Research

/research — end-to-end pipeline conductor

You (the main agent) are the orchestrator. Drive the pipeline below in order. The filesystem is the message bus: pass the run directory path in every subagent task message; expect each subagent to end with {"ok": ..., "outputs": [...], "notes": ...}.

Arguments received: $ARGUMENTS

0. Scaffold the run

python3 .claude/scripts/new_run.py $ARGUMENTS

Capture the printed run id; RUN = workspace/runs/<run-id>. Read RUN/run-config.json for B (branches), I (iterations), offline flag, topic. If new_run.py fails, report and stop.

1. Investigate

Launch the problem-investigator subagent: task message = run directory, topic, offline flag; ask for brief.md + bibliography.jsonl + literature/ + investigation/ per its protocol.

• If it returns ok: false (relevance gate ABORT), report the reason to the user and STOP the pipeline.
• Then: python3 .claude/scripts/bib_validate.py RUN must pass.
• Ledger: python3 .claude/scripts/ledger.py append '{"event":"stage_investigate","detail":"complete"}'

2. Discover — I iterations of Parallel Explore-Exploit

For each iteration iN in i1..iI:

1. Ideate: launch ideator (run dir, iteration, B; from i2 also point it at all previous distilled-feedback.md files — the ideator then applies top-K retention: p1..pK refine the previous iteration's top-K surviving branches). Expect iterations/iN/ideas.md + proposals/p1..pB.md. Ledger event stage_ideate.
2. Solve — PARALLEL: launch ALL B solver subagents IN ONE SINGLE MESSAGE (one Agent call per branch). Each task message: run dir, its own branch dir iterations/iN/branches/bK/, its proposal proposals/pK.md. Never launch solvers sequentially.
3. Evaluate — PARALLEL: after all solvers finish, launch B evaluator subagents in one message (one per branch). Expect eval.json per branch.
4. Audit — PARALLEL: launch B auditor subagents in one message. Expect audit.md per branch.
5. Rank + Select (you, the orchestrator): read every branch's eval.json + audit.md. Disqualify audit-FAIL branches. Write iterations/iN/ranking.md: table (branch, proposal one-liner, score, audit verdict, rank) + selection rationale.
6. Distill feedback (you): invoke the distill-feedback skill (Skill tool, skill: distill-feedback) and write iterations/iN/distilled-feedback.md with exactly its section template — do not improvise the format from memory; the next ideator round consumes this file. Ledger event stage_iteration_iN.

After the last iteration: pick the best surviving (audit-PASS) branch across ALL iterations by official score. Copy its solution.py, eval.json, solve.log, plan.md into RUN/best/ and write RUN/best/SELECTED.json: {"iteration": "iN", "branch": "bK", "score": <score>}. Ledger event stage_discover complete. If ZERO branches survive audits in all iterations, report and stop.

Then launch ablation-analyst (run dir) → best/ablations/{ablations.json, ablation.md}; ledger event stage_ablation. Ablations are secondary evidence — on ok: false, record and continue; never block the pipeline.

3. Write paper

1. Launch paper-writer with stage=conceive → paper/research-representation.md.
2. Run python3 .claude/scripts/ground_check.py RUN — if exit 1, relaunch paper-writer (stage=conceive, "fix every non-SUPPORTED check in paper/ground-report.json") once and re-run ground_check. If it still fails, STOP and show the user the unsupported list — never compose a draft on an ungrounded representation.
3. Launch paper-critic → paper/critic-report.md.
4. Launch paper-writer with stage=resolve-compose → paper/draft.md.
5. Convergence round (Ground→Critic→Resolve, max 2 rounds total): if the critic report contained any BLOCKER, relaunch paper-critic against paper/draft.md. If BLOCKERs remain, relaunch paper-writer (stage=resolve-compose) once more with them, then stop regardless (plateau) and report residual issues honestly. Ledger event stage_write_paper.

4. Verify claims

1. python3 .claude/scripts/extract_claims.py RUN
2. Launch claim-verifier → verdicts in paper/claims.jsonl + paper/verification-report.md. If run-config.json#verifier_votes is N ≥ 2, follow the best-of-N vote protocol in the verify-claims skill instead of a single launch (sequential verifier launches interleaved with verdict_votes.py snapshot, then verdict_votes.py merge).
3. If any claim is FAIL: launch refiner; then re-run python3 .claude/scripts/verify_claims.py RUN; if FAILs remain, one more refiner round (max 2), then accept honestly reported residuals.
4. Copy the final draft: paper/draft.md → final/paper.md.
5. python3 .claude/scripts/chain_of_evidence.py RUN — the four integrity checks. Ledger event stage_verify.

5. Report

Summarize for the user: run id, best branch + official score, grounding ratio (paper/ground-report.json), claim tally, chain-of-evidence table result, and the path RUN/final/paper.md. Mention /paper-to-latex if they want a .tex/PDF.