distill

name	distill
description	Distill the current session into a rich, scoped memory entry — for Codex sessions, append to ~/.codex/memories/rollout_summaries/ in the established task-group format; for Claude sessions, write to the auto-memory directory. Use when you've finished a non-trivial unit of work that future-you (or another agent) should benefit from.
allowed-tools	Bash, Read, Write, Glob, Grep

Why this exists

Both Claude Code and Codex have built-in memory, but Codex's ~/.codex/memories/MEMORY.md follows a richer pattern than Claude's default auto-memory: scoped task groups with explicit applies_to / reuse_rule, user preferences with citations, reusable knowledge as bulleted facts, and "failures and how to do differently" sections. This skill produces that shape on demand, regardless of which harness you're in.

Task

1. Identify session context:

   • Current working directory: pwd
   • Branch and recent commits: git rev-parse --abbrev-ref HEAD && git log --oneline -10
   • Detect harness from environment: if $CLAUDE_PROJECT_DIR set → Claude; if $CODEX_HOME set → Codex. Default to Claude when ambiguous.

2. Identify what's worth distilling (skip if none apply):

   • User preferences expressed during the session ("I prefer X over Y", corrections to your earlier behavior)
   • Reusable knowledge about this checkout (file locations, build commands, conventions that aren't obvious from the code)
   • Failures and recoveries (something didn't work, here's what did)
   • Non-obvious gotchas worth flagging next time

   If nothing meets the bar, report "nothing material to distill" and stop. Over-saving is worse than under-saving.

3. Write the entry in this exact shape, appended to the appropriate target file:

   # Task Group: <cwd> <short topic>
   
   scope: <one sentence on what this group covers>
   applies_to: cwd=<cwd>; reuse_rule=<checkout-specific|broadly-reusable> for <reason>
   
   ## Task <N>: <short title>, <success|partial|failed>
   
   ### keywords
   
   - <comma-separated retrieval anchors>
   
   ## User preferences
   
   - when <situation>, the user <preference quote or paraphrase> -> <how to apply next time>. [Task <N>]
   
   ## Reusable knowledge
   
   - <fact about this checkout>. [Task <N>]
   
   ## Failures and how to do differently
   
   - Symptom: <what went wrong>. Cause: <root cause>. Fix: <what to do next time>. [Task <N>]
   

   Sections may be omitted if they have no content for this distillation.

4. Target file selection:

   • Codex: append to ~/.codex/memories/MEMORY.md (or create a per-session rollout summary at ~/.codex/memories/rollout_summaries/$(date -u +%Y-%m-%dT%H-%M-%S)-<slug>.md and link from MEMORY.md).
   • Claude: write under ~/.claude/projects/$(pwd | sed 's|/|-|g')/memory/ as individual <slug>.md files using the auto-memory frontmatter format (see ~/.claude/CLAUDE.md), and add one-line entries to that project's MEMORY.md index.

5. Convert relative dates in the content to absolute dates (Thursday → 2026-MM-DD).

6. Report: which file you wrote to, how many entries, and one-sentence summary of each.

Constraints

• Do not save code patterns, conventions, or file paths that can be derived by reading the project — those rot. Save WHY and HOW-NEXT-TIME, not WHAT-EXISTS.
• Do not save in-progress work or current-conversation state — those belong in a plan or task list, not memory.
• Update or remove existing entries that this session contradicted rather than stacking conflicting facts.