{
  "markdown": "# Parley Deck Skill\n\n<!-- Hook base: claude-1. Grafts: repository-file proof from kimi-1; five-skill close from codex-1. -->\n\n> **One model playing four reviewers is still one model.**\n\nParley Deck requires separate participants to write their own files, write round\none before reading the others, and cross-review what the others wrote. Disagreements\nstay on disk, and recorded signoffs gate what becomes final.\n\nThe working state lives in files in your repository that you can read, diff, and\nresume — not a chat log you have to trust.\n\nThis package includes six skills: the core cooperation protocol and five add-ons\nfor design, design enforcement, tracker-ready tickets, parallel worktrees, and\nsoftware-procurement bidding.\n\n## What's in the box\n\nThe installer places six skills into each detected runtime. The first is the protocol; the\nother five build on it. All six install by default — `--no-addons` takes just the core skill,\n`--only <name>[,<name>]` picks specific add-ons.\n\n- **`parley-deck`** — the multi-agent cooperation protocol.\n- **`parley-bidding`** — evidence-backed bidding on software tenders, with the irreversible steps held by a human.\n- **`parley-design`** — collaborative design that refuses to read as machine-made.\n- **`parley-design-check`** — that doctrine's rules, enforced against files on disk.\n- **`parley-tracker`** — tickets a stakeholder, a reviewer, and an agent can all read.\n- **`parley-worktrees`** — parallel agents over one repo, without silent corruption.\n\n> [!IMPORTANT]\n> **`parley-bidding` installs by default, including on upgrade.** A routine\n> `install --force` places a procurement-portal skill into every runtime this installer\n> covers — including runtimes belonging to people who never asked for a bidding tool. What\n> expands is *availability*, not permission: every gate in the skill still binds, it performs\n> no portal action without an action-specific human approval, and it never handles\n> credentials. Use `--no-addons`, or `--only` without it, to leave it out.\n\n<!-- Base: codex-1. Graft: the closing \"more than one model's first answer\" line from kimi-1. -->\n\n### [`parley-deck`](./skills/parley-deck/SKILL.md) — make multi-agent work inspectable\n\nUse the core skill when a design, plan, implementation, or review deserves\nindependent analysis. Every participant owns its canonical artifact; one agent\ndoes not proxy-write another agent's round, review, or signoff.\n\nThe protocol records kickoff, independent round one, cross-review, consensus,\n`FINAL.md`, `IMPLEMENTATION.md`, code review, and fix-up. The `fast`, `standard`,\nand `deliberation` tracks scale the route to the risk. Canonical files remain\nauthoritative whether the working surface is a local directory, GitHub pull\nrequests, or GitLab merge requests. Reach for it when the work is worth more than\none model's first answer.\n\n### [`parley-bidding`](./skills/parley-bidding/SKILL.md) — bid on software tenders without ever guessing that it worked\n\nPublic software procurement punishes the two failure modes agents are worst at: acting on\ncontent that reads like an instruction, and reporting success it cannot evidence. This add-on\ntreats tender and portal text as untrusted evidence rather than direction, defaults to\nread-only, and classifies every operation by effect before it happens — from a passive read to\nthe one irreversible submission — so each consequential step needs its own single-use human\napproval, bound to the exact payload, price, deadline and target it was granted for.\n\nIt refuses to inflate proof. Upload is not submission and a green screen is not a receipt: a\nbid claims content-verified only when the submitted bytes match the frozen SHA-256 manifest,\nand an ambiguous outcome becomes `unknown-possibly-submitted` with no retry until a read-only\nreconciliation says what actually happened. Seven local, deterministic Python tools build the\nmanifest, lint the package and the buyer's completeness requirements separately, and drive the\nlifecycle. None of them logs in, browses, uploads or submits — and the skill never handles a\npassword, cookie, token or MFA code at all.\n\nIt ships adapters for Cosinex/DTVP and NRW, subreport ELViS, and a manual profile whose proof\nceiling is an operator's word. Adapter maturity never grants permission.\n\n> **Runtime availability, stated precisely.** The payload installs into and validates in every\n> runtime this installer covers — measured across all fifteen destinations. Whether a given\n> runtime then *exposes* it as an invocable skill is that runtime's own behaviour and is\n> **NOT TESTED** here.\n>\n> **Every** packaged skill ships a `parley-addon.json` integrity manifest, so `doctor` can\n> verify any of them byte for byte — including when another installer put them there, which is\n> what the universal `skills` command above does. Such a tree reports `valid-unmanaged`:\n> healthy, and not owned by this installer. See the changelog for what that status means.\n\n<!-- Base: kimi-1. Grafts: the alongside/never-instead relationship from hermes-1; the bounded-graft constraint from claude-1. -->\n\n### [`parley-design`](./skills/parley-design/SKILL.md) — choose one visual direction without averaging it away\n\nPDS/1.0 makes participants diverge on directions, critique them, choose one whole,\nbind it as a contract, apply it, and audit what shipped. It is markdown doctrine\nwith no runtime, network, or framework; load it alongside `parley-deck`, never\ninstead of it.\n\nIts refusals are the point: no numeric aesthetic score, no house look, and no\n\"good default aesthetic\" guessed from the category. One direction wins whole;\nzero to three bounded grafts may come from losing directions, but none may modify\nthe winner's token file. Use it for a new visual world, a changed design rule, or\nan audit against a ratified contract instead of taste.\n\n<!-- Base: codex-1. Graft: the \"says so instead of passing it\" line from claude-1. -->\n\n### [`parley-design-check`](./skills/parley-design-check/SKILL.md) — enforce only what the evidence can prove\n\nThis add-on runs the checkable PDS/1.0 rules over design artifacts, DTCG token\ndocuments, stylesheets, and markup. It uses Node built-ins, carries no fallback\nregistry, and emits stable `rule-id — violation — remedy` findings.\n\nWith no registry it refuses rule checks and exits `3`. What it cannot decide is\nreported `UNJUDGEABLE`; a run that judged nothing reportable, or left a conformance\nclaim unverified, exits `4`, not `0`. Its capability declaration is generated from\nits detector modules, so it says what it cannot check instead of passing it.\n\n<!-- Base: codex-1. Grafts: \"the tracker is a mirror\" and the migration consequence from claude-1. -->\n\n### [`parley-tracker`](./skills/parley-tracker/SKILL.md) — write tickets for the business, the builder, and the agent\n\nThis skill authors canonical markdown epics, stories, and subtasks with `At a\nglance`, `[B] Business`, `[T] Technical`, and `[A] Agent directives` sections.\nAcceptance criteria carry audience tags, and the Definition of Done points back\nto those criteria with verification commands. Its gap-scan reports the full\nreadiness list; `claim` refuses to mark a ticket `in-progress` when that scan fails.\n\nThe tracker is a mirror; the markdown file is canonical. Sync is one-way by\ndefault, and pull reconciliation may write back only fields declared\n`mirror-owned`. The skill defines neutral projections for Jira, Linear, GitHub\nIssues, GitLab, Trello, and plain boards; live create/update requires an opt-in\nconnector. Change trackers and you lose a projection, not a requirement.\n\n<!-- Base: kimi-1. Graft: the no-stack-trace failure framing from claude-1. -->\n\n### [`parley-worktrees`](./skills/parley-worktrees/SKILL.md) — isolate concurrent work before it collides\n\nThis is protection against the concurrency failure that leaves no stack trace:\ntwo agents writing the same files and producing a result nobody intended. The\nbranch + worktree + file-set discipline turns that invisible corruption into a\nconflict Git can show.\n\nThe worktree-allocation table in `IMPLEMENTATION.md` is the lock manifest. Before\na second concurrent worktree is provisioned, its file set is compared with every\nclaimed boundary; an intersection is refused unless an explicit override is\nrecorded. Each implementer gets a sibling worktree. Git gives that worktree its own\nworking tree, index, `HEAD` and branch — it does not give it its own ports, databases or\ncaches, so the manifest records those overrides too. Use it when two or more sessions\nor Phase-5 implementers work in one repository at once.\n\n## Install\n\n> [!TIP]\n> **One command, most agents.** The universal skill installer from\n> [`vercel-labs/skills`](https://github.com/vercel-labs/skills) installs all six skills into\n> the coding agents it supports — a longer list than this package's own installer covers, and\n> theirs to state, not ours:\n>\n> ```bash\n> npx -y skills add feci/parley-deck-skill\n> npx -y skills list\n> ```\n>\n> It detects your agents and asks which to install into. `--agent <name>` picks them\n> explicitly; `--list` shows what the repository offers without installing anything.\n\nThis package's own installer covers fifteen named runtimes and adds health checks\n(`doctor`, `status`) and project-metadata sync that the universal one does not:\n\n```bash\nnpx -y parley-deck-skill@latest install --target all\nnpx -y parley-deck-skill@latest doctor --target all\n```\n\nIf your runtime does not pick the change up, follow its own instructions for reloading\nskills. The full command reference is in [Install, update, and remove](#install-update-and-remove).\n\n## Use Parley Deck\n\nAsk your agent for the skill by name. In Codex, use `$parley-deck`.\n\n```text\nUse $parley-deck to design this feature.\nDiscover available CLI agents, use the default participants, and write the Parley Deck artifacts.\n```\n\n```text\nUse $parley-deck to implement the accepted plan for <idea-slug>.\nFollow FINAL.md, record IMPLEMENTATION.md, run review rounds, and do not merge until consensus is ready.\n```\n\n```text\nUse $parley-deck to continue the current Parley Deck workflow.\nStart with the session-start checklist, read inbox and open ideas, then tell me the next required action.\n```\n\nSubstitute freely: *\"review this branch against the idea's FINAL.md\"*, *\"compare two\napproaches, arguing each independently in round 1\"*, or *\"use GitHub PR transport\"* — the\nshape of the request is the same.\n\nIf your runtime does not support skills directly, attach the skill and the protocol as\ninstruction context. In a repository checkout they are `skills/parley-deck/SKILL.md` and\n`skills/parley-deck/references/COOPERATION.md`; in an installed skill directory they are\n`SKILL.md` and `references/COOPERATION.md`. The skill is plain Markdown by design.\n\n## Install, update, and remove\n\nThe installer checks known user-level agent directories and CLI commands, then installs into\nthe runtimes it detects. A marker-only\ndirectory created by this installer is not treated as a real runtime.\n\nNative targets are **fifteen named runtimes** — Codex, Claude Code, Antigravity CLI (plugin\nmode), legacy Gemini CLI (extension mode), Hermes, Qwen, CodeBuddy, Goose, Kimi, Factory\nDroid, Vibe, Cursor, OpenCode, AionRS and zcode — **plus `generic`, a destination you point at with\n`--dest`.**\n\n```bash\n# every detected runtime, then verify\nnpx -y parley-deck-skill@latest install --target all\nnpx -y parley-deck-skill@latest doctor --target all\n\n# update in place\nnpx -y parley-deck-skill@latest install --target all --force\n\n# preview without writing\nnpx -y parley-deck-skill@latest install --target all --dry-run\n\n# one runtime, a project scope, or an explicit directory\nnpx -y parley-deck-skill@latest install --target codex\nnpx -y parley-deck-skill@latest install --scope project --target all --project . --include-undetected\nnpx -y parley-deck-skill@latest install --target generic --dest /path/to/skills/parley-deck\n\n# seed every supported path even where the runtime is not detected\nnpx -y parley-deck-skill@latest install --target all --include-undetected\n\n# remove managed copies\nnpx -y parley-deck-skill@latest uninstall --target all\n```\n\nCommands: `install`, `paths`, `doctor`, `status`, `sync-project`, `uninstall`, `--version`.\n\n```text\n--target auto|all|codex|claude|agy|gemini|hermes|qwen|codebuddy|goose|kimi|droid|vibe|cursor|opencode|aionrs|zcode|generic\n--scope user|project     --project <path>     --dest <path>\n--force  --dry-run  --json  --include-undetected\n--yes                    (sync-project only: without it, sync-project is a dry run)\n--no-addons              --only <name>[,<name>]\n```\n\nOther channels:\n\n```bash\nbrew install feci/parley/parley-deck-skill && parley-deck-skill install --target all\nwinget install Feci.ParleyDeckSkill        # Windows; standalone binaries also on GitHub releases\nnpm install -g parley-deck-skill && parley-deck-skill install\ngemini extensions install https://github.com/feci/parley-deck-skill   # legacy Gemini only\n```\n\nFor legacy Gemini use `--target gemini`, which writes `~/.gemini/extensions/parley-deck`.\n`gemini extensions install <repo-url>` is the other manager of that same destination — use one\nor the other, never both. The repository manifest points at `skills/parley-deck/SKILL.md` for\nthat path, and a native install rewrites its staged copy to the flat destination shape;\n**we have not been able to run the Gemini CLI to confirm it end to end.** Antigravity is a\nseparate target: prefer `--target agy`, and validate with\n`agy plugin validate ~/.gemini/config/plugins/parley-deck`.\n\nRun `parley-deck-skill paths` for the install directory of every *detected* target, or\n`parley-deck-skill paths --target all --include-undetected` for all fifteen. The installer writes\n`.parley-deck-skill-install.json` into each managed destination; updates replace marked\ninstalls safely, and unmarked directories are never overwritten or removed without `--force`.\n\n## Local agent contract\n\nParley Deck hardcodes no agent names. Any CLI agent can participate if it can run headlessly,\nreceive a prompt through stdin or a configured argument, read the workspace, write exactly the\nrequested protocol file, and report enough failure information for recovery.\n\nThe facilitator builds a capability matrix before each workflow. By default it uses a bounded\nparticipant set — normally two to four, including at least one non-facilitator when one is\navailable — the strongest discovered model and thinking mode per agent, a 30-minute timeout,\nand YES for sending the task brief plus necessary repository context to external CLI backends.\nObvious secrets and clearly sensitive customer data still require explicit handling.\n\n## Transports\n\n- `local-dir` — canonical files in the repository\n- `github-pr` — canonical files plus GitHub PR ergonomics\n- `gitlab-mr` — canonical files plus GitLab MR ergonomics\n\nCanonical files are the source of truth. PR and MR comments are mirrors.\n\n## Repository layout\n\n```text\nparley-deck-skill/\n├── skills/\n│   ├── parley-deck/              # the core skill\n│   │   ├── SKILL.md              # canonical entrypoint for agents\n│   │   ├── references/           # COOPERATION.md, compatibility.json, WORKED_EXAMPLES.md\n│   │   └── agents/               # manifest.yaml (neutral), openai.yaml (Codex UI metadata)\n│   ├── parley-bidding/           # bidding: scripts/, references/, assets/ + parley-addon.json\n│   ├── parley-design/            # doctrine + PDS/1.0 (markdown only)\n│   ├── parley-design-check/      # the checker: bin/, lib/, test/\n│   ├── parley-tracker/\n│   └── parley-worktrees/\n├── bin/  lib/                    # dependency-free Node installer\n├── test/  packaging/  scripts/\n├── plugin.json                   # Antigravity CLI plugin mode\n├── gemini-extension.json         # legacy Gemini extension mode\n└── NOTICE.md  RELEASING.md  LICENSE\n```\n\n## Related repositories, and what this one owes\n\nThis repository is only the skill layer, and the skill implements **manual facilitation**:\nan agent follows it, invokes other CLI agents, and verifies canonical files. Deterministic\nautomated orchestration is not part of it and requires separate tooling.\n\nParley Deck did not invent the ideas it runs on. The protocol lineage recorded here is\n**RHO** (Retrospective Harness Optimization), credited in `skills/parley-deck/references/COOPERATION.md` §13,\nwhere RHO's single-model self-preference is deliberately replaced by the deck's multi-agent\nquorum. `NOTICE.md` records `hallmark` and `impeccable` as the prior art studied for the\ndesign add-ons. Reference is for attribution and lineage only; no endorsement, sponsorship,\nor affiliation is implied.\n\n## License\n\nApache-2.0. See `LICENSE`.\n",
  "bytes": 16721,
  "sha": "f2edcfb1d69e32e7e9b028419a43c5ddad6defcb0332040b5c056b269a10c00b",
  "repo_slug": "feci/parley-deck-skill",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_feci_parley_deck_skill_5dbef3bd/readme"
}