dfg-inc/praxis-architect/council

Council

When a choice is hard to roll back, force independent viewpoints and cross-examination, then a compact verdict for the human. The council advises; the human decides. Never auto-commit an ADR from a council verdict alone.

When to use

• error-cost-gate returned council-required (mandatory — cannot skip).
• error-cost-gate returned council-optional and human chose to run it.
• Human asks to “run a council” / stress-test a design fork.
• Schema, data migration, security boundary, or cross-system coupling choices (WBS 4.6.1).

Do not run council for package-local reversible choices (error-cost-gate = skip) unless the human explicitly wants theater — prefer grill instead.

Inputs

Panel defaults

• Size: 3 members unless human asks for more.
• Lenses: reliability, security, delivery (or classic + architecture if unspecified).
• Name members explicitly in the verdict file (role labels are enough; full personas optional).

Protocol

Round 1 — Blind independent analysis

Each member sees only the problem packet + their lens. No peer text.

Each produces:

• Verdict (which option / conditional)
• Confidence (high/med/low)
• Where they may be wrong
• Top risk if their advice is ignored

Preferred orchestration: parallel subagents / forked contexts. If unavailable → Fallback below.

Round 2 — Cross-examination

Share Round 1 outputs. Sequentially, each member:

1. Names the peer position they most disagree with and why.
2. Names one insight that strengthened them.
3. Restates position (changed or not).

Round 3 — Final stance

Short restatement only (≤5 lines each). No new evidence dumps.

Synthesis

Compact verdict for the human — do not force consensus:

• Options considered
• Majority lean (if any) vs dissent
• Recommended path + conditions
• Residual risks that survive any choice
• Explicit: “Council does not decide.”

Fallback (single-session)

If multi-agent orchestration is unavailable:

1. Simulate rounds as clearly separated persona sections (### Member: Reliability, etc.).
2. Disclose at top: orchestration: simulated-single-session.
3. Keep blind Round 1 discipline (write all three analyses before cross-exam edits).

Steps (operational)

1. Confirm error-cost-gate status; if somehow skip, ask human whether to proceed anyway.
2. Write problem packet to:

design/<workPackageId>/council/<slug>-problem.md

(or architecture/audits/<workPackageId>-council-<slug>-problem.md if design folder not yet created).

1. Run Rounds 1–3 per protocol.
2. Write verdict:

design/<workPackageId>/council/<slug>-verdict.md

Required structure:

• ## Verdict first (human-readable, ≤20 lines)
• ## Dissent
• ## Residual risks
• ## Recommended path
• Appendix: round notes (optional / collapsible)

Also persist the machine outcome:

``bash node ${CLAUDE_PLUGIN_ROOT}/tools/architecture-governance.mjs record-council --in <json> --out design/<wp>/council/<id>.json ``

Approved outcomes require humanApproval.actor. Council never self-approves or writes an ADR.

1. Human gate (required) — AskUserQuestion with the verdict and the concrete options. Wait for explicit choice (option id / hybrid / reject-all).
2. Record human choice on the verdict file under ## Human decision (date + choice + rationale if given).
3. Hand off to record-decision with chosen + rejected sets. Do not write the ADR inside this skill.

Human gates

Always required. Present verdict; ask the human to choose. Silence ≠ consent. Council recommendation is not an ADR.

Done when

• Verdict file exists with Verdict-first layout + human decision section filled.
• Independent positions were produced (or fallback disclosed).
• Human choice recorded; record-decision may run next.
• No ADR auto-written from synthesis alone.

Failure modes