loop-solution-council

Loop: solution-council

Follow the ordered workflow below. Respect every human-approval gate, iteration bound, and stop condition.

# `mode: parallel-when-supported` means: run independently in parallel only if the
# platform provides isolated subagents; otherwise run sequentially or as simulated
# role passes, and never claim independence the target cannot guarantee.
name: solution-council
description: Produce an evidence-backed implementation recommendation through bounded specialist review.
max_iterations: 1
steps:
  - id: discovery
    name: discovery
    agent: product-discovery
    output: decision-brief
    instruction: >-
      Reuse .agent/context/ via context-discovery, inspect repository evidence,
      and produce a decision brief (facts, assumptions, preferences, blockers).
      Build one task context capsule for every specialist delegation.
      Ask at most 3 (max 5) grouped, decision-changing questions; otherwise state
      defaults and continue.

  - id: route
    name: route
    agent: council-router
    input: decision-brief
    output: selected-specialists
    instruction: >-
      Select the smallest relevant specialist set (always architecture and
      product; add others only on evidence), select each role's model tier via
      model-routing, and justify both choices. Skip the full council for trivial,
      low-risk changes. Keep chair synthesis heavy/high.

  - id: independent-review
    name: independent-review
    mode: parallel-when-supported
    agents_from: selected-specialists
    input: decision-brief
    output: specialist-reports
    instruction: >-
      Each selected specialist produces a first-pass report from the brief without
      seeing the others' reports. Use the strongest isolation the platform supports
      (parallel subagents, else sequential subagents, else simulated role passes)
      and record which mode was used.

  - id: challenge
    name: challenge
    mode: parallel-when-supported
    agents_from: selected-specialists
    input:
      - decision-brief
      - specialist-reports
    max_rounds: 1
    output: rebuttals
    instruction: >-
      Exactly one bounded round. The devil's advocate and selected specialists
      challenge the leading proposal with evidence-based objections, unsupported
      assumptions, and premature-consensus checks. No further rounds.

  - id: synthesis
    name: synthesis
    agent: council-chair
    input:
      - decision-brief
      - specialist-reports
      - rebuttals
    output: decision-record
    instruction: >-
      Synthesize (do not vote) into one decision record: recommended path,
      confidence with deliberation mode, rejected alternatives, risks, preserved
      dissent, and unresolved questions. Cite only evidence present in the reports.

  - id: approve
    name: approve
    gate: human-approval
    instruction: >-
      Stop and present the decision record for explicit human approval. Do not
      implement production code.

  - id: handoff
    name: handoff
    command: /spec
    instruction: >-
      Only after approval, hand off to /spec to enter spec-to-implementation.

Apply context-discovery before broad scanning: use a supplied context capsule first; otherwise check context frontmatter freshness, read .agent/context/routing.md, load only relevant summaries, and verify critical claims against source.