lum1104/uni-code/cross-agent-recall

Cross-agent recall

This machine has uni-code installed: a local, read-only index of every Claude Code and Codex session, exposed through the uni-code MCP server.

When to reach for it

• The user references past work you don't have in context: "that bug we fixed", "the command from last time", "why did we choose X".
• The user mentions another agent: "I started this in Codex", "Claude tried something yesterday".
• You are about to re-derive something (a config, a fix, an investigation) the user has plausibly done before — one cheap search first can save the whole rework.

How to use it

1. resume_session when the user wants to CONTINUE earlier work in earnest ("resume the Codex session", "continue exactly where I left off") — it reconstructs the session's effective context: the source agent's own compaction summary plus everything after it. Pass project: "."; if the result turns out to be this very conversation, pass the other agent's name or an explicit id.
2. get_handoff for a quick one-page orientation instead (goal, where it ended, files touched, commands, errors) when the user just wants you up to speed, not a full continuation.
3. search_history with distinctive terms — code identifiers and error strings beat prose; Chinese/Japanese/Korean queries work. Add project: "." to scope to the current project.
4. read_session on the best hit's id. Default returns the END of the session (outcomes, conclusions); pass from_start: true to see the original goal.
5. list_sessions when there is no search term — e.g. "what was I doing yesterday?".

Notes

• Results may come from a different agent than the current one; treat them as the user's own history, and say which agent/session an answer came from.
• Everything is local and read-only; the index refreshes automatically on each query. If the very first query on a machine is slow, that is the one-time initial indexing.
• If a search returns nothing, retry once with fewer or different terms before concluding the history doesn't exist.