Reflect

End-of-session reflection. Re-read the current conversation, surface what slipped through and what would be lost, and act on what the user chooses — all before the session wraps.

Scope and stance

• Conversation-scoped. Reflect over the conversation that is in context. Do NOT parse transcript files or git history. This is not gstack-retro (commit-history retro) and not gstack-learn (learnings store) — stay in the forward-action lane.
• Assume the session dies after this. Every fact established here that isn't written to a file, an issue, or a PR is gone. That makes homeless knowledge the headline job of this skill, not a footnote: a hard-won gotcha that never reaches CLAUDE.md costs the next session the same debug loop.
• Compaction honesty. If the session was compacted, reflect on what survived and say so plainly. Do not fabricate recall of summarized turns.
• No artifact by default. Findings live in the conversation. The surfaced list is throwaway — the user reacts or ignores it. Write a file ONLY if the user picks an action that writes one, or explicitly asks.
• Under-report, never spam. A wrong "loose end" or a skill suggestion for one-off work is worse than silence. Apply the confidence bars below. When in doubt, drop it.
• Verify before you propose. Never claim something is undocumented or untracked without checking. Read the destination file (CLAUDE.md, AGENTS.md, README.md) and search open issues first, so what you surface is a real edit against real current content — not a duplicate of a line that's already there.

Step 1 — Reflect

Re-read the conversation end to end. Hold the whole arc in mind before flagging anything.

Step 2 — Run the detectors