research-pipeline

Full Research Pipeline: Idea → Experiments → Submission

External cadence is fire-control only. An overnight scheduler may check process/file progress, update a heartbeat, and nudge a stalled phase. It must never rerun or replace a reviewer verdict. Register the state file with watchdog.py, unregister on completion, and use iteration_log.py to trigger structural pivots after repeated no-progress iterations. See external-cadence.md.

End-to-end autonomous research workflow for: $ARGUMENTS

Optional formal proof route

Before creating an experimental run, route theorem-centered requests for Lean to /lean-formalize and report its actual proof result. Run the experimental stages below when the task includes empirical research. In a mixed project, use Lean for a user-requested formal proof or a specific mathematical obligation whose formal verification would materially help; return its checked scope and remaining obligations to the research summary and applicable proof audit. Ordinary mathematics stays in the existing theory skills. Difficulty alone does not make Lean mandatory, and a formal sublemma does not establish an empirical claim.

Constants

• AUTO_PROCEED = true — When true, every selection checkpoint is informational: report the choice and continue in the same turn. When false, ask for explicit user confirmation and end the turn at the checkpoint.
• ARXIV_DOWNLOAD = false — When true, /research-lit downloads the top relevant arXiv PDFs during literature survey. When false (default), only fetches metadata via arXiv API. Passed through to /idea-discovery → /research-lit.
• HUMAN_CHECKPOINT = false — When true, the auto-review loops (Stage 3) pause after each round's review to let you see the score and provide custom modification instructions before fixes are implemented. When false (default), loops run 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 repo directly via codex exec + memory + debate. Passed through to /auto-review-loop.
• CODE_REVIEW = true — GPT-6-Astra xhigh reviews experiment code before deployment. Catches logic bugs before wasting GPU hours. Set false to skip. Passed through to /experiment-bridge.
• BASE_REPO = false — GitHub repo URL to use as base codebase. When set, /experiment-bridge clones the repo first and implements experiments on top of it. When false (default), writes code from scratch or reuses existing project files. Passed through to /experiment-bridge.
• COMPACT = false — When true, generates compact summary files for short-context models and session recovery. Passed through to /idea-discovery and /experiment-bridge.
• AUTO_WRITE = false — When true, automatically invoke Workflow 3 (/paper-writing) after Stage 4. VENUE is needed only when Stage 5 begins — a missing venue defers paper writing; it never blocks Stages 1-4. When false (default), Stage 4 generates NARRATIVE_REPORT.md and stops — user invokes /paper-writing manually.
• VENUE = (unset) — Target venue for paper writing; bound only when Stage 5 begins. Options: ICLR, NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_CONF, IEEE_JOURNAL. No default: a missing venue defers paper writing — it never blocks Stages 1-4 and is never guessed.
• RENDER_HTML = true — When true (default), auto-render NARRATIVE_REPORT.md to HTML at Stage 4 completion via /render-html. Uses --no-review because Stage 3 already produced a traced same-family provisional review. Set false to skip. Rendering failure is non-blocking.
• RESUMABLE = true — Record per-stage state under .aris/runs/ and resume from the first non-terminal phase. Same-family Codex review produces provisional; deterministic or overlay gates produce accepted.

💡 Override via argument, e.g., /research-pipeline "topic" — AUTO_PROCEED: false, human checkpoint: true, difficulty: nightmare, code review: false, base repo: https://github.com/org/project, auto_write: true, venue: NeurIPS.

Checkpoint execution rule

Resolve AUTO_PROCEED once from $ARGUMENTS before Stage 1 and pass that resolved value to nested workflows.

• AUTO_PROCEED=true is non-blocking. A checkpoint is a progress update, not a question. State the result and the automatically selected next action, then continue executing in the same turn. Do not ask for confirmation, request user input, sleep, wait for silence, or end the turn at a checkpoint.
• AUTO_PROCEED=false is blocking. Present the options, ask the user, and end the turn. Resume only after an explicit reply.

Never implement auto-proceed as “ask, then continue if there is no response.” Once a turn ends, silence cannot resume the pipeline. The user can still interrupt a non-blocking run at any time.

This rule governs only AUTO_PROCEED-controlled selection checkpoints. If the user explicitly enables a Feishu interactive gate, that external approval or reply is an intentional blocking exception; wait for that user-controlled gate rather than treating it as a silence timeout. Feishu off/push-only modes remain non-blocking under AUTO_PROCEED=true.

Overview

This skill chains the entire research lifecycle into a single pipeline:

/idea-discovery → /experiment-bridge → /auto-review-loop → /paper-writing (optional)
├── Workflow 1 ──┤├── Workflow 1.5 ──┤├── Workflow 2 ───┤ ├── Workflow 3 ──┤

It orchestrates up to four major workflows in sequence. Workflow 3 (paper writing) is optional and controlled by AUTO_WRITE.

Resumable runs and heartbeat

When RESUMABLE=true, resolve helpers through the Codex manifest:

if [ -z "${

Warn-and-skip state tracking if RUN_STATE cannot be resolved; never pretend it was persisted. Phases are idea-discovery,experiment-bridge,auto-review-loop,summary,paper-writing.

• New run: python3 "$RUN_STATE" start . "$RUN_ID" --executor codex-gpt-6-astra --provisional-advances --phases "idea-discovery,experiment-bridge,auto-review-loop,summary,paper-writing" (the --provisional-advances policy is what lets a same-family provisional verdict close a phase for resume — without it, mainline semantics apply and provisional phases stay open).
• Resume: python3 "$RUN_STATE" resume . "$RUN_ID"; restart the returned phase.
• Each phase: mark running, then done --artifact <path>.
• A fresh Codex reviewer PASS uses mark-provisional --reviewer gpt-6-astra --verdict-id <trace-or-agent-id>. This is terminal for resume but not accepted.
• A cross-family overlay or deterministic verifier uses accept.
• If AUTO_WRITE=false, mark paper-writing as skipped after summary.

For an unattended loop, touch the run state at the start of every tick, register it once with watchdog.py --register as type loop, and unregister on completion. After each tick run iteration_log.py note <root> <run_id> <phase> <new-finding-count>: pivot=structural requires a genuinely different approach; pivot=human surfaces the stall. Neither result is a quality verdict. See resumable-runs.md.

Pipeline

Stage 1: Idea Discovery (Workflow 1)

If RESEARCH_BRIEF.md exists in the project root, it will be automatically loaded as detailed context (replaces one-line prompt). See templates/RESEARCH_BRIEF_TEMPLATE.md.

Invoke the idea discovery pipeline:

This internally runs: /research-lit → /idea-creator → /novelty-check → /research-review

Output: idea-stage/IDEA_REPORT.md with ranked, validated, pilot-tested ideas.

Review Tracing follows the downstream review skills. Stage 1 and Stage 3 preserve reviewer prompts/responses through their own trace protocols so the final handoff can be audited.

🚦 Gate 1 — Idea Selection:

After idea-stage/IDEA_REPORT.md is generated, present the top ideas.

If AUTO_PROCEED=true (non-blocking): report the selection and continue immediately in the same turn. Do not phrase the update as a question:

If AUTO_PROCEED=false (blocking): present the same ranking, ask Recommended: Idea 1. Shall I proceed with implementation?, then end the turn. The user may:

• Approve the idea → proceed to Stage 2. /experiment-bridge reads refine-logs/EXPERIMENT_PLAN.md already generated by /idea-discovery.
• Request changes (e.g., "combine Idea 1 and 3", "focus more on X") → update the idea prompt with user feedback, re-run /idea-discovery with refined constraints, and present again.
• Reject all ideas → collect feedback on what's missing, re-run Stage 1 with adjusted research direction. Repeat until the user commits to an idea.
• Stop here → save current state to idea-stage/IDEA_REPORT.md for future reference.

⚠️ This gate waits for user confirmation when AUTO_PROCEED=false. When true, it auto-proceeds after presenting results. The rest of the pipeline (Stages 2-3) is expensive (GPU time + multiple review rounds), so set AUTO_PROCEED=false if you want a final review checkpoint before committing GPU resources.

Stage 2: Experiment Bridge (Workflow 1.5)

Once the idea is selected (automatically or by the user), delegate implementation and deployment to /experiment-bridge:

💡 Queue routing is automatic: /experiment-bridge Phase 4 routes each milestone by job count — ≤5 jobs → /run-experiment, ≥10 jobs or teacher→student phase dependencies → /experiment-queue (with OOM retry, wave gating, crash-safe state). No manual override is needed.

What this does (fully autonomous):

1. Parses refine-logs/EXPERIMENT_PLAN.md — extracts milestones, run order, compute budget
2. Implements experiment code — extends pilot to full scale, follows existing codebase conventions
3. Fresh-agent code review — GPT-6-Astra xhigh reviews the implementation in a new context; base result is same-family provisional
4. Sanity check — runs the smallest experiment first to verify the environment; auto-debugs failures (up to 3 attempts, with /codex:rescue fallback)
5. Deploys full experiments — auto-routes by job count (≤5 → /run-experiment, ≥10 → /experiment-queue with OOM retry, wave gating, crash-safe state)
6. Collects initial results — parses outputs, updates refine-logs/EXPERIMENT_TRACKER.md, runs /training-check if W&B is configured
7. Auto-plans ablations via /ablation-planner if main results are positive

Output:

• refine-logs/EXPERIMENT_RESULTS.md — structured results by milestone
• refine-logs/EXPERIMENT_TRACKER.md — updated run-by-run status
• EXPERIMENT_LOG.md (when COMPACT=true) — session-recovery-friendly log

Monitor progress (while experiments run):

Wait for /experiment-bridge to complete and report its handoff summary before proceeding.

Stage 3: Auto Review Loop (Workflow 2)

Once initial results are in, start the autonomous improvement loop:

What this does (up to 4 rounds):

1. GPT-6-Astra xhigh reviews the work (score, weaknesses, minimum fixes)
2. Claude Code implements fixes (code changes, new experiments, reframing)
3. Deploy fixes, collect new results
4. Re-review → repeat until score ≥ 6/10 or 4 rounds reached

Output: review-stage/AUTO_REVIEW.md with full review history and final assessment.

Stage 4: Research Summary & Writing Handoff

After the auto-review loop completes, prepare the handoff for paper writing.

Step 1: Write a final research status report (same as before).

Step 2: Generate NARRATIVE_REPORT.md from:

• IDEA_REPORT.md (chosen idea, hypothesis, novelty justification)
• Implementation details from the repo
• Experiment configs and final results
• AUTO_REVIEW.md (review history, weaknesses fixed, remaining limitations)

The narrative report must contain:

• Problem statement and core claim
• Method summary
• Key quantitative results with evidence for each claim
• Figure/table inventory (which exist, which need manual creation)
• Limitations and remaining follow-up items

Output: NARRATIVE_REPORT.md + research pipeline report.

Stage 5 / Stage 6: Paper Writing (Workflow 3 — Optional)

This is the Stage 6: Paper Writing handoff in the broader research lifecycle; it is numbered Stage 5 here because this consolidated pipeline counts the writing handoff after the Stage 4 narrative report.

Skip this stage if AUTO_WRITE=false (default). Present the /paper-writing command for manual use:

If AUTO_WRITE=true:

🚦 Gate 2 — Writing Checkpoint:

Checks before proceeding (venue binds HERE — Stages 1-4 are venue-independent):

• If VENUE is missing: with AUTO_PROCEED=false, ask now. With AUTO_PROCEED=true, do not guess and do not wait — stamp "VENUE NOT SPECIFIED — paper writing deferred" in the report and checkpoint, leave the paper-writing phase pending, and finish the run cleanly for a later resume. Never silently pick a venue.
• If manual figures are required: with AUTO_PROCEED=false, pause and list them. With AUTO_PROCEED=true, record "paper writing deferred (manual figures: )" and finish cleanly the same way.

Then invoke:

Pass the resolved AUTO_PROCEED explicitly so Workflow 3 cannot silently fall back to its own default mode.

This delegates to Workflow 3 which handles its own phases: /paper-plan → /paper-figure → /paper-write → /paper-compile → /auto-paper-improvement-loop

When Workflow 3 finishes, update the pipeline report with:

• Paper writing completion status
• Final PDF path (paper/main.pdf)
• Improvement scores (round 0 → round N)
• Remaining issues

Output: paper/ directory with LaTeX source, compiled PDF, and PAPER_IMPROVEMENT_LOG.md.

Render HTML view (auto, when RENDER_HTML = true)

After Stage 4 finalizes NARRATIVE_REPORT.md (before paper writing branches), invoke /render-html on the narrative report:

--no-review is intentional: this is an internal handoff doc, not reviewer-facing — the claims already received a traced same-family provisional review in Stage 3. Output: NARRATIVE_REPORT.html next to the MD, with embedded source SHA256.

Non-blocking: if /render-html fails (helper missing, file write error, etc.), log the failure and continue Stage 4 — the HTML view is a convenience artifact, not a pipeline prerequisite.

Skip this step if RENDER_HTML = false.

Output Protocols

Follow these shared protocols for all output files:

• Output Versioning Protocol — write timestamped file first, then copy to fixed name
• Output Manifest Protocol — log every output to MANIFEST.md
• Output Language Protocol — respect the project's language setting

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.

• The Stage 1 checkpoint is controlled by AUTO_PROCEED. When false, do not proceed without user confirmation. When true, report the top selection and continue in the same turn without asking or waiting.

• Stages 2-3 can run autonomously once the idea is selected. This is the "sleep and wake up to results" part.

• If Stage 3 ends at round 4 without positive assessment, stop and report remaining issues. Do not loop forever.

• Budget awareness: Track total GPU-hours across the pipeline. Flag if approaching user-defined limits.

• Documentation: Every stage updates its own output file. The full history should be self-contained.

• Fail gracefully: If any stage fails (no good ideas, experiments crash, review loop stuck), report clearly and suggest alternatives rather than forcing forward.

Typical Timeline

Sweet spot: Run Stage 1 in the evening, launch Stage 2-3 before bed, wake up to a reviewed paper.