parley-deck-cli

Real multi-agent deliberation with a durable, reviewable audit trail — not three agents in three terminals, and not one model role-playing a committee.

parley runs Parley Deck: a transport-agnostic protocol for getting several AI agents to genuinely cooperate on a hard change — plus the CLI that makes it usable instead of just specified. Each agent writes its own analysis, they cross-review, reach a recorded consensus, implement, and review the implementation — every step a file you can read, diff, and resume.

Why not just spawn three agents in three terminals? Ad-hoc multi-agent gives you no audit trail, no conflict discipline, no consensus step, and no way to resume. Parley Deck ships those as protocol.

Why not ask one model to play a committee? That's solo reasoning in a costume — there's no second voice to actually disagree, and it reintroduces the single-model self-preference that quorum-gated review exists to defeat. Parley Deck is non-solo by design.

What you get

• An 8-phase idea lifecycle (§4) — kickoff → independent analysis → cross-review → consensus → FINAL.md → IMPLEMENTATION.md → code review → fix-up. Append-only, and resumable from the documents alone.
• Non-solo by design (§1) — stable agent IDs (§2), one file per agent per round (§6); no agent overwrites another.
• Compare, don't merge — the consensus "Comparison & blind spots" lens rates confidence by agreement and surfaces contradictions and blind spots instead of averaging them away.
• Transport-agnostic (§0, §11) — local-dir, github-pr, or gitlab-mr: the same protocol whether agents share a filesystem or review each other through a PR.
• Vendor/model-agnostic roster — Claude, Codex, Gemini, GLM, and more by stable ID (subject to the CLIs you have installed and authorized).
• Readiness preflight (§9.0, parley preflight) — protocol-freshness plus a live roster ping before an idea starts; any exclusion is user-confirmed.
• Advisory retrospectives (§13, parley retro) — a quorum-gated pass over the deck's own history that proposes improvements through the normal workflow, never applies them automatically.
• Supervised automation — a live TUI and an auto-drive driver advance protocol phases; agent supervision (watchdog, stall guard, validated-artifact-beats- nonzero-exit) catches hung agents; code implementation and side effects stay gated.

Inspired by — adopted & adapted

Parley Deck didn't invent these ideas; it wired them into one repository-backed, quorum-gated protocol:

• OpenRouter Fusion → the compare-not-merge consensus lens (confidence-by-agreement, blind-spots), applied to asynchronous multi-round markdown instead of a real-time API ensemble.
• OpenAI ExecPlans / PLANS.md → resume-from-the-doc state, split into a static FINAL.md and a living IMPLEMENTATION.md governed by review-consensus.
• RHO (Retrospective Harness Optimization) → §13 retro, but advisory-only and quorum-gated instead of single-model self-preference.
• kindly → strict gates, stopping judgment, no-suppression review dispositions, and artifact-wins supervision.
• Preflight readiness → §9.0 protocol-freshness and roster liveness before each idea.

Reference to these projects is for attribution and lineage only; no endorsement, sponsorship, or affiliation is implied.

Install

Install the current checkout into ~/.parley-deck while developing:

cd /path/to/parley-deck-cli
scripts/install-local.sh

The binary is installed as:

~/.parley-deck/bin/parley

Add it to your shell path if needed:

export PATH="$HOME/.parley-deck/bin:$PATH"

Verify:

parley version
parley version --all
parley help

The release version follows semantic versioning and is recorded in VERSION. parley version --all also reports parley-deck-skill installer, runtime skill, and project metadata status when the skill installer is available.

Re-run scripts/install-local.sh after pulling or building new changes to replace the installed binary with the latest local version.

Options:

scripts/install-local.sh --dry-run
scripts/install-local.sh --prefix /tmp/parley-test
scripts/install-local.sh --bin-dir "$HOME/bin"

Homebrew users can install or update from the tap:

brew update
brew install feci/parley/parley-deck-cli
brew upgrade feci/parley/parley-deck-cli

Using Parley Deck

Initialize a repository:

parley init --dir .

Inspect and verify available agents:

parley agents list --dir .
parley agents verify --dir . --agent claude
parley agents verify --dir . --full --agent codex --yes

Start a multi-agent round:

parley run --dir . --participants claude,agy --yes "Plan the next CLI slice"

Run without opening the live TUI:

parley run --dir . --no-tui --participants claude,agy --yes "Review the repo-map MVP"

Resume or inspect work:

parley status --dir .
parley status --dir . --idea repo-map-mvp
parley continue --dir . 20260517T120000.000000000Z
parley resume --dir . 20260517T120000.000000000Z

continue prints the planner-derived next safe workflow action for an existing run or idea. resume remains the read-only detail/live-view command.

Answer a human-in-the-loop question:

parley answer --dir . 20260517T120000.000000000Z q1 "Use the conservative default"

Generate repository context for prompts or tools:

parley context repo-map --dir . --format markdown --max-files 50
parley context repo-map --dir . --format json --max-files 10

Manage consensus:

parley consensus status --dir . repo-map-mvp
parley consensus draft --dir . --round 1 --by codex repo-map-mvp
parley consensus request-signoffs --dir . --participants claude,agy --yes repo-map-mvp
parley consensus signoff --dir . --agent codex --status accept --notes "Ready" repo-map-mvp

Open the TUI:

parley tui --dir .

For the full command and flag reference:

parley help

Documentation

• Agent runtime configuration
• CLI reference

Participant dropout and review gates (1.53.0)

New ideas save participant-failure-v1 under the existing [defaults].quota_auto_exclude boolean. A non-protected participant gets the original attempt and one retry after five seconds at the same timeout. Two child failures or invalid own outputs can remove it permanently from that idea; valid BLOCK/disagreement prevents dropout. Attempts remain bound across restarts. Retained vetoes, disputes and findings remain binding.

Use parley run --quota-auto-exclude=false ... to opt out for a new idea. Saved legacy quota policies retain their trigger; upgrades do not widen them. parley quota recover repairs interrupted projections/receipts, including kickoff notices, without changing history. Permanently dropped IDs cannot return even after opt-out or owner revision; the next idea probes them normally.

Reduction still requires two positively usable non-organizers, including any designated or pinned implementer. A recorded automatic loss from at least two independent reviewers to one may now satisfy the numeric review/auto-close gate. The remaining reviewer must have a known distinct model in the saved roster snapshot. Manual exclusion markers do not establish this exception. Current signers, retained dissent, strict review and independent acceptance evidence remain required; the same reviewer may goal-check in a fresh process. See evidence and timing rules.

Eligible streaming steps, including headless signoffs, default to first output 120s and stall 300s, with one retry after 5s. Buffered Zcode and default Claude text output use their hard ceiling; silence alone cannot diagnose a hang. Explicit configuration remains honored. Windows CLI assets remain experimental.

Legacy accounting and goal checks (1.54.0)

parley budget legacy inspect|apply records a one-time attended declaration for exact, structural-only legacy run bytes. Subsequent ideas retain unknown-history provenance without repeating a per-idea declaration. Changed, missing, richer or undeclared history still refuses; charges and caps remain. See operator workflow. Installing the release alone does not activate this declaration.

Goal checks use the lesser of their track's 5/15/30-minute bound and positive checker configuration; a malformed track refuses. Their original-plus-one execution attempts share a frozen deadline, including protected checkers; valid FAIL is final and retry cannot grant dropout authority.

Default output text (1.55.0)

New decks use ASD-STE100 Issue 9 for new English protocol prose. Read the working summary in Appendix B. Preserve exact quotations, code, syntax, and evidence. Pure style feedback remains advisory.

An idea can use ste_output: false with an English reason at kickoff. Later changes need owner direction or participant consensus. Existing decks adopt the rule through their normal authority update. A newer CLI cannot override an older live authority. Existing artifacts remain unchanged. This summary does not certify full ASD conformance.

Headless prompt arguments at or above 120 KiB use a private file in the invocation directory. The agent must read the entire file and verify its SHA256 before the task. This avoids the Linux argument limit and preserves the full prompt. File delivery does not prove that a model read it.