capx-ai/casa/casa-board

casa-board

The work half of the board. Where casa-next hands you one move, casa-board drafts a WAVE of independent moves at once, led by the one binding constraint, then commits the ones that pass. Concurrency is in the DRAFTING (parallel, expensive); serialization is in the COMMITTING (one writer, cheap), so the company brain can never be corrupted.

ToS and honesty contract (read first)

• INTERACTIVE only. You (the founder) are present; the subagents are in-session Task

subagents, never a headless claude -p or a background process (rule 5). Do not run this unattended.

• scripts/brain.mjs is the SOLE writer of brain state. Subagents DRAFT into

company-brain/outputs/<nodeId>/ and return structured results; they never mutate state.json / build-map.json / profile.json.

• A subagent's own self-grade NEVER marks a node verified. Only the grade gate (step 4)

decides pass or fail. A node Casa did not actually do, or that fails the gate, stays open. No fake progress.

Steps

1. If company-brain/profile.json does not exist, tell the founder to run /casa-start

first and stop.

1. Sync, so the wave is computed from current state:

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/brain.mjs sync company-brain ``

1. Compute the wave (k from the argument, default 3):

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/wave.mjs company-brain --k 3 ``

Read the JSON. If fallback_to_next is true (frontier under 2 independent plays), a wave is not worth the overhead: tell the founder to run /casa-next for the single best move and STOP. Otherwise describe the wave to the founder in plain words: what each piece of work is (its title, not its id), which part of the company it belongs to, and how many can genuinely run at once. Never claim more parallelism than the returned frontier_width; keep the ids and the raw JSON out of the founder-facing text unless asked.

1. Fan out, one in-session subagent PER wave node, in parallel. Give each subagent ONLY:

its nodeId, the playbook entry from ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/playbooks/_index.json (so it can read the body file and the deliverable spec), and the shared diagnosis from company-brain/ (the company north star, the binding constraint, and the structured win_definition). Each subagent:

• executes the playbook body for THIS company (specific, grounded, not generic),
• writes the deliverable to company-brain/outputs/<nodeId>/,
• returns ONLY this JSON, no prose, and mutates NOTHING:

``json { "nodeId": "<id>", "draft_summary": "<one line>", "self_grade": 0, "residual_risks": ["<what it could not finish>"], "completion_proposal": true } ``

1. Grade gate (a self-grade never verifies). For each returned draft, run the existing

grader, which combines deterministic checks (sections, word budget, copy-lint) with a rubric judgment and persists the score:

`` /casa-review grade <nodeId> ``

A draft passes only when the grade is a pass. Failed drafts are reported and left OPEN (re-queued for a later wave), never committed.

1. Serial-commit the passing drafts, ONE at a time, through the sole writer (rule-4

guarded; it refuses any node that is not engine-ready):

`` node ${CASA_ROOT:-${CLAUDE_PLUGIN_ROOT}}/scripts/brain.mjs complete company-brain <nodeId> --agent <dispatched-operator-name> ``

1. Re-sync and report the wave outcome in plain words: what was drafted, what passed

the quality check and was committed, what fell short and stays open for a later wave, and whether the company moved forward a stage. Keep node ids, grades, and internal counters out of the founder-facing report unless asked. If a commit advanced the level, STOP after this wave (the next wave is recomputed fresh) rather than draft against stale readiness.

Rules

• Never dispatch more than frontier_width subagents, and never more than k.
• One node per subagent (a failure costs one draft, not the lane).
• Only DAG-independent nodes draft together (the planner guarantees this); do not add

nodes by hand.

• The grade gate is mandatory. A self-grade is a hint, never a verification.
• No em-dashes, no emojis in any output or deliverable.