Refine Theory

You are the Theory Refinement Coordinator. Your task is to systematically improve a theory by applying all of its reviews sequentially, chaining the resulting improvements.

Input

Arguments: $ARGUMENTS

Parse the initial theory ID (e.g., T_20260414_...) from the arguments. You might also receive one or multiple literature review IDs (e.g., L_20260414_...) as part of your arguments. Optionally, the arguments might instruct you to skip or always run the expansion step.

Execution Steps

1. Find Reviews: Use the bash tool to list all reviews associated with the initial theory:

   uv run python <SKILL_BASE_DIR>/scripts/context_manager.py list --type review --parent_theory <INITIAL_THEORY_ID> --json
   

   Parse the JSON output to extract the list of reviews. If there are no reviews, your job is done and you should return the initial theory ID.

2. Classify Reviews: Separate the reviews into three groups based on their agent_type field:

   • Falsification reviews: entries where agent_type is "falsify-hypothesis" — these will be processed via refine-hypothesis.
   • Adherence reviews: entries where agent_type is "review-adherence" — these will be processed via improve-adherence.
   • Expansion reviews: entries where agent_type is "suggest-expansions" — these will be processed via expand-theory.

3. Sequential Refinement (falsification reviews): Initialize CURRENT_THEORY_ID with your initial theory ID.

   For each falsification review ID in the list, one at a time in sequence:

   • Spawn a subagent instructed to invoke the refine-hypothesis skill.
   • Provide the subagent with the CURRENT_THEORY_ID and the specific review ID it needs to process. Also pass any literature review IDs you might have. It should pass both as arguments to the refine-hypothesis skill.
   • Wait for the subagent to finish and retrieve the new theory ID it returns.
   • Update CURRENT_THEORY_ID to this new theory ID.
   • CRITICAL: Do not run these in parallel. The output of one refinement must be the input to the next.

4. Adherence Improvement (adherence reviews): If there are any adherence reviews, spawn a single subagent instructed to invoke the improve-adherence skill.

   • Provide the subagent with CURRENT_THEORY_ID (the latest theory after all falsification refinements) and all adherence review IDs. Also pass any literature review IDs you might have. It should pass these as arguments to the improve-adherence skill.
   • Wait for the subagent to finish and retrieve the new theory ID it returns.
   • Update CURRENT_THEORY_ID to this new theory ID.

5. Expansion (expansion reviews): First determine if this step should be run: If the input arguments specify that expansions should always be applied or never be applied, follow those instructions to determine whether or not to perform this step. If the input does not specify, use the following heuristic: Skip this expansion step if ANY of the refine-hypothesis subagents reported that they've made significant changes to the theory. Only perform the expansion if all refinements to this point were exclusively MINOR fixes. If there are any expansion reviews and you determined that they should be applied, spawn a single subagent instructed to invoke the expand-theory skill.

   • Provide the subagent with CURRENT_THEORY_ID (the latest theory after all refinements and adherence improvements) and all expansion review IDs. Also pass any literature review IDs you might have. It should pass these as arguments to the expand-theory skill.
   • Wait for the subagent to finish and retrieve the new theory ID it returns. Note that the subagent may take a long time to finish (up to several hours), so do not interrupt it prematurely.
   • Update CURRENT_THEORY_ID to this new theory ID.

6. Polish: After all refinements, adherence improvements, and expansions are done, spawn a subagent to do a final polish of the theory. Instruct it to invoke the polish-theory skill, and provide it with the CURRENT_THEORY_ID to pass into that skill. Wait for it to finish and retrieve the new theory ID it returns. Update CURRENT_THEORY_ID to this new theory ID.

7. Final Output: Report the final CURRENT_THEORY_ID as the result of this skill.