workflow-orchestrator

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

Pass when parse, classification, data inventory, success criteria, and human framing exist. A placeholder in a human-owned field blocks the gate.

G2 — METHOD_SCREENED

Pass when:

• qx_method_card.md defines a main candidate and usable baseline;
• the baseline completes the real task with comparable output;
• risk_probe_summary.json covers applicable checks, including output degeneracy;
• main and baseline verdicts are PASS or justified CONDITIONAL;
• any fallback has a concrete trigger.

Do not require a fixed number of candidates, universal PoCs, or a source-line limit.

G2.5 — METHOD_CHOSEN_BY_HUMAN

Pass when qx_decisions.jsonl contains a human DECIDED method choice citing probe evidence. While blocked, allow data preparation but not model code generation.

G3 — CODE_AND_EXPERIMENT_REVIEWED

Pass when:

• approved main and baseline executed;
• latest run_summary.json is complete;
• language review contains passing named checks for syntax, input contract, method alignment, reproducibility, and output contract.

Accept legacy Markdown review artifacts during migration, but prefer JSON for new work.

G4 — RESULTS_JUDGED_AND_FROZEN

In lean, pass the result-judgment subgate when final-result and stability decisions cite computed evidence. Continue iterating without freezing when the human selects adjust or fallback.

In submission, additionally require:

• final method explanation;
• final result analysis;
• robustness report;
• package sign-off in the decision ledger;
• solution package;
• current frozen_numbers.json.

G5 — PAPER_SECTION_READY

Require the three writer rules, frozen-number sourcing, human-confirmed interpretation/claim scope, and verified figures.

G6 — FINAL_AUDIT_PASSED

Evaluate only in submission. Require passing consistency, completeness, and QA artifacts. Never infer that one auditor covers another.

Routing

Choose one primary next action:

• missing framing → parser/classifier or human framing card;
• missing data profile → data-auditor-cleaner;
• missing method card/probe → method-selector;
• missing human method choice → decision-prompt-builder;
• approved method without implementation plan → model-code-analyzer;
• code/review incomplete → language generator or reviewer;
• meaningful experiment awaiting judgment → result choice card;
• final results without robustness → robustness-checker;
• submission package incomplete → final explainer, result report, or package builder;
• paper ready but unaudited → the earliest missing final auditor.

Do not invoke several judgment-bearing skills speculatively.

Change Impact

Classify changes before scheduling checks:

• NONE: scratch, formatting, comments, non-semantic docs.
• LOCAL: exploratory code or method-card updates before freeze.
• CANONICAL: schema/units, symbols, equations, parameters, official values, figure paths.
• FROZEN: changes affecting frozen values or paper claims.

Route checks:

• NONE: none.
• LOCAL: local tests/review.
• CANONICAL: scoped consistency for affected Qx.
• FROZEN: thaw log, rerun affected work, re-freeze, scoped consistency.

Never schedule a full-workspace consistency audit solely because more than one file changed.

Lean vs Submission

In lean:

• require only manifests, method card, decision ledger, probe summary, and run summaries;
• do not require per-round Markdown reports, full success logs, frozen numbers, paper artifacts, or final audits;
• persist a detailed report only at a human decision point or final round.

In submission:

• require final explanations, reviews, analyses, robustness, package, freeze, paper, and G6;
• run the full three-auditor layer once before final assembly.

Compatibility

Read legacy artifacts when new ones are absent:

• planning/progress_dashboard.md
• qx_method_candidates.md
• qx_method_iteration_log.md
• qx_decision_log.md
• decisions/*_modeler_decision.md
• Markdown code reviews

Mark them legacy_source in the manifest and recommend migration at the next material edit. Do not regenerate legacy files for new work.

Output

Return a compact state report:

• profile;
• per-question current gate and blocker;
• artifacts changed or missing;
• change-impact class;
• one next action;
• optional runners-up only when they can proceed independently.

Do not paste a full dashboard or large JSON structure unless the user asks.

Verification

• State is derived from current canonical artifacts.
• Lean requirements are not confused with submission requirements.
• Human decisions were not inferred from AI suggestions.
• Code generation, freeze, paper writing, and final assembly flags match the gates.
• Audit scope matches semantic impact.
• Manifest and reported next action agree.