| name | collab-board |
| description | Orchestrate strict, turn-based PLAN→IMPL collaboration between a PRIMARY and one or more SECONDARY AI reviewers over a bounded .collab-board session tree. Use for multi-agent co-design/implementation, skeptical peer review, point tracking, phase gates, or whenever the user references COLLAB_BOARD, a collab board, or a two-agent review loop. Default roles are Claude PRIMARY and Codex SECONDARY, and registered CLI harnesses and models are configurable. |
Collab-board
Act as PRIMARY/orchestrator. Alternate one PRIMARY and one SECONDARY seat through PLAN, then IMPL.
Track disagreements as points; require both gates.
Core posture: protocol §0 and Rules 9/11 govern. Narrate in the user's language.
CLI
Run from the target project root. $SKILL is this skill directory.
node "$SKILL/scripts/collab-board.mjs" new --type FEATURE --slug jwt-auth
node "$SKILL/scripts/collab-board.mjs" lint --session <id>
node "$SKILL/scripts/collab-board.mjs" explain L26
node "$SKILL/scripts/collab-board.mjs" status --all
node "$SKILL/scripts/collab-board.mjs" activate --session <id>
node "$SKILL/scripts/collab-board.mjs" advance --session <id>
node "$SKILL/scripts/collab-board.mjs" terminal --session <id> --status COMPLETED
node "$SKILL/scripts/collab-board.mjs" points --session <id>
node "$SKILL/scripts/collab-board.mjs" doctor
node "$SKILL/scripts/collab-board.mjs" --help # the situational commands are documented here
node "$SKILL/scripts/test.mjs"
new bootstraps the root catalog and writes an immutable PROTOCOL.md inside the new session.
Legacy sessions may still declare ../../PROTOCOL.md; follow the path in HEAD, never a hardcoded
location. The script preserves an existing compatibility root but creates none for new boards.
Reset archives the old tree, never deletes it.
Claude slash commands ship under commands/. Install project-local tooling with /collab-install
or scripts/install.mjs .; globally for Codex with scripts/install.mjs --codex-global.
Start or resume
- For a new board run
new --type <BUG_FIX|FEATURE|REFACTOR|META|INVESTIGATION> --slug <slug>. Override --primary, --secondary, or --adapter only when needed.
- Read
SESSION.md before other state on the first turn; fill Topic/Goal/Done; it is write-once.
- Read HEAD; resolve
HEAD.PROTOCOL relative to the session directory and read it once per persistent thread. A fresh-dispatched SECONDARY pays this cost each turn.
- A new session is IDLE with both hands ON_HOLD. PRIMARY alone bootstraps P1: set ACTIVE, run
activate, append STATE_SET, then take and hand off the turn.
- On resume, lint first. Act only when HEAD gives your actor START (except P1 bootstrap).
Bounded turn read-set
Never bulk-read turns/ or log.md. Lint may read them in a subprocess; its contents do not enter the model context.
PLAN: HEAD; the protocol; SESSION on first turn; points; the single RESPONDS_TO shard; your agent file.
IMPL: HEAD; the protocol; points; frozen plan/context.md; impl/code_state.md; the single predecessor shard — at I1 plan/context.md stands in its place, because the phase boundary is where the frozen plan replaces the PLAN chain and the last PLAN shard is a gate attestation IMPL does not act on; your agent file; only project sources required for the assigned work.
Cadence (§1). HOT — HEAD, points, the predecessor shard: another hand may have changed them, so read them every turn. COLD — the protocol, SESSION, plan/context.md, impl/code_state.md, your agent file: frozen, or written by you, so a persistent thread reads each once and again only once its context has been renewed. A fresh dispatch holds no context and reads the whole set every turn.
HEAD is procedural truth: hands, phase, cursor, next actor/id, gates, open-plan count, protocol path. The log is authoritative derivation but is not a turn read.
Author a direct-write turn
For a non-terminal turn, HANDOFF is the commit point. Write HEAD before HANDOFF so a crash leaves a recoverable prefix. In order:
- Create
turns/<NEXT_TURN_ID>-<actor>.md in turn/v1 format: Header; FINDINGS/CHALLENGE/PROPOSAL; Evidence; Handoff; PREV; NEXT: pending. Keep it lean (§0). Cite real evidence when resolving a point; preserve overridden dissent in one - DISSENT: line. PRIMARY IMPL adds top-level - Impl: BRANCH=… BASE_COMMIT=… LATEST_COMMIT=…; SECONDARY never does.
- After the shard exists, replace only the predecessor's
NEXT: pending with the new link.
- Update points rows and Resolved In links; a point is for what changes an outcome (§0).
- Ensure log ends in newline; append TURN_COMMIT, then POINT_SET/GATE_SET events as needed.
- Update your agent mirror and recovery/thread keys.
- Atomically rewrite HEAD: hands, gates, cursor, SEQ, open count, LAST_UPDATE.
- Append HANDOFF. A final PRIMARY turn uses TERMINAL as its commit point instead.
- Run lint and fix every FAIL before proceeding.
Produce in one turn whatever one review can cover. Splitting a deliverable across turns buys an extra review round at the price of a full read-set for both seats and a fresh dispatch for the SECONDARY; §0 counts settlement, not turns. This bounds the work a turn takes on, never the length of its shard, which §0 taxes separately.
A gate turn states the challenge raised or what was checked and why no objection remains. IMPL agreement cites an applicable command plus outcome, or states none applies. A gate cannot be withdrawn, so PRIMARY gates last (§4).
Delegate SECONDARY
When SECONDARY has START, do not take its turn. Which files a dispatch reads, and how often, is the References map below. The adapter registry — ids, actor families, aliases, model/effort defaults — is owned by PROTOCOL §9 and references/adapters.md's table.
Confirm after every dispatch: shard exists, log contains TURN_COMMIT and later HANDOFF, HEAD gives PRIMARY START with bumped SEQ, then lint. Never trust process narration. If lint FAILs, recovery.md's "Lint correction" routes it by what it names; only an all-L10 FAIL remains landed for PRIMARY's forced decision.
NOT_MY_TURN also re-delegates after re-reading HEAD; never alter state to fit the dispatch.
Phase and terminal transitions
PLAN→IMPL requires no OPEN P point, both PLAN gates YES, and a frozen plan/context.md that lists every implementation file. PRIMARY keeps START, writes the digest, and runs advance; do not delegate an extra turn after both gates agree. Only PRIMARY edits project files.
When both IMPL gates are YES, run terminal --status COMPLETED (or ABORTED). Both hands become DONE.
For L26/non-convergence, stop incremental turns and run Rule 11's step-back.
A known SECONDARY limit is scheduling, not a lower review bar; protocol §4 governs the pause and resume.
References
Each file, and when to read it:
references/protocol.md: session protocol source; rules, schemas, log grammar, state machine — cadence in Start or resume step 3.references/adapters.md: common dispatch interface and direct-write prompt — needed at every delegated turn, but cold like the protocol: a skill file cannot change mid-session, so read it once per orchestrator thread and again once your context has been renewed. Compose each dispatch from the file, never from memory of it.references/executors/<adapter-id>.md: probes/argv/result/vendor facts for a CLI adapter id — the one file that id names, once per orchestrator thread and again on resume or an adapter change. A subagent:/manual id has none; it reads references/non-cli-adapters.md.references/hosts/claude-code.md, references/hosts/codex-cli.md: PRIMARY runtime mechanics — the current host, once per orchestrator thread.references/non-cli-adapters.md: subagent:<name>, manual, peer mode — a CLI thread never reads it.references/relay.md: scribe, sole-writer relay, fusion, panels, cross-product — only for PRIMARY_ONLY, a panel, or scribe fallback.references/recovery.md: failure classification, partial recovery, resume, limits — only after an invalid, timed-out, killed, limited, lint-failed, partial, or deliberately resumed dispatch.references/manager.md: optional PRIMARY fan-out and durable panel preference — only when fanning out or persisting a panel choice.references/workers.md: delegating one well-defined job outside the board, verified by diff — only for a Rule 7 worker dispatch.references/lint-spec.md: lint checks through L30 (L17 retired); resolve a finding via explain <code>.
