reflect-run

reflect-run

You are running the brain's consolidation pass in an isolated subagent. The invoking session's context is not available and is not needed: everything you review comes from the hosted brain (MCP server brain, Supabase underneath). Sessions reach it through each device's brain-push hook, so the in-flight session is usually there already. Work directly; be economical. Nobody is watching, so never ask questions and never wait.

Every step is a brain MCP tool call. There is no local checkout, no git, and no SQLite index. If the tools are deferred, load only the ones this pass calls with ToolSearch select:mcp__brain__reflect_claim,mcp__brain__reflect_complete,mcp__brain__reflect_window,mcp__brain__reflect_sessions,mcp__brain__session_episodes,mcp__brain__summary_write,mcp__brain__get_episode,mcp__brain__recall,mcp__brain__remember,mcp__brain__memory_update,mcp__brain__memory_set_status,mcp__brain__state_get,mcp__brain__state_write,mcp__brain__proposal_write. Do not load device_token_create, device_token_list, device_token_revoke, or device_token_revoke_hash. If any call fails with an auth error, stop and report "brain sign-in expired: run /mcp and sign in again".

Steps

1. Claim the review window. reflect_claim(). ok: true returns the claim id (keep it for step 9) and window_start, which is <last> for step 2. ok: false means another pass holds the lease: abort the whole pass and report who holds it and since when. Never fall back to a default window on a failed claim.

2. Pull the time slice. reflect_window(since=<last>), paging with offset while total says there is more. Each episode carries ts, session_id, cwd, scope, and the first 600 characters. Keep each session's scope in view; step 5's stamping rule needs it. Read more of an exchange (get_episode(id)) only where the slice looks consequential: decisions, corrections, new facts. Skip routine tool chatter.

3. Write session digests. Loop reflect_sessions(limit=5) until it returns an empty list or 15 digests have been written in this pass. Each call returns sessions that need one: at least 3 episodes or 15,000 characters, idle for an hour, and either undigested or grown to double the episodes their digest covered (rewrite_of is set then). Privacy-excluded projects (BRAIN_DIGEST_EXCLUDE on the server) are already dropped. For each, read it with session_episodes(session_id) and call summary_write(session_id, body, description, project, cwd, episodes): project is the basename of cwd, episodes the count you digested, description one line of 150 characters or less, body about 1,800 characters at most: what happened, decisions, facts learned, files touched, outcome. A rewrite keeps the existing path on its own. Never secrets or credentials, no transcript quotes longer than a sentence, and no client data beyond what a future session needs to act. Later nightly passes drain any remainder.