Add agn replay for recorded tool actions

[manojsinghnegiwd]

Summary

Add an agn replay command that can rerun a previously recorded AGN run by replaying the concrete tool actions it performed, such as write_file, read_file, shell commands, patches, etc. The replay should be based on what AGN actually did, not on the original chat/messages.

This allows the same exact changes/actions from a prior run to be applied again deterministically.

Task

Implement replay support with entry points such as:

agn replay --environment <environment-id>
agn replay --conversation <conversation-id>

The implementation should support recording enough structured execution data during an AGN run to later replay the meaningful side-effecting actions. At minimum, the replay data should capture the ordered tool/action calls, relevant inputs, outputs or status where needed, and enough metadata to identify and validate the target workspace/environment.

Use Cases

1. Repeat the same run: Re-run the same sequence of actions again and again without re-prompting the model or relying on message history.
2. Apply environment-tested changes: Run AGN in a disposable environment, inspect/approve the resulting behavior, then replay the recorded actions against another target such as the real working directory.

Acceptance Criteria

• AGN records the concrete actions/tool calls performed during a run in a replayable format.
• agn replay --conversation <conversation-id> replays actions from a recorded conversation/run.
• agn replay --environment <environment-id> replays actions associated with a disposable environment run.
• Replay applies the recorded actions in the original order.
• Replay does not depend on the original prompt/messages to regenerate behavior.
• Replay has clear behavior for reads, writes, patches, shell commands, failures, and mismatched preconditions.
• Replay should provide useful output showing which actions were replayed and whether each succeeded or failed.
• Documentation explains the two main workflows: deterministic repeat runs and applying changes that were first tested in an isolated environment.

Open Questions

• Should replay be fully automatic, interactive/confirming each side-effecting step, or support both modes?
• Should read_file and other non-mutating actions be replayed, skipped, or validated against expected prior outputs?
• How should shell commands be classified and guarded, especially destructive or environment-specific commands?
• Should replay validate file preconditions/hashes before applying writes or patches?
• What should happen if the target workspace differs from the original environment?
• Should the CLI option be named --environment rather than --environement?

References

• Support AGN disposable Docker environments #7 — Describes AGN disposable Docker environments. Replay should integrate with this workflow by allowing a run performed safely in an isolated environment to be replayed/applied later outside that environment after inspection.