workflows-compound

Capture learnings from the current workflow: $ARGUMENTS

Purpose

You are executing Stage 7 (Compound) of the development workflow. Implementation is reviewed and revised. Your job is to extract reusable learnings from this workflow and persist them so future workflows benefit. You are NOT writing code — you are documenting what was learned.

Pre-Conditions

Before starting, verify:

1. .workflow/state.json exists and currentStage is "compound"
2. Review has passed (check stages.review.status is "completed")
3. If revision happened, it's also completed

If pre-conditions are not met, report what's missing and halt.

Process

Step 1: Gather Evidence

Read all workflow artifacts:

1. Workflow state: .workflow/state.json — timeline, iterations, stage notes
2. Sub-agent summaries: .workflow/summaries/*.md — what was built
3. Review report: .workflow/review-report.md — what was found
4. Spec and ADR: The original design documents
5. Git log: What actually changed

git log --oneline --since="$(jq -r .startedAt .workflow/state.json)" -- .

Step 2: Extract Learnings

Identify three categories of learnings:

Solutions — Reusable patterns for solving specific problems:

• What problem was solved?
• What approach worked?
• What approach was tried and didn't work?
• What would you do differently next time?

Discoveries — Things learned about the codebase or domain:

• Unexpected behaviors encountered
• Undocumented constraints discovered
• Performance characteristics measured

Process — What worked or didn't in the workflow itself:

• Which stages were smooth vs. painful?
• Where did revision loops happen and why?
• What spec assumptions were wrong?

Step 3: Write Solution Document

If a reusable solution was produced, write it to docs/solutions/:

# <Problem Title>

## Problem
[What problem this solves, in 2-3 sentences]

## Solution
[The approach that worked, with code references]

## Context
- Workflow: <id>
- Spec: <path>
- Date: <ISO date>

## Key Decisions
- [Decision 1]: [Why this choice was made]
- [Decision 2]: [Why this choice was made]

## What Didn't Work
- [Approach that was tried and abandoned, with brief explanation]

## Related
- [Links to specs, ADRs, or other solutions]

Step 4: Update Agent Memory

If significant patterns or discoveries should persist across sessions:

1. Check existing memory files for related entries
2. Update or add entries as appropriate
3. Include fitness tags (HOT/WARM/COLD) per the DGM calibration rules

Step 5: Record and Handoff

1. Update workflow state (.workflow/state.json):
   • Set stages.compound.status to "completed"
   • Set stages.compound.completedAt to current ISO timestamp
   • Set stages.compound.artifacts.solution to the solution doc path (if created)
   • Set currentStage to "reflection"
   • Update updatedAt
2. Present the handoff:

   COMPOUND COMPLETE
   ==================
   
   Solutions captured: N
   Discoveries: N
   Process notes: N
   Solution doc: <path or "none — no reusable pattern identified">
   
   Next: Stage 8 - Reflection (/workflow-reflection)

What Makes a Good Learning

A learning is worth capturing if it meets ANY of these:

• It would save >30 minutes if encountered again
• It contradicts documentation or common assumptions
• It reveals a non-obvious interaction between components
• It's a pattern that applies beyond this specific feature

A learning is NOT worth capturing if:

• It's specific to this feature with no broader applicability
• It's already documented in the codebase or specs
• It's a trivial fix that anyone would find quickly

Anti-Patterns

• Do NOT capture every detail — focus on what's reusable
• Do NOT write vague learnings like "testing is important" — be specific
• Do NOT skip this stage because "nothing interesting happened" — every workflow teaches something
• Do NOT write code — this is a documentation stage
• Do NOT create solution docs for trivial changes — only for patterns worth reusing