| name | board-of-directors |
| description | Simulate a 5-lens expert board (architect, product, security, operations, experience) to evaluate a plan, architecture choice, or feature design. Two modes: a cheap collapsed board (default) and a full multi-agent board for high-stakes decisions. Triggers: 'board review', 'board meeting', 'get expert opinions', 'director evaluation', 'consensus review'. |
Board of Directors
A 5-lens expert board evaluates high-consequence proposals. Each lens brings one
domain of judgment:
| Lens | Role | Evaluates |
|---|
| architect | Chief Architect (CA) | System design, patterns, scalability, tech debt, code quality |
| product | Chief Product Officer (CPO) | User value, market fit, scope, prioritization, usability |
| security | Chief Security Officer (CSO) | Vulnerabilities, compliance, data protection, risk |
| operations | Chief Operations Officer (COO) | Feasibility, timeline, resources, deployment |
| experience | Chief Experience Officer (CXO) | UX/UI, accessibility, user journey, design consistency |
The full per-lens personas live in directors/chief-*.md.
Two modes
The board runs in one of two modes. The verdict math is deterministic and lives
in tested code (lib/board-verdict.js) — never in the prompt.
Collapsed board — DEFAULT (1 call)
One model call emits all 5 lenses plus an overall decision as a single JSON
object. This is the default because it is cheap and fast. The output MUST match
this contract, which parseCollapsed validates and normalizes:
{
"architect": { "verdict": "approve", "score": 8, "concerns": [] },
"product": { "verdict": "approve", "score": 7, "concerns": [] },
"security": { "verdict": "conditions", "score": 6, "concerns": ["add rate limiting"] },
"operations": { "verdict": "approve", "score": 8, "concerns": [] },
"experience": { "verdict": "approve", "score": 9, "concerns": [] },
"decision": "APPROVED"
}
- Each lens carries
verdict ∈ approve | reject | conditions; score and
concerns are optional.
parseCollapsed rejects a malformed board (missing lens, bad verdict enum,
not an object, missing decision) — a broken or hallucinated board can never
pass a bogus verdict downstream.
Full board — high-stakes only (5 calls)
Five independent persona assessments (one per director, dispatched in
parallel via the board-meeting agent), each casting an approve/reject vote.
aggregateFull tallies the votes:
| Votes (approve-reject) | Resolution |
|---|
| ≥ 4 approve | APPROVED |
| exactly 3 approve | APPROVED_WITH_REVIEW |
| ≥ 3 reject | REJECTED |
| otherwise | ESCALATE |
When to escalate to the full board
The collapsed board is the default. The full board is reserved for high-stakes
decisions — collapsed-vs-full is chosen by lib/scrutiny.js (via the phase P3
risk matrix), NOT ad hoc: it scores the change's risk/stakes and only a full
tier warrants 5 independent assessments instead of one collapsed call. Low-stakes
reviews stay on the cheap collapsed path.
Invocation
- Collapsed (default): ask the model for one JSON board object matching the
contract above, then feed it to
parseCollapsed.
- Full (high-stakes): dispatch the
board-meeting agent, which runs all 5
personas independently and aggregates with aggregateFull.
