undrift-full

Undrift Full — Session Pattern Mining → CLAUDE.md Rules

You are acting as a meta-learning agent for this project, running entirely in the main agent — no subagents. Your mission is to scan past sessions, extract recurring corrections, and distil them into durable CLAUDE.md rules — so the same mistakes never have to be corrected twice.

Step 0 — Locate & Scope Sessions

1. Run pwd to identify the current project directory
2. If $ARGUMENTS is provided, treat it as a path override for the target project
3. Validate the path: resolve it to an absolute path and confirm it exists inside ~/.claude/projects/. If it does not match any known project directory, stop immediately and tell the user — do not proceed.
4. Locate the matching session directory under ~/.claude/projects/
5. List all available sessions, sorted by modification date (newest first)
6. Cap scope at the last 50 sessions. If fewer than 10 exist, proceed and note the limited sample size.
7. Report: "Found N sessions for [project]. Analysing the most recent 50."

Step 1 — Filter + Extract

Run the following bash loop in the main agent. It processes sessions newest-first, skips sessions with fewer than 3 text-based user turns, extracts user text content, and stops when the global 2000-line cap is reached.

Verified JSONL structure: each line in a session file is a JSON object. User messages have "type": "user" with a nested "message" object whose "content" is either a plain string or an array of content blocks. Text turns have content blocks with "type": "text"; tool results have "type": "tool_result" and do not count as user turns.

lines=0
for f in $(ls -t <session_dir>/*.jsonl 2>/dev/null | head -50); do
  # Count text-based user turns only (excludes tool results)
  turn_count=$(jq -r 'select(.type == "user") |
    .message.content |
    if type == "string" then "T"
    elif (map(select(.type == "text")) | length) > 0 then "T"
    else empty
    end' "$f" 2>/dev/null | wc -l | tr -d ' ')
  [ "${turn_count:-0}" -lt 3 ] && continue
  # Extract user text content
  extracted=$(jq -r 'select(.type == "user") |
    .message.content |
    if type == "string" then .
    else (.[]? | select(.type == "text") | .text)
    end' "$f" 2>/dev/null)
  new_lines=$(echo "$extracted" | wc -l)
  if [ $((lines + new_lines)) -gt 2000 ]; then
    echo "CAP_HIT:$(basename $f)"
    break
  fi
  echo "=== SESSION: $(basename $f) ==="
  echo "$extracted"
  lines=$((lines + new_lines))
done