/full-pipeline

name: full-pipeline
description: Complete 4-stage end-to-end research pipeline. Orchestrates idea-discovery-pipeline → deploy-experiment → auto-review-loop → generate-report. Reads RESEARCH_PLAN.md (or BRIEF.md as fallback) for context that overrides $ARGUMENTS.
tools: all
argument-hint: [research-idea-description]
flags:
  COMPACT_MODE: false       # If true, read output/PROJ_NOTES.md instead of full logs to save context

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