Purpose

Act as the gate-driven scheduler and state reader. Do not solve models, write model code, or draft paper sections.

../../AGENTS.md is the packaged policy source. Prefer a project-root AGENTS.md when one exists; otherwise read the packaged copy relative to this SKILL.md. Apply that policy without reproducing large reports or dashboards.

Session Start

Before orchestration in a new workspace:

• show git status --short;
• check the chosen runtime and required core packages;
• verify the workspace skeleton needed for the current request;
• read planning/session_config.json, accepting legacy mode.

Report warnings concisely. Do not create the full project skeleton unless the user is initializing a project.

State Sources

Prefer, in order:

1. planning/manifests/Qx.json
2. canonical artifacts on disk
3. legacy dashboard and legacy method/decision artifacts

Never trust a dashboard over newer canonical artifacts.

Manifest Contract

Maintain one compact JSON manifest per subquestion:

{
  "schema_version": 1,
  "question_id": "Q1",
  "rigor_profile": "lean",
  "current_gate": "G2",
  "status": "method_screened_waiting_human",
  "artifacts": {
    "method_card": "methods/Q1/q1_method_card.md",
    "decision_ledger": "methods/Q1/q1_decisions.jsonl",
    "risk_probe": "methods/Q1/probes/risk_probe_summary.json",
    "latest_run": null
  },
  "allowed": {
    "code_generation": false,
    "freeze": false,
    "paper_writing": false,
    "final_assembly": false
  },
  "blockers": [],
  "next_action": {
    "owner": "human",
    "skill": "decision-prompt-builder",
    "reason": "method choice not recorded"
  },
  "updated_at": "ISO-8601"
}

Update only fields affected by the current state change. Generate a human dashboard on request or at a milestone; otherwise derive status directly from manifests.

Gate Evaluation

Evaluate each Qx independently.

G1, PROBLEM_FRAMED