Reflection

# Reflection Prompt

> Usage: at the end of a substantial session, send the following to the agent (or wrap it as the injected content of an enforced hook).
> It scans signals by confidence, dedupes against existing memory, and outputs a diff proposal — nothing is written without approval.

---

The substantial work of this session is done. Now run a reflection. Follow every step, don't skip:

1. Signal scan: review the interaction and list candidate signals with confidence —
   - High (user corrections: "no", "use X instead", "from now on always...")
   - Medium (user confirmations: "great", "just like that", "exactly what I wanted")
   - Low (patterns that recur but were never stated outright)

2. Dedupe: compare each signal against existing entries in memory/INDEX.md and label:
   - [new] a brand-new lesson -> draft a new memory file
   - [confirms] re-verification of existing memory -> +1 to that entry's verified_count
   - [conflict] contradicts existing memory -> do NOT edit; list it separately for the user to decide

3. Output a proposal: present everything to be written/modified as a diff, and wait for confirmation.
   Do not write to any memory file without approval.

4. If this session has nothing worth keeping, say "nothing to record this session" and explain why.
   Do not manufacture fake signals just to produce output.