Catchup

Catchup

Rebuild context for the current session, then deliver a crisp summary so the user can resume work immediately.

Getting the history

Use the first source that works, in this order:

1. In-context history — only if verifiably complete. Use this only when you can see the session's first user message verbatim and no part of the conversation has been compacted or summarized. A compaction/summary block is not the earlier turns — it is lossy and silent (you cannot tell what detail was dropped), and the long sessions that prompt a catchup are exactly the ones most likely to have been compacted. If any earlier history has been summarized, treat this source as unavailable and fall through to the transcript — do not summarize from a compaction block.
2. A session-query tool. If your platform exposes one (e.g. a session store / SQL tool), query it for recent turns, milestones, and any referenced PRs/issues/commits.
3. The on-disk transcript. Otherwise locate this session's transcript file and read it. Agent CLIs store sessions under a per-user config dir, typically keyed by working directory and a session id. Discover it:
   • Check env vars for a session id and config/home path (names vary: *SESSION_ID*, *_HOME, XDG_CONFIG_HOME).
   • Look under ~/.<tool>/ or ~/.config/<tool>/ for a sessions, projects, or history directory; the current session is usually the most-recently-modified file there.
   • Transcripts are commonly JSONL (one event per line) or SQLite. Extract user/assistant turns plus any file-edit or tool-call records. Do not dump the whole file into context — pull just the recent turns and the list of files touched.

If none of these yield history, say so plainly and summarize only what's in context.

Known platforms (shortcut for step 3)

• Claude Code: ~/.claude/projects/<cwd>/$CLAUDE_CODE_SESSION_ID.jsonl, where <cwd> is the working directory with / and . replaced by - (e.g. /Users/me/dev → -Users-me-dev). JSONL; conversational turns have type: "user" / type: "assistant", and file edits appear as tool_use blocks with name in Edit/Write. The file also contains non-conversational events (attachment, file-history-snapshot, mode, system, etc.) — ignore them. A record's message.content may be a plain string or an array of typed blocks (text, tool_use, tool_result); a user record is often a tool result or a large injected attachment, not a real prompt. Filter to genuine prompts and text blocks before summarizing.
• GitHub Copilot CLI: use the session_store_sql tool. Query checkpoints, recent turns, session refs, and session files.

What to extract

• Recent turns (last ~15–20) — the freshest state of play. Filter to real conversational turns first, then take the last N. In transcript formats, tool calls, tool results, and attachments each occupy their own line, so a raw "last N lines" slice is not the last N exchanges.
• Files / areas touched — from edit or tool-call records; summarize by module or directory, not individual files.
• Refs — PRs, issues, or commits mentioned in the turns. Skip if none.

Ignore the /catchup request itself — it is usually the most recent entry. Summarize the work that came before it, not the act of asking for the summary.

Output

Deliver exactly this format:

---

Where we left off

[1–2 sentence prose: what we were working on, the goal, and the current state]

Key decisions

• [decision or conclusion reached, with brief rationale if captured]

What's next

• [the next logical step or open item, if determinable]

Areas touched

[One sentence on modules/directories affected — skip if only one file or trivial]

---

Do not ask follow-up questions. Output the summary and stop.