catchup

catchup

Reconstruct a project's working state from the brain and hand the main session one distilled brief to act on — without spending the main window on the retrieval.

This is the DEEP path. The routine path is automatic: the brain-session-start SessionStart hook injects memories/<project>-state.md into context at session start (startup / --resume / post-/clear), stamped with freshness. Reach for /catchup only when that brief is missing (no state note yet), carries the STALE banner, or visibly disagrees with the repo — its subagent re-derives state from the ground up (sync, ingest, recall, git cross-check), which the injection deliberately does not do.

All the expensive legwork (sync, ingest, state-note read, recall, freshness cross-check, pointer verification) happens inside ONE subagent; the main window only holds the brief. This is the /lean discipline applied to resuming: hold the brief, never the raw recall.

1. Resolve the project

• Use the /catchup <project> argument if given.
• Else infer from the conversation topic or cwd (basename of the working dir, e.g. ~/src/brain -> brain).
• If genuinely ambiguous (no arg, no clear cwd/topic), ask once, then proceed. Never silently guess between two live candidates.
• Headless (claude -p, no user to ask): infer from cwd, proceed, and record the inference in the brief's Staleness warnings so it is auditable.

2. Spawn ONE subagent (model: opus) with this exact brief

Model opus — this is retrieval legwork (house model economy: Opus for scoping/searching, Fable only for verification). Give the subagent this brief verbatim, substituting <project>:

You are reconstructing the working state of the <project> project from the brain (~/src/brain). Use Bash only — no MCP tools (the CLI always runs fresh code; a live MCP server may hold a stale schema). Run, in order:

1. uv run --directory ~/src/brain brain-sync pull — best-effort; on failure (offline / no remote) note "brain not synced" and continue.
2. uv run --directory ~/src/brain brain-ingest — incremental, fast; makes episodes current through the last session.
3. Read ~/src/brain/memories/<project>-state.md if it exists — the rolling state note reflect maintains. Missing → say so and fall back to digests + episodes.
4. uv run --directory ~/src/brain brain-recall "<project> state decisions next steps", then again with --scope summaries. Fetch full text of the top 1–3 hits with brain-recall --get <id> — snippets are too thin to brief from. If brain-recall is not installed, fall back to the MCP brain tools (recall / get_episode) and note the fallback.
5. Cross-check freshness: git -C ~/src/brain log -5 --oneline AND the target repo's own git log/git status. The brain is only current through the last reflect; if the target repo moved past it, trust the repo and flag the drift.
6. Verify every file path and command you are about to cite still exists.

Return ONLY a distilled brief, max ~400 words, in exactly this structure:

• Task — what this project is and what we were doing.
• State — done / in-progress / blocked, each with evidence (commit hashes, verdicts, test results).
• Decisions already made — calls not to re-litigate.
• Single next action — the one concrete first step.
• Verified pointers — exact file paths / commands, each confirmed to exist.
• Staleness warnings — only if the brain lagged reality, sync failed, the state note was missing, or a pointer failed to verify. Omit if clean.

Do not return raw recall output, transcripts, or file dumps — only the brief.