Skill: Plan Council

Multi-model planning powered by a council of model-pinned agents. Each agent independently researches the codebase and proposes an implementation plan for the same task. The orchestrator then synthesizes the proposals into a consensus plan — surfacing approaches the models agree on, alternatives where they diverge, and risks multiple models flagged independently.

This catches blind spots that a single-model plan misses: one model might notice a reusable pattern another overlooks, or flag an architectural risk the others didn't consider.

Council Models

Spawn subagents with the model parameter to pin each to a different model. Use all three when available, at least two otherwise:

• GPT-5.5
• Claude Opus 4.6
• GPT-5.3-Codex

Each subagent receives the same system preamble:

You are an independent subagent for read-only code research. MUST stay read-only. Stay within the requested scope. Do not speculate. Do not suggest patches. For every finding, cite exact files and lines. Form your own view from first principles. Do not anchor to the provided context.

Workflow

Phase 1 — Scope

Establish what needs to be planned:

• Gather the user's goal, constraints, and any relevant context (issue, PR description, conversation history).
• If the task is vague, ask the user to clarify scope before fanning out — council time is expensive.
• Run an Explore subagent to gather baseline codebase context: relevant files, existing patterns, analogous features. This context goes into the planner prompt so agents start informed.

Phase 2 — Build a Planning Brief

Before fanning out, build a preliminary orientation that each agent will receive. This is a starting point, not ground truth — agents are expected to contradict it if their own research leads elsewhere.

The brief should include only:

1. Goal — what needs to be built or changed, and why.
2. Constraints — technical constraints explicitly stated by the user (not inferred).
3. Codebase context — file and directory paths discovered in Phase 1. No conclusions, no risk assessments — just locations.
4. Open questions — unresolved areas where the user has not stated a preference.

Keep the brief under ~50 lines. Do not answer the open questions — leave them open so agents form independent views. The agents will use search, read and terminal tools to dig into specifics.

Phase 3 — Independent Planning

Spawn all subagents in parallel at once:

• One subagent per model using the model parameter. Prepend the system preamble to the planner prompt below.
• Do not share one agent's proposal with another before this phase completes.

Phase 4 — Consensus Synthesis

Compare the three proposals and classify into:

1. Consensus approach — two or more agents independently propose the same approach, architecture, or sequencing. These form the backbone of the plan. High confidence.
2. Alternatives — agents propose different viable approaches for the same aspect. Present each with trade-offs and a recommendation. Let the user decide or pick the approach with the strongest rationale.
3. Consensus risks — risks or concerns flagged by two or more agents independently. These are high-confidence and should be addressed in the plan.
4. Single-agent insights — a useful idea only one agent surfaced. Keep if well-evidenced, note it came from one model.

Drop vague suggestions, over-engineering proposals, and speculative concerns.

Phase 5 — Debate (optional)

Trigger this phase when the user asks to "debate", "discuss", or "cross-plan", or when Phase 4 produces significant alternatives that could benefit from deliberation.

1. Share the anonymized synthesis (without attributing which model said what) back to each council agent.
2. Ask each agent to evaluate the alternatives and state which approach they'd recommend, with evidence.
3. Spawn all subagents in parallel at once.

…