusted evidence rather than direction, defaults to
read-only, and classifies every operation by effect before it happens — from a passive read to
the one irreversible submission — so each consequential step needs its own single-use human
approval, bound to the exact payload, price, deadline and target it was granted for.
It refuses to inflate proof. Upload is not submission and a green screen is not a receipt: a
bid claims content-verified only when the submitted bytes match the frozen SHA-256 manifest,
and an ambiguous outcome becomes unknown-possibly-submitted with no retry until a read-only
reconciliation says what actually happened. Seven local, deterministic Python tools build the
manifest, lint the package and the buyer's completeness requirements separately, and drive the
lifecycle. None of them logs in, browses, uploads or submits — and the skill never handles a
password, cookie, token or MFA code at all.
It ships adapters for Cosinex/DTVP and NRW, subreport ELViS, and a manual profile whose proof
ceiling is an operator's word. Adapter maturity never grants permission.
> Runtime availability, stated precisely. The payload installs into and validates in every
> runtime this installer covers — measured across all fourteen destinations. Whether a given
> runtime then exposes it as an invocable skill is that runtime's own behaviour and is
> NOT TESTED here.
>
> parley-bidding is the one skill that ships a parley-addon.json integrity manifest, so it
> is the one whose payload doctor can verify even when another installer put it there. See
> the changelog for what valid-unmanaged means and for the residual affecting the other five.
parley-design — choose one visual direction without averaging it away
PDS/1.0 makes participants diverge on directions, critique them, choose one whole,
bind it as a contract, apply it, and audit what shipped. It is markdown doctrine
with no runtime, network, or framework; load it alongside parley-deck, never
instead of it.
Its refusals are the point: no numeric aesthetic score, no house look, and no
"good default aesthetic" guessed from the category. One direction wins whole;
zero to three bounded grafts may come from losing directions, but none may modify
the winner's token file. Use it for a new visual world, a changed design rule, or
an audit against a ratified contract instead of taste.
This add-on runs the checkable PDS/1.0 rules over design artifacts, DTCG token
documents, stylesheets, and markup. It uses Node built-ins, carries no fallback
registry, and emits stable rule-id — violation — remedy findings.
With no registry it refuses rule checks and exits 3. What it cannot decide is
reported UNJUDGEABLE; a run that judged nothing reportable, or left a conformance
claim unverified, exits 4, not 0. Its capability declaration is generated from
its detector modules, so it says what it cannot check instead of passing it.
parley-tracker — write tickets for the business, the builder, and the agent
This skill authors canonical markdown epics, stories, and subtasks with At a glance, [B] Business, [T] Technical, and [A] Agent directives sections.
Acceptance criteria carry audience tags, and the Definition of Done points back
to those criteria with verification commands. Its gap-scan reports the full
readiness list; claim refuses to mark a ticket in-progress when that scan fails.
The tracker is a mirror; the markdown file is canonical. Sync is one-way by
default, and pull reconciliation may write back only fields declared
mirror-owned. The skill defines neutral projections for Jira, Linear, GitHub
Issues, GitLab, Trello, and plain boards; live create/update requires an opt-in
connector. Change trackers and you lose a projection, not a requirement.
This is protection against the concurrency failure that leaves no stack trace:
two agents writing the same files and producing a result nobody intended. The
branch + worktree + file-set discipline turns that invisible corruption into a
conflict Git can show.
The worktree-allocation table in IMPLEMENTATION.md is the lock manifest. Before
a second concurrent worktree is provisioned, its file set is compared with every
claimed boundary; an intersection is refused unless an explicit override is
recorded. Each implementer gets a sibling worktree. Git gives that worktree its own
working tree, index, HEAD and branch — it does not give it its own ports, databases or
caches, so the manifest records those overrides too. Use it when two or more sessions
or Phase-5 implementers work in one repository at once.
Install
> [!TIP]
> One command, most agents. The universal skill installer from
> vercel-labs/skills installs all six skills into
> the coding agents it supports — a longer list than this package's own installer covers, and
> theirs to state, not ours:
>
> bash > npx -y skills add feci/parley-deck-skill > npx -y skills list >
>
> It detects your agents and asks which to install into. --agent picks them
> explicitly; --list shows what the repository offers without installing anything.
This package's own installer covers fourteen named runtimes and adds health checks
(doctor, status) and project-metadata sync that the universal one does not:
npx -y parley-deck-skill@latest install --target all
npx -y parley-deck-skill@latest doctor --target all
If your runtime does not pick the change up, follow its own instructions for reloading
skills. The full command reference is in Install, update, and remove.
Use Parley Deck
Ask your agent for 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 .
Follow FINAL.md, record IMPLEMENTATION.md, run review rounds, and do not merge until consensus is ready.
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.
Substitute freely: "review this branch against the idea's FINAL.md", "compare two
approaches, arguing each independently in round 1", or "use GitHub PR transport" — the
shape of the request is the same.
If your runtime does not support skills directly, attach the skill and the protocol as
instruction context. In a repository checkout they are skills/parley-deck/SKILL.md and
skills/parley-deck/references/COOPERATION.md; in an installed skill directory they are
SKILL.md and references/COOPERATION.md. The skill is plain Markdown by design.
Install, update, and remove
The installer checks known user-level agent directories and CLI commands, then installs into
the runtimes it detects. A marker-only
directory created by this installer is not treated as a real runtime.
Native targets are fourteen named runtimes — Codex, Claude Code, Antigravity CLI (plugin
mode), legacy Gemini CLI (extension mode), Hermes, Qwen, CodeBuddy, Goose, Kimi, Factory
Droid, Vibe, Cursor, OpenCode and AionRS — plus generic, a destination you point at with
--dest.
# every detected runtime, then verify
npx -y parley-deck-skill@latest install --target all
npx -y parley-deck-skill@latest doctor --target all
# update in place
npx -y parley-deck-skill@latest install --target all --force
# preview without writing
npx -y parley-deck-skill@latest install --target all --dry-run
# one runtime, a project scope, or an explicit directory
npx -y parley-deck-skill@latest install --target codex
npx -y parley-deck-skill@latest install --scope project --target all --project . --include-undetected
npx -y parley-deck-skill@latest install --target generic --dest /path/to/skills/parley-deck
# seed every supported path even where the runtime is not detected
npx -y parley-deck-skill@latest install --target all --include-undetected
# remove managed copies
npx -y parley-deck-skill@latest uninstall --target all
--target auto|all|codex|claude|agy|gemini|hermes|qwen|codebuddy|goose|kimi|droid|vibe|cursor|opencode|aionrs|generic
--scope user|project --project --dest
--force --dry-run --json --include-undetected
--yes (sync-project only: without it, sync-project is a dry run)
--no-addons --only [,]
Other channels:
brew install feci/parley/parley-deck-skill && parley-deck-skill install --target all
winget install Feci.ParleyDeckSkill # Windows; standalone binaries also on GitHub releases
npm install -g parley-deck-skill && parley-deck-skill install
gemini extensions install https://github.com/feci/parley-deck-skill # legacy Gemini only
For legacy Gemini use --target gemini, which writes ~/.gemini/extensions/parley-deck.
gemini extensions install is the other manager of that same destination — use one
or the other, never both. The repository manifest points at skills/parley-deck/SKILL.md for
that path, and a native install rewrites its staged copy to the flat destination shape;
we have not been able to run the Gemini CLI to confirm it end to end. Antigravity is a
separate target: prefer --target agy, and validate with
agy plugin validate ~/.gemini/config/plugins/parley-deck.
Run parley-deck-skill paths for the install directory of every detected target, or
parley-deck-skill paths --target all --include-undetected for all fourteen. The installer writes
.parley-deck-skill-install.json into each managed destination; updates replace marked
installs safely, and unmarked directories are never overwritten or removed without --force.
Local agent contract
Parley Deck hardcodes no agent names. Any CLI agent can participate if it can run headlessly,
receive a prompt through stdin or a configured argument, read the workspace, write exactly the
requested protocol file, and report enough failure information for recovery.
The facilitator builds a capability matrix before each workflow. By default it uses a bounded
participant set — normally two to four, including at least one non-facilitator when one is
available — the strongest discovered model and thinking mode per agent, a 30-minute timeout,
and YES for sending the task brief plus necessary repository context to external CLI backends.
Obvious secrets and clearly sensitive customer data still require explicit handling.
Transports
local-dir — canonical files in the repository
github-pr — canonical files plus GitHub PR ergonomics
gitlab-mr — canonical files plus GitLab MR ergonomics
Canonical files are the source of truth. PR and MR comments are mirrors.
This repository is only the skill layer, and the skill implements manual facilitation:
an agent follows it, invokes other CLI agents, and verifies canonical files. Deterministic
automated orchestration is not part of it and requires separate tooling.
Parley Deck did not invent the ideas it runs on. The protocol lineage recorded here is
RHO (Retrospective Harness Optimization), credited in skills/parley-deck/references/COOPERATION.md §13,
where RHO's single-model self-preference is deliberately replaced by the deck's multi-agent
quorum. NOTICE.md records hallmark and impeccable as the prior art studied for the
design add-ons. Reference is for attribution and lineage only; no endorsement, sponsorship,
or affiliation is implied.