synthesize

name	synthesize
description	Read all per-framework analysis fragments in a decision workspace and produce a cross-framework synthesis — agreement/disagreement matrix, score table, weighted verdict, and the open questions any individual framework couldn't resolve. Run after /decision:analyze, or invoke directly to re-synthesize.

Synthesize

Aggregates per-framework fragments into one document.

Inputs

A workspace folder produced by /decision:analyze, with decision.md and a populated fragments/ directory. The user may pass the path or the slug; if neither, list recent workspaces and ask.

Procedure

1. Read decision.md (for context) and every file in fragments/.

2. Extract the score (0-100) from each fragment. Each framework guide instructs the subagent to end with ## Score: <N>.

3. Build the score table:

   Framework	Score	One-line verdict

4. Build the agreement matrix. Identify points where frameworks converge (e.g., "5/7 frameworks flag stakeholder-A as a blocker") and where they diverge (e.g., "MCDA prefers Option A; pre-mortem prefers Option B"). Convergence is signal; divergence is where the real work is.

5. Surface unresolved questions — what did multiple frameworks list as "would need to know X to decide"? These are the things to investigate before committing.

6. Weighted verdict. Compute a weighted average of the framework scores. By default, weight each framework equally. If the user has indicated a bundle bias (e.g., "personal" bundle suggests they care about regret-minimization more), reflect that. Always show both the equal-weighted and (if applicable) bias-weighted result.

7. Write synthesis.md in the workspace root. Structure:

   # Synthesis: <Decision title>
   
   ## Score table
   | Framework | Score | Verdict |
   
   ## Convergence
   <where frameworks agree>
   
   ## Divergence
   <where they disagree, and why>
   
   ## Unresolved questions
   <items that need investigation before committing>
   
   ## Weighted verdict
   - Equal-weighted: <score>/100, leaning <option/direction>
   - <Other weightings if applicable>
   
   ## Recommendation
   <one paragraph: what the integrated picture says, and what the user should do next>
   
   ## Confidence
   <how robust the recommendation is — wide framework agreement vs split>
   

8. Report the synthesis path back to the user.

Notes

• Don't override the per-framework analyses — synthesis is an aggregation, not a re-evaluation.
• If frameworks badly conflict (>30 point spread between top and bottom), say so explicitly and recommend the user examine the divergent fragments rather than blindly trusting the average.