novafabric.ai/demo

Replay & structural diff

interactive

Pick a replay mode and read what it guarantees. Below, a structural diff of two runs of the same agent with a different prompt version and a different output.

Runs in your browser on a fixture from the NovaFabric repository. It reads files; it does not talk to a running server. The real interface is the CLI.

Re-run the agent's Python process with the recorded OpenAI / Anthropic chat replies served from the capsule.

No live model call. Tool calls still run live — this is not a sandbox.

Mocked replay — RUN_A

replay_id: 01KT2P4Y8WGTHE9MRP4ZRX7QP3
  • ✓openai.chat.completions · served from the capsule (1240ms recorded → 0ms replay)
  • ✓openai.chat.completions · served from the capsule (980ms recorded → 0ms replay)
  • ↯mcp.read-file · ran live
  • ↯mcp.run-tests · ran live
2/2model calls mocked
0/2tool calls mocked — tools run live
0env warnings

Only synchronous, non-streaming OpenAI and Anthropic chat calls are served from the capsule today, including the tool-call requests the model made. Tool results are not substituted: the tools themselves run live.

Structural diff

01KR5SQZPD… vs01KS9K8R2N…

2changed
0added
0removed
Environment0 changes

No environment changes.

Model calls1 change

2 aligned / 1 changed / 0 added / 0 removed

  • model_call[0]call_A_001 → call_B_001

    no changes

  • model_call[1]call_A_002 → call_B_002
    • request.messages[0].contentmajor
      −You are a thorough code reviewer. For the diff below, identify: (1) bugs, (2) edge cases the author may have missed, (3) style inconsistencies. Be specific. Quote line numbers.
      +Review this code diff. Note bugs and style issues.
Tool calls0 changes

No tool calls recorded.

Outputs1 change
  • outputs/review.md
    before: a6911021ffa38800…after: 833f562a971d0ace…

On your machine

nova — bash
$ nova replay <run-id> --mode forensic
$ nova replay <run-id> --mode mocked
$ nova diff <run-a> <run-b>
$ nova diff <run-a> <run-b> --output-format github-annotation