MartinoPolo/mp-harvest-decisions

Harvest Decisions

Scan Claude Code session transcripts for grilling and design-refinement discussions. Extract decisions, categorize them, and update .mpx/CONTEXT.md and .mpx/DECISIONS.md.

Step 1: Locate Sessions

1. Determine the Claude Code projects directory: ~/.claude/projects/
2. Identify the current project's session directory by matching the working directory path
3. Also identify all worktree session directories (pattern: *worktrees* in the project name)
4. Parse $ARGUMENTS as days-back (default 30). Filter to session files modified within that window.

Step 2: Find Grill Sessions

Search session JSONL files for actual skill invocations — not just mentions in available-skills lists.

Primary signal (high confidence): "name":"Skill" paired with "mp-grill" or "gk-design-brief" or "gk-design-refine" in the same file.

Secondary signal (check if primary yields <5 results): Look for AskUserQuestion tool calls with option arrays containing design/architecture decisions.

Record the file path and size for each matching session.

Step 3: Extract Decisions (Parallel)

Spawn Sonnet sub-agents to read session files in parallel. Group ~3-4 sessions per agent.

Each agent's prompt:

Read these JSONL session files. Each line is a JSON object with role and content fields. Find all grill/design discussion segments by locating AskUserQuestion tool calls and the user's responses.

For each decision found, extract:

• Topic: What was being decided (e.g., "issue card color identity", "session chat layout")
• Question: The question that was asked
• Answer: The user's chosen answer
• Rationale: Why this was chosen (if stated)
• Category: One of: Platform, UI-Design, Data-State, Session-Providers, Workflow, Domain-Language

Skip: tool call details, code output, implementation work. Only extract decision Q&A pairs.

Output as a markdown list, one entry per decision.

Step 4: Read Existing Docs

Read .mpx/CONTEXT.md and .mpx/DECISIONS.md to understand what's already documented. If either doesn't exist, create it using the templates from skills/shared/DOCUMENTATION_STRATEGY.md.

Also read any existing memory files in ~/.claude/projects/*/memory/project_*.md for previously summarized decisions.

Step 5: Merge & Deduplicate

With all raw extractions in hand:

1. Categorize by module/domain (Platform, UI, Data, Session, Workflow, Domain Language)
2. Deduplicate — same decision found across multiple sessions keeps the most detailed version
3. Identify conflicts — if two sessions decided differently on the same topic
4. Cross-reference — mark decisions already captured in CONTEXT.md or DECISIONS.md
5. Flag gaps — decisions in sessions not yet in any doc

Step 6: Resolve Conflicts

If conflicts exist, present each one to the user via AskUserQuestion:

Conflict: [topic] Session A (date): chose X because Y Session B (date): chose Z because W Which decision stands?

Step 7: Update Docs

Present a summary of proposed changes before writing:

CONTEXT.md updates:

• New domain terms for § Domain Language — show each candidate entry (**Term** — Definition.) individually and ask the user whether to add it. Write only confirmed terms.
• New or updated feature entries for § Core Features
• New constraints for § Key Constraints

DECISIONS.md updates:

• New decision entries grouped by domain
• Format: ### Title + Decided: date + What: + Why: + Rejected:

Ask user to confirm before writing. Apply edits to existing files (do not overwrite).

Step 8: Clean Up Memory

If decisions were previously stored in memory files (project_prd*_decisions.md, project_*_grilling.md), suggest which memory files are now redundant since their content has been captured in the project docs.

Report

Summarize:

• Sessions scanned / sessions with decisions
• Decisions extracted / new / already documented / conflicts resolved
• Docs updated (CONTEXT.md, DECISIONS.md)
• Memory files suggested for cleanup