Boardstate Dev

Boardstate development

Read AGENTS.md §5 first (layout, commands, read-first list). This skill adds the process that isn't obvious from the tree.

The loop

1. Branch off main (worktree for anything non-trivial). Never commit to main directly except docs-only follow-ups.
2. Scoped gates locally, full matrix in CI. pnpm --filter @boardstate/<pkg> test for what you touched + root pnpm typecheck + pnpm lint (oxlint zero-warnings + prettier). Do not run the full suite locally — CI runs Node 22 + 24.
3. Changeset per behavior change (.changeset/*.md, patch/minor per touched package). Docs/examples-only changes need none.
4. PR → CI green → merge. The changesets bot opens a Version Packages PR; merging it IS the release train (publishes all bumped packages with npm provenance). One train per batch — let related PRs ride together.
5. After the train: verify dist.attestations on the registry for each published version (registry curl, not npm view — this machine pins min-release-age).

Verify discipline (this repo's record: 10 real defects caught this way)

Green tests prove the happy path; the seams need adversaries. Before merging anything touching grants, the pending-action engine, the sandbox, bindings, or the agent tools:

• Run an adversarial pass per named invariant (SPEC §11/§17/§18): one skeptic per invariant, prompted to REFUTE with concrete inputs and file:line citations.
• Gate every caller. A guarded mutation reachable via >1 path (RPC / agent tool / import / CLI) is only as strong as its weakest caller — grep every call site; don't trust the gate's own "holds for all callers" comment. (The agent workspace.replace self-grant bypass shipped exactly this way.)
• Reads ≠ actions. A read path must never reuse a side-effecting verb. If a "read" can park, queue, or mutate, it's the wrong verb (dashboard.connector.read exists because a read binding routed through action.invoke parked mutations on every refresh).
• Claim-before-await for single-shot actions (mark terminal synchronously before the first await) and compute-under-the-lock for read-modify-write unions (inside the store's mutate producer, never from an earlier unlocked read).
• Functional ≠ visually correct. After UI-adjacent changes, look at the rendered board (preview server / demo), not just the render model — a missing stylesheet passes every test.

House rules

• Wire-contract tests at every client↔server seam — assert the exact param shape that crosses the wire ({ tab, id, patch }, not each side vs. its own mock).
• External/untrusted strings render as text bindings. No unsafeHTML/innerHTML outside the vetted markdown path. Tool results, descriptions, and refusals carry the untrusted-data framing.
• Comments explain constraints (why this order, what breaks otherwise), not narration.
• Secrets are ${ENV} refs, node-side only — never literals in configs/docs, never anything the browser or document can see.
• The widget catalog is honesty-gated — any new builtin or prop change updates packages/core/src/widget-catalog.ts, and its example must pass validateWorkspaceDoc.

Where things live

• Protocol: packages/schema/SPEC.md (§11 invariants, §17 capabilities/tool grants, §18 connector broker). Architecture seams: docs/ARCHITECTURE.md.
• The operate loop, runnable: examples/operational-demo/ (keyless).
• Release checklist: docs/release-checklist.md. Roadmap: docs/ROADMAP.md.