chemqa-review

Use when an agent needs the ChemQA react_reviewed control protocol as an installable multi-agent skill bundle layered on top of a sibling debateclaw-v1 skill, with fixed proposer plus four reviewer la

SKILL.md

GitHub

ChemQA Review

This directory is an installable skill bundle.

Treat the directory containing this SKILL.md as <skill-root>. This bundle must be installed beside these sibling skill bundles under the same skills/ root:

  • debateclaw-v1
  • paper-retrieval
  • paper-access
  • paper-parse
  • rdkit
  • pubchem
  • opsin
  • chem-calculator

Purpose

chemqa-review recreates ChemQA's react_reviewed control protocol as a portable skill bundle:

  • one fixed main proposer lane
  • four fixed reviewer lanes
  • DebateClaw V1 handles the low-level debate transport
  • this bundle injects ChemQA-specific role constraints, prompt assets, and artifact reconstruction
  • prompt routing can hand deterministic chemistry subtasks to sibling provider skills for numeric and structure-heavy questions

Root Rule

Always pass --root <skill-root> when a script accepts --root.

Do not use repo-relative paths such as ../debateclaw-v1. Resolve sibling skills from <skill-root>/../.

Standard Flow

1. Check runtime and sibling skill availability

python3 <skill-root>/scripts/check_runtime.py \
  --skill-root <skill-root> \
  --json

2. Compile and materialize a launch-ready run

python3 <skill-root>/scripts/launch_from_preset.py \
  --root <skill-root> \
  --preset chemqa-review@1 \
  --goal "Question: does Pt/C improve HER activity in 1 M KOH?" \
  --launch-mode print

3. Rebuild react_reviewed-style artifacts after the debate completes

python3 <skill-root>/scripts/collect_artifacts.py \
  --skill-root <skill-root> \
  --source-dir <run-output-dir> \
  --output-dir <artifact-dir> \
  --json

Role Topology

  • debate-coordinator: protocol coordinator and final artifact aggregator
  • proposer-1: main ChemQA proposer
  • proposer-2: search_coverage reviewer
  • proposer-3: evidence_trace reviewer
  • proposer-4: reasoning_consistency reviewer
  • proposer-5: counterevidence reviewer

Important:

  • Only proposer-1 is allowed to own the candidate submission.
  • Reviewer lanes must not drift into independent final-answer proposals.
  • The coordinator must emit chemqa_review_protocol.json so the artifact collector can rebuild the react_reviewed protocol surface.

Prompt Routing Notes

  • Numeric questions should prefer chem-calculator before web search when the prompt already provides the needed givens.
  • Structure questions should extract available SMILES or name text first, then route to rdkit, opsin, and pubchem as appropriate.
  • Reviewer lanes should cite provider result JSON artifact paths or structured tool_trace entries when challenging numeric or structural claims.
  • This phase does not include a dedicated image-reading or OCSR skill.

References

  • Runtime checks and bridge behavior live in scripts/.
  • Prompt contracts live in prompts/contracts/.
  • Shared policy and artifact requirements live in prompts/modules/.
  • The fixed control defaults live in control/config-snapshots/react-reviewed-default.json.