Skill: full-pipeline

End-to-end research pipeline. You are the pipeline conductor — you sequence the four sub-skills, check gates, and persist state across sessions.

The pipeline orchestrates exactly four subskills, in order:

idea-discovery-pipeline → deploy-experiment → auto-review-loop → generate-report

---

Constants

• AUTO_PROCEED = true — When true, Gate 1 auto-selects the top-ranked idea (highest pilot signal + novelty confirmed) and continues to implementation. When false, always waits for explicit user confirmation before proceeding. Passed through to /idea-discovery-pipeline.
• ARXIV_DOWNLOAD = false — When true, /lit-review (inside /idea-discovery-pipeline) downloads the top relevant arXiv PDFs. When false (default), only fetches metadata via arXiv API. Passed through to /idea-discovery-pipeline.
• HUMAN_CHECKPOINT = false — When true, the auto-review loop (Stage 3) pauses after each round's review to let you see the score and provide custom modification instructions before fixes are implemented. When false (default), the loop runs fully autonomously. Passed through to /auto-review-loop.
• REVIEWER_DIFFICULTY = medium — How adversarial the reviewer is. medium (default): standard MCP review. hard: adds reviewer memory + debate protocol. nightmare: GPT reads the repo directly via codex exec + memory + debate. Passed through to /auto-review-loop.
• COMPACT_MODE = false — When true, prefer compact summaries (output/PROJ_NOTES.md, output/experiment/EXPERIMENT_RESULT.md) over full logs (output/experiment/EXPERIMENT_LOG.md, output/AUTO_REVIEW_REPORT.md) on session recovery. Passed through to /auto-review-loop and /generate-report.

Override via argument, e.g., /full-pipeline "topic" — AUTO_PROCEED: false, human checkpoint: true, difficulty: nightmare.

---

Startup: Load Context, Check & Resume

Before any subskill is invoked, load the project context. All discovered files override $ARGUMENTS; $ARGUMENTS is only the primary driver when none of the context files exist or carry usable information.

1. Read CLAUDE.md — project dashboard, control flags, canonical output paths.
2. Read handoff.json if it exists — pipeline.stage, recovery.resume_skill, recovery.read_first.
3. Read memory/MEMORY.md for the current pipeline stage and prior token usage.
4. Load research context (priority order):
   1. RESEARCH_PLAN.md in the project root — if present and non-empty, this is the authoritative research brief (problem, method, constraints, success criteria, prior attempts). Treat it as ground truth.
   2. BRIEF.md in the project root — if RESEARCH_PLAN.md is missing, read BRIEF.md (the 12-section research brief — see CLAUDE.md Key Files). Treat it as the authoritative brief.
   3. $ARGUMENTS — if neither file exists, or both are empty / contain only template placeholders, fall back to $ARGUMENTS as the research direction.
   4. Both file and $ARGUMENTS present: the file's details (problem, constraints, success criteria) override; $ARGUMENTS may sharpen scope but cannot contradict the brief. If they conflict materially, surface the conflict to the user and ask which one to follow before proceeding.
5. Read output/REVIEW_STATE.json if it exists — resume from saved round (see /auto-review-loop resume rules).
6. Read the experiment status file appropriate for COMPACT_MODE:
   • COMPACT_MODE = true → output/PROJ_NOTES.md and output/experiment/EXPERIMENT_RESULT.md
   • COMPACT_MODE = false → output/experiment/EXPERIMENT_LOG.md and output/experiment/EXPERIMENT_RESULT.md
7. Display the resolved context (which brief was loaded, current stage, last action) and confirm with the user before proceeding.

💡 If neither RESEARCH_PLAN.md nor BRIEF.md exists and $ARGUMENTS is also empty, stop and ask the user for a research direction or to create one of these files (templates: templates/RESEARCH_PLAN_TEMPLATE.md).

---

Stage 1 — Idea Discovery (/idea-discovery-pipeline)

Goal: Identify a tractable, novel research direction and produce an experiment plan.

Invoke the subskill, passing through control flags:

/idea-discovery-pipeline "$ARGUMENTS"

The subskill internally chains: /lit-review → /generate-idea → /novelty-check → /idea-review → /experiment-design-pipeline (which itself runs /refine-research → /experiment-design). It will read RESEARCH_PLAN.md directly in its Phase 0 — do not re-paste its contents into the prompt.

Outputs (canonical paths produced by /idea-discovery-pipeline):

• output/LIT_REVIEW_REPORT.md — literature landscape and ranked gaps
• output/IDEA_REPORT.md — ranked, novelty-checked, pilot-tested ideas (top-level report)
• output/NOVELTY_REPORT.md — deep novelty verdicts
• output/IDEA_REVIEW_REPORT.md — external critical review
• output/refine-logs/FINAL_PROPOSAL.md + output/refine-logs/REFINE_REPORT.md — refined proposal
• output/EXPERIMENT_PLAN.md + output/EXPERIMENT_TRACKER.md — experiment roadmap
• output/EXP_PIPELINE_SUMMARY.md — pipeline summary

Gate 1: If AUTO_PROCEED = false, stop here and present the top ideas to the user:

📋 Idea Discovery Pipeline completed! Top ideas:

1. [Idea 1 title] — Pilot: POSITIVE (+X%), Novelty: 9/10
2. [Idea 2 title] — Pilot: WEAK POSITIVE (+Y%), Novelty: 8.3/10
3. [Idea 3 title] — Pilot: NEGATIVE, eliminated

Recommended: Idea 1. Shall I proceed with deployment?

The user may approve, pick a different idea, request changes (re-run the subskill with refined constraints), reject all (collect feedback, re-run), or stop and save state.

If AUTO_PROCEED = true, the subskill auto-selects the top idea after presenting results.

After selection, ensure the chosen idea and supporting context are recorded in RESEARCH_PLAN.md (template: templates/RESEARCH_PLAN_TEMPLATE.md) so downstream stages and future sessions inherit it.

Path bridge before Stage 2: /deploy-experiment reads output/refine-logs/EXPERIMENT_PLAN.md and output/refine-logs/FINAL_PROPOSAL.md. /idea-discovery-pipeline writes the plan to output/EXPERIMENT_PLAN.md. Before Stage 2, verify output/refine-logs/EXPERIMENT_PLAN.md exists; if it does not, copy output/EXPERIMENT_PLAN.md to that location so the deploy skill finds it. Do not relocate the original.

Update memory/MEMORY.md.

---

Stage 2 — Autonomous Experiment Execution (/deploy-experiment)

Goal: Run all experiments described in the plan and collect results.

Invoke the subskill:

/deploy-experiment

The subskill will:

1. Read output/refine-logs/EXPERIMENT_PLAN.md (authoritative plan) and output/refine-logs/FINAL_PROPOSAL.md (context).
2. Run the mandatory local GPU availability check (Step 0 of /deploy-experiment) before any classification. This applies to every pilot or full experiment in this pipeline. If a local GPU (CUDA or MPS) is detected, all ML/DL runs MUST execute on it; only when LOCAL_GPU=none may runs route to remote/Modal/CPU.
3. Classify the experiment into Track A (ML/DL on GPU) or Track B (spatial/GIScience on CPU) — or both for mixed GeoAI.
4. Acquire data via /data-download if any required dataset is missing from data/raw/.
5. Pre-flight environment checks (GPU re-confirmed from Step 0 for Track A, package availability for Track B).
6. Launch runs (long-running jobs go to background or remote screen sessions).
7. For Track B, route execution through /spatial-analysis per claim.
8. Periodically invoke /training-check (Track A with W&B) to monitor for stalls/failures.

Outputs (canonical paths produced by /deploy-experiment):

• output/experiment/EXPERIMENT_RESULT.md — final per-claim results, tables, key numbers
• output/experiment/EXPERIMENT_LOG.md — chronological run log (commands, timings, failures)
• output/experiment/data/ — visualization-ready artifacts (CSV, GeoPackage, predictions, metrics, sidecar metadata)
• output/experiment/figures/ — quick-look figures
• output/experiment/scripts/ — runner scripts created during execution

Wait for runs to finish, then confirm EXPERIMENT_RESULT.md and EXPERIMENT_LOG.md are populated before advancing.

Update memory/MEMORY.md.

---

Stage 3 — Autonomous Review Loop (/auto-review-loop)

Goal: Iteratively improve work quality through adversarial review until all per-criterion floors are met and the weighted score is acceptable, or MAX_ROUNDS = 4 is reached.

Invoke the subskill, passing the chosen idea and difficulty:

/auto-review-loop "[chosen idea title] — difficulty: $REVIEWER_DIFFICULTY"

The subskill reads output/experiment/EXPERIMENT_RESULT.md, output/experiment/EXPERIMENT_LOG.md, output/refine-logs/FINAL_PROPOSAL.md, and prior review state. Each round runs: independent reviewer (Codex MCP / codex exec / Claude subagent fallback) → parse → (optional debate in hard/nightmare) → fix verification → re-evaluate. State is persisted to output/REVIEW_STATE.json at every phase boundary.

Outputs (canonical paths produced by /auto-review-loop):

• output/AUTO_REVIEW_REPORT.md — cumulative review log (one entry per round)
• output/review-rounds/round_<N>_raw.md — verbatim raw reviewer responses
• output/REVIEW_STATE.json — per-round scores, open weaknesses, threadId, status
• output/METHOD_DESCRIPTION.md — concise final method description (consumed by /paper-figure-generate and /paper-draft)
• memory/REVIEWER_MEMORY.md (hard and nightmare only) — persistent reviewer memory across rounds
• output/PROJ_NOTES.md (when COMPACT_MODE = true) — appended one-line discoveries per round

After the loop completes, read the terminal status and final score from output/REVIEW_STATE.json:

• status = "completed" (POSITIVE_THRESHOLD met) → proceed to Stage 4.
• status = "max_rounds_reached" → proceed to Stage 4 with the remaining open_weaknesses flagged as limitations in the narrative.
• status = "blocked" (Phase B.7 circuit breaker, contract violation, or no reviewer backend) → STOP. Surface output/CONTRACT_VIOLATION.md (if any) and the blocker reason. Do not proceed to /generate-report until a human resolves it.

Update memory/MEMORY.md.

---

Stage 4 — Report Writing (/generate-report)

Goal: Consolidate all pipeline artifacts into a single narrative rich enough to drive the downstream paper-writing-pipeline (paper-plan → paper-figure-generate → paper-draft → paper-review-loop → paper-convert).

Invoke the subskill (optionally with a focus hint):

/generate-report

The subskill reads, at minimum:

• output/LIT_REVIEW_REPORT.md
• output/IDEA_REPORT.md and RESEARCH_PLAN.md (or BRIEF.md if that was the active brief)
• output/refine-logs/FINAL_PROPOSAL.md
• output/refine-logs/EXPERIMENT_PLAN.md
• output/experiment/EXPERIMENT_RESULT.md and output/experiment/EXPERIMENT_LOG.md
• output/AUTO_REVIEW_REPORT.md and output/REVIEW_STATE.json
• output/METHOD_DESCRIPTION.md
• output/PROJ_NOTES.md (when COMPACT_MODE = true)
• data/DATA_MANIFEST.md, output/spatial-analysis/, output/figures/ (as needed)

…and produces output/NARRATIVE_REPORT.md — the single source-of-truth narrative for the paper-writing pipeline.

After /generate-report returns, append a short pipeline summary block to output/NARRATIVE_REPORT.md (or write it alongside as output/PIPELINE_SUMMARY.md):

# Research Pipeline Report

**Direction**: $ARGUMENTS (and/or RESEARCH_PLAN.md / BRIEF.md)
**Chosen Idea**: [title]
**Date**: [start] → [end]
**Pipeline**: idea-discovery-pipeline → deploy-experiment → auto-review-loop → generate-report

## Journey Summary
- Ideas generated: X → filtered to Y → piloted Z → chose 1
- Implementation: [brief description of what was built]
- Experiments: [number of runs, total compute time, GPU-hours]
- Review rounds: N/4, final score: X/10, status: [completed|max_rounds_reached|blocked]

## Final Status
- [ ] Ready for paper-writing-pipeline / [ ] Needs manual follow-up

## Remaining TODOs (if any)
- [items from output/REVIEW_STATE.json open_weaknesses]

## Files Changed
- [list of key files created/modified]

Gate 4 (if HUMAN_CHECKPOINT = true): Present output/NARRATIVE_REPORT.md to the user for approval before handing off.

Handoff: Once approved, the next call in the workflow is /paper-writing-pipeline using output/NARRATIVE_REPORT.md as input.

Update memory/MEMORY.md.

---

Recovery Instructions

If context overflows mid-pipeline:

1. Read handoff.json for pipeline.stage, recovery.resume_skill, recovery.read_first (fastest recovery).
2. Read output/REVIEW_STATE.json to find current review round and per-criterion scores (Stage 3 only).
3. Read memory/MEMORY.md for pipeline state flags.
4. Read output/PROJ_NOTES.md (when COMPACT_MODE = true) or output/experiment/EXPERIMENT_LOG.md (full).
5. Resume from the interrupted stage — do NOT re-run completed stages or experiments marked SUCCESS in output/experiment/EXPERIMENT_LOG.md.

---

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.
• Brief precedence: RESEARCH_PLAN.md > BRIEF.md > $ARGUMENTS. The first one with usable content wins; the others provide supplementary scope only.
• Mandatory local GPU check before any pilot or full experiment. Stage 1 (/idea-discovery-pipeline → /generate-idea Phase 5.0) and Stage 2 (/deploy-experiment Step 0) MUST each run the local GPU presence check (nvidia-smi, then MPS) before launching. If a local GPU is detected, all ML/DL runs MUST execute on it; remote/Modal/CPU is only allowed when LOCAL_GPU=none.
• Gate 1 is controlled by AUTO_PROCEED. When false, do not proceed past Stage 1 without user confirmation. When true, the subskill auto-selects the top idea after presenting results.
• Stages 2–4 can run autonomously once the user confirms the idea. This is the "sleep and wake up to results" part.
• Respect the canonical output paths above — do NOT relocate subskill outputs. Downstream skills (and the paper-writing pipeline) read them at fixed locations.
• If Stage 3 ends blocked, stop and surface the blocker. Do not proceed to /generate-report or paper writing.
• If Stage 3 ends max_rounds_reached, proceed to Stage 4 but flag remaining open_weaknesses as limitations.
• Budget awareness: Track total GPU-hours across the pipeline. Flag if approaching user-defined limits.
• Documentation: Every stage updates its own canonical output files. The full history must be self-contained for the next session to recover.
• Generator-evaluator separation (CLAUDE.md): never let the entity that wrote a section score it. /auto-review-loop enforces this — do not bypass it.
• Fail gracefully: If any stage fails (no good ideas, experiments crash, review loop blocked), report clearly and suggest alternatives rather than forcing forward.