Replayability

Every run is replayable. The engine checkpoints the full message/tool-call history at every node boundary, and any step can be re-executed against a different model, prompt, or tool output.

This is what makes ALDO AI a debugger-grade orchestrator rather than a "fire and forget" SDK.

#What's checkpointed

At each step the engine writes:

• The system + user messages going into the call.
• The model output (assistant message + any tool calls).
• The tool results that came back.
• The routing decision (which model was chosen, which were skipped, why).
• The guard verdicts (which guards ran, what they decided).

The full bundle is signed, content-addressed, and durable. The control plane keeps it forever (subject to your retention policy); local-only deployments keep it on disk.

#What you can do with it

• Replay — rerun the entire run, deterministically, against the same models. Useful when reproducing a regression.
• Edit-and-resume — change the prompt or tool output at step N, re-execute from N+1. The UI offers a side-by-side diff of the original and edited bundle.
• Swap model mid-trace — pick a step, choose a different capability class or model, and watch the alternate timeline run to completion. The "what would have happened on a local model instead of the frontier?" question becomes trivial.
• Breakpoints — set a breakpoint on a tool call name; the next run pauses there, shows the proposed call, and waits for you to approve / edit / skip.

#API

The replay surface is exposed under /v1/runs/:id/:

• GET /v1/runs/:id/events — full event stream.
• GET /v1/runs/:id/breakpoints — list breakpoints.
• POST /v1/runs/:id/breakpoints — set a breakpoint.
• POST /v1/runs/:id/continue — release a paused run.
• POST /v1/runs/:id/edit-and-resume — fork from a step with edits.
• POST /v1/runs/:id/swap-model — fork with a different model.

See the API reference for full schemas.