Reflect

One job: end the session with an honest written record — what happened (episodic memory) and what to do differently (improvement signal).

This is the heart of self-improvement. A reflection that never gets written, or is written as vague praise, breaks the whole loop.

All vault/… paths below resolve against the vault base (see CLAUDE.md → "Resolving the vault path"); on a clone that's the repo's vault/, on a plugin install it's the setup-recorded path. Resolve the base before writing.

Steps

1. Log the session — append to today's vault/Daily/YYYY-MM-DD.md under ## Sessions:

   ### HH:MM — <short session title>
   
   - Goal: <what this session set out to do>
   - Did: <what actually happened, including links to any [[Knowledge notes]] captured>
   - Outcome: <done / partial / blocked — and why>
   - Open: <what the next session should pick up, or "nothing">

2. Write the reflection — check whether vault/Reflections/YYYY-MM-DD.md already exists and APPEND a session section if it does. Never create it blind: the note is per-day but reflections are per-session, and a create against an existing path silently lands as YYYY-MM-DD 1.md, hiding a parallel session's reflection from every future improve pass.

   ## Session: <title> (HH:MM)
   
   ### What worked
   - <specific behavior that helped>
   
   ### What didn't
   - <specific behavior that hurt — be honest, this is the improvement signal>
   
   ### Lesson (one sentence, actionable)
   <a rule a future session could actually follow>
   
   ### Proposed change
   - target: <skills/<name>/SKILL.md | CLAUDE.md | vault/MEMORY.md | config.yaml | none>
   - change: <one sentence describing the edit>
   - status: proposed

3. Update MEMORY.md — this is the session's FIRST MEMORY.md write (it's read-only until now for prompt-cache stability): mark the goal done or update its one-line status, and add any Knowledge index lines queued by capture. Stale goals rot the index.
4. Commit memory — commit in the vault's own git repo if it has one (<base> = cat ~/.config/loop-and-gate/vault, or the clone's vault/). Stage the paths this session actually wrote — Daily/YYYY-MM-DD.md, Reflections/YYYY-MM-DD.md, MEMORY.md, plus any capture targets — never add -A, which sweeps a parallel session's or an auto-sync's work in under your message: git -C <base> add <those paths> && git -C <base> commit -m "chore: memory update — <short summary>" If the vault isn't under git, skip. Every session that can leaves an auditable trail.

Rules

• Reflect does NOT apply changes. It only proposes them (status: proposed). The improve skill applies them later, when a signal has repeated. This gate is deliberate: unverified self-modification makes agents worse, not better.
• "What didn't" must contain something real. A session where nothing went wrong still had something inefficient. If truly nothing: write "no signal" — but that should be rare.
• One lesson per session, not five. Pick the one that matters most.
• If the same lesson already appears in a previous reflection, say so explicitly: "REPEAT SIGNAL — also in YYYY-MM-DD". Repeats are what authorize the improve skill to act.