Reflect

Mine the current conversation for durable learnings, then route them into skill edits.

When to invoke

Run when the user says "reflect" or "/reflect". If the conversation is trivial, off-topic, or already covered by a skill that the parent followed correctly, report that there is no durable learning and stop. One-offs are not learnings.

Process

1. Locate the active transcript

Use the runtime's current-session transcript API or session store when it exposes one. Resolve only the transcript for the current project and session. Do not scan unrelated projects or sessions. Confirm that the opening user message matches this session. If the runtime does not expose a safe transcript path, write a tight digest of the active conversation and pass that instead. This step is complete when all reviewers can read the same session evidence.

2. Spawn three reviewers in parallel

Start three general-purpose review subagents in parallel when the runtime permits it. Use full-tool agents because reviewers can need read access to context named in the session. The prompt forbids writes; the parent applies edits. If parallel dispatch is unavailable, run the three subagents in sequence. If subagents are unavailable, run three separate manual review passes yourself against the same transcript evidence and keep the outputs separate by lens.