novafabric.ai/novafabric
experimental

Use NovaFabric for local-first capture, replay, diff, validation, and audit evidence workflows. Interfaces and schemas may change before v1.0. Production-scale and compliance-sensitive deployments require independent validation.Not claimed: million-agent readiness · HSM/X.509 enterprise PKI · compliance guarantee · stable v1.0 schema

Your agent ran for 3 hours and failed at step 47. You have no trace, no replay, no way to know if it was the model call or the tool call or the prompt. You rerun it. It fails differently.

This is a novafabric problem.

Capture every run.
Replay it anywhere.

Wrap any command. novafabric captures the environment, every model call, every tool exchange — without touching your code. One command. A portable, secret-redacted, signed capsule.

$pip install novafabricpip
$uv add novafabricuv

requires Python 3.12+

nova — bash
01/
what gets captured

One command.
Everything recorded.

Every captured run produces a structured directory — a capsule — that contains everything needed to understand, validate, and replay the execution. No instrumentation required.

capsule structure
.novafabric/runs/01HXAY7M5JZ8R7K4P9DPBYK2WX/
capsule.yaml← run manifest — id, status, timing
trace.jsonl← execution spans
model-calls.jsonl← every LLM call
tool-calls.jsonl← every tool invocation
env.lock← full environment snapshot
redaction-proof.json← proof no secrets leaked
replay.yaml← replay policy
inputs/
outputs/
stdout.txt
stderr.txt
⟶

Model-call evidence

Model name, prompt, completion, token counts, latency — for API calls where capture is enabled.

▦

Tool invocations

MCP exchanges, function calls, shell commands — the tool-call chain where capture hooks are active.

◎

Environment snapshot

Python version, installed packages, env vars (secrets redacted) — so you can reproduce the execution context.

⊚

Execution spans

OpenTelemetry-compatible spans for the execution tree. Queryable, exportable, visualisable.

◈

Redaction proof

Cryptographic proof that no secrets appear in the capsule. Auditable by anyone, no access to secrets required.

⬡

Tamper-evident seal

DSSE signature + RFC 3161 timestamp. Designed for audit evidence workflows; regulatory fit requires independent review.

02/
interactive demo

See what a capsule contains.

Click "run" to simulate a capture. Explore the files a capsule produces.

nova — bash
$ nova capture python agent.py
03/
replay modes

Four ways to replay.

forensic
$ nova replay --mode forensic

Read-only. Replays the exact execution environment without re-running model calls. Safe for debugging, auditing, and forensic investigation.

mocked
$ nova replay --mode mocked

Uses recorded LLM responses. No API calls, no cost. Deterministic replay for CI pipelines and regression testing.

semantic
$ nova replay --mode semantic

Re-runs with live model calls but validates that outputs are semantically equivalent. Detects prompt drift and model behavior changes.

exact
$ nova replay --mode exact

Full re-execution. Every tool call, every model call, live. Compares outputs byte-for-byte. Strictest reproduction guarantee.

04/
four commands

The complete lifecycle.

$nova capture python my_agent.py

Wrap any command — script, agent, training run. No code changes required.

$nova replay runs/01HXAY7M5 --mode forensic

Re-execute any capsule. Forensic mode is read-only. Mocked mode uses recorded LLM responses.

$nova validate runs/01HXAY7M5

Validate a capsule against schema. Checks integrity, redaction proof, and signature.

$nova seal verify runs/01HXAY7M5

Verify DSSE signature and RFC 3161 timestamp. Works offline, forever.

full documentation on GitHub →
novafabric

The flagship system: capture, replay, validate, and govern AI agent executions.

← back to lab