/session-insights — Extract insights from past sessions

Parse .jsonl session transcripts and extract structured insights: problems, decisions, patterns, friction points. Works on sessions outside the current context.

Step 1: Find sessions

Resolve memory directory, then find sessions by argument:

# Transcripts are stored per launch cwd (a monorepo subdir gets its own dir), memory per git root
SESSIONS_DIR="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/projects/$(pwd | sed 's/[^a-zA-Z0-9]/-/g')/"

• Ticket/keyword — grep across all .jsonl files
• Session ID — specific file by UUID
• Workstream — use node "${CLAUDE_PLUGIN_ROOT}/scripts/memory.js" workstream <name> for context, match by dates/keywords
• No argument — most recent session

grep -l "<query>" "$SESSIONS_DIR"*.jsonl 2>/dev/null | head -10

Also check session index if available:

cat "$(node "${CLAUDE_PLUGIN_ROOT}/scripts/memory.js" dir)/sessions.jsonl" 2>/dev/null

Step 2: Extract user messages

For each session found:

import json, sys

with open(sys.argv[1]) as f:
    for i, line in enumerate(f):
        try:
            obj = json.loads(line.strip())
            if obj.get('type') != 'user':
                continue
            c = obj.get('message', {}).get('content', obj.get('content', ''))
            if isinstance(c, list):
                for item in c:
                    if isinstance(item, dict) and item.get('type') == 'text':
                        text = item.get('text', '').strip()
                        if text and not text.startswith('<') and len(text) > 5:
                            print(text[:300])
            elif isinstance(c, str):
                text = c.strip()
                if text and not text.startswith('<') and len(text) > 5:
                    print(text[:300])
        except:
            pass

Step 3: Analyze

Extract structured insights from the user messages:

1. Problems — what didn't work, was redone, caused friction
2. Decisions — what was chosen and why (architectural, process)
3. Insights — non-obvious conclusions, patterns, reassessments
4. Friction points — where the agent repeatedly failed, what required many iterations

Format:

## Session insights: <date range or session id>

### Problems
- <description> → <resolution or status>

### Decisions
- <what> — <why>

### Insights
- <insight> — <evidence from session>

### Friction
- <what> (×N iterations) — <root cause>

Step 4: Save

MUST present all options as a numbered menu and wait for user response before writing anything:

What to do with these insights?
  1. Save insights to memory (feedback/ or decisions/)
  2. Add friction points to backlog
  3. Mark as DOC: for /docs-reflect
  4. Skip — display only

Choose (1-4, or comma-separated):

Then execute only what the user chose.

Rules

• Filter system messages, ide_opened_file, tool calls — only USER text
• Large sessions (>50KB user text) — split into chunks
• Don't dump raw output — format into readable blocks
• If multiple sessions found for a query — combine chronologically
• If session index (sessions.jsonl) exists — use it for faster lookup