Time Travel

Rewind a run to an earlier point and resume from there. The boundary is chosen with continue_from:

"end" — keep the whole transcript (a normal follow-up)
"last_user" — keep through the last user message, drop the post-user tail (tools re-invoke)
K (int) — keep messages[:K], an exact message-index boundary

For a COMPLETED run, /continue auto-forks into a new sibling run so the "1 run = 1 model loop" contract holds — the source stays intact. Passing fork=True makes that explicit and also works on non-completed runs: it creates a sibling instead of resuming in place. So run-level fork is just time-travel that keeps both the original and the new line, side by side in one session.

Boundaries are snapped to a tool-call-safe index automatically: a cut that would orphan an assistant tool_call from its result is moved back to the start of that exchange, so the resumed transcript is always valid.

Example	What it shows
`01_continue_from.py`	`continue_from="end"`, `"last_user"`, and the numeric `K` form. COMPLETED runs auto-fork.
`02_fork_run.py`	`fork=True` + a boundary to create an explicit non-destructive sibling run — for evals / A-B exploration from a known state.

bash

.venvs/demo/bin/python cookbook/02_agents/20_time_travel/01_continue_from.py
.venvs/demo/bin/python cookbook/02_agents/20_time_travel/02_fork_run.py

Redo just the last answer (keep tool exchanges): ../19_regenerate/
Copy a whole session instead of one run: ../21_fork_session/
The persistence layer this resumes from: ../18_checkpointing/

Time Travel

Time Travel

Related