Session state, resumption, and forking

Long work doesn't fit in one sitting, and how you carry state across the gap changes the answers you get.

Resuming

--resume <session-name> continues a specific prior conversation with its context intact. For investigation work spanning days, this is the default.

Resuming picks up where the session stopped. It cannot return to an earlier turn or remove later reasoning from the conversation. If you need to preserve an earlier state, create a fork before continuing in a new direction.

The staleness problem

A resumed session carries old tool results. If files changed while you were away, the agent is reasoning over a snapshot of a codebase that no longer exists, and it has no way to know that.

You photographed the bulletin board last week and you're still reading the photo. Everything in it is exactly what was posted then, and nothing in it can tell you a notice came down on Tuesday.

Choose based on how much of the saved context is still reliable:

• Prior context is mostly still valid — resume, and explicitly tell the agent which files changed so it re-reads those rather than everything.
• Prior tool results are stale — start fresh and provide a structured summary of what was learned. This avoids carrying outdated results into the new work.

Forking

fork_session creates independent branches from a shared baseline. Analyze a codebase once, then fork to explore two refactoring strategies without either branch contaminating the other, and without paying for the analysis twice.

The branches have independent conversation state, but normally still see the same filesystem. If an experiment may edit files, pair the fork with a checkpoint, worktree, or other isolated workspace. A conversation fork isn't a file rollback.

Resumption continues one line of work. Forking splits one into several. They aren't alternatives.

Forking is also not delegation. A fork gives you a second conversation to carry on yourself, and whatever it produces stays in that branch. A subagent spawned with the Agent tool runs its own task and returns a result to the coordinator. Fan-out that has to be gathered into one report is delegation.