Session Recall

Session recall

Use agent-session-find as the first recall step when previous agent work may answer the current question. The goal is to recover the right full session with minimal tokens, not to dump transcripts into context.

Keep session contents local. Do not upload session contents or paste whole session logs into external tools.

Workflow

1. Confirm the command setup in command-setup.md.
2. Start with a recent, bounded fuzzy query: agent-session-find --index-since 14d --max-sources 80 "<query>".
3. Add repo or cwd context when the project is known: agent-session-find --cwd "<repo-name>" --since 30d "<query>".
4. Search one source when the likely harness is known: agent-session-find --source codex "<query>" or agent-session-find --source claude "<query>".
5. If the user mentions a handoff, worker, subagent, delegated implementation, reviewer pass, branch, commit, or PR opened by another agent, retry with --workers.
6. If nothing matches, widen gradually: increase --index-since, remove --cwd, try synonyms, then omit --max-sources for a fuller local refresh.
7. If running several follow-up searches against the same index, add --no-refresh after the first successful refresh.
8. Use query-strategy.md for search terms.
9. Inspect low-token result cards first, then use result-handling.md before opening logs.

Defaults

Use these defaults unless the task suggests otherwise:

agent-session-find --index-since 14d --max-sources 80 "<query>"
agent-session-find --cwd "<repo-name>" --since 30d "<query>"
agent-session-find --limit 5 "<query>"

For stale or long-running projects, prefer --index-since 90d over an unbounded first pass. Run agent-session-find status when you need to see index size before widening.

Context pointers

• Use command-setup.md for binary, wrapper, privacy, database, and parallel-refresh rules.
• Use query-strategy.md for query terms and widening strategy.
• Use result-handling.md for interpreting result cards and deciding when to open local JSONL.