Everything on this page uses artifacts from the repository — a real capture fixture and the published run-capsule.schema.json. Where something is experimental, it says so. Nothing here is a mock-up of a feature that does not exist.

Prefer your own terminal? pip install novafabric then nova capture python your_agent.py. The whole tour below takes about five minutes for real.

01/

Capture any command

works today

NovaFabric wraps a process. Not a framework integration, not a decorator you add to every call site — a wrapper around whatever you already run. A shell script, a notebook cell, an agent, a SLURM training job.

nova — bash

$ nova capture python review.py --diff examples/sample.diff

✓ Capsule written: ~/.novafabric/capsules/01KR5SQZPDGTKE3MDP3ZRX8WP1
  trace.jsonl          ✓  spans recorded
  model-calls.jsonl    ✓  2 LLM calls
  tool-calls.jsonl     ✓  2 tool invocations
  env.lock             ✓  environment snapshot
  redaction-proof.json ✓  no secrets found

The run exits with its own exit code. If NovaFabric fails, your workload still runs — never blocking the user's work is an architectural invariant, not a setting.

02/

Inspect the capsule — and try to break it

works today

Below is the manifest from that capture, validated against the real JSON Schema 2020-12 file from the repository. Ajv compiles the schema in your browser and reports what it finds.

A green check a website could have hard-coded proves nothing, so: break it. Delete a required field, change a type, put a nonsense value in an enum. If the check stayed green, this demo would be worthless — and for a project selling verifiable evidence, an unfalsifiable demo would be the wrong thing to ship.

TRY TO BREAK IT

invalid — undefined errors

{
  "schema_version": "0.1.0",
  "novafabric_version": "0.6.12",
  "run_id": "01KR5SQZPDGTKE3MDP3ZRX8WP1",
  "created_at": "2026-04-15T10:00:00+00:00",
  "finished_at": "2026-04-15T10:00:04+00:00",
  "duration_ms": 4127,
  "command": [
    "python",
    "review.py",
    "--diff",
    "examples/sample.diff"
  ],
  "exit_code": 0,
  "status": "success",
  "capture_mode": "cli-wrapper",
  "working_directory": "~/projects/code-review-agent",
  "host": {
    "arch": "x86_64",
    "os": "linux",
    "python": "3.12.4",
    "cpu_count": 16,
    "memory_bytes": 67108864000,
    "gpu": [],
    "hostname_redacted": true
  },
  "trace_ref": "trace.jsonl",
  "trace_root_span_id": "0123456789abcdef",
  "model_calls_ref": "model-calls.jsonl",
  "tool_calls_ref": "tool-calls.jsonl",
  "assets_ref": "assets.jsonl",
  "environment_ref": "env.lock",
  "replay_policy_ref": "replay.yaml",
  "redaction_proof_ref": "redaction-proof.json",
  "lineage_ref": "lineage.jsonl",
  "model_call_count": 2,
  "tool_call_count": 2,
  "mutating_tool_count": 0,
  "inputs": [
    {
      "name": "diff",
      "path": "inputs/sample.diff",
      "content_hash": "sha256:2a8439e85cae440ff7e28be7009b7056ccff0597d706a093a7e0c44e90f022b7",
      "size_bytes": 1240,
      "media_type": "text/x-diff"
    }
  ],
  "outputs": [
    {
      "name": "review",
      "path": "outputs/review.md",
      "content_hash": "sha256:5d93970a885c44d5babc3cf3dabcbcbd275e1c3df1fdd020f76126bc14bf5819",
      "size_bytes": 612,
      "media_type": "text/markdown"
    }
  ],
  "metadata": {
    "agent": "code-review-agent@0.1.0",
    "prompt": "code-review-prompt@0.1.0"
  }
}

The same check runs locally as nova validate <run-id>.

03/

Replay it

works today

A replay re-executes or inspects the capsule with every external call under NovaFabric's control — and produces a new capsule, so you can diff a replay against the original.

Mode	Network	Best for
`forensic`	none	Audit and post-incident inspection
`mocked`	served from cache	CI and regression testing
`semantic`	re-executes	Remote LLMs that drift — judges meaning
`exact`	controlled	Local / on-prem byte-exact re-run

The limitation, stated up front. NovaFabric does not claim byte-exact replay of remote LLM calls. That needs a deterministic environment and a per-call seed — realistic for a local model, not for a hosted endpoint that can change under you. For remote models that drift, semantic mode scores similarity of meaning on a 0.0–1.0 scale. If you have seen “deterministic replay” advertised for hosted models, this is what the honest version of that claim looks like.

04/

Diff two runs

works today

Something changed between Tuesday and Wednesday and the git diff is empty. Structural diff compares two capsules — model version, environment, tool responses, outputs — rather than diffing log text.

nova — bash

$ nova diff RUN_A RUN_B

Diff: 01KR5SQZPDGTKE3MDP3ZRX8WP1 → 01KR5T2A9WKPN6M3T8VZCJDR4Y
  changed=2  added=0  removed=0

Outputs:
  ~ outputs/review.md
  ~ model-calls.jsonl

05/

Trace lineage

works today

A directed provenance graph over runs, assets, and artifacts. It answers the three questions people actually ask:

Provenance — what produced this artifact?
Blast radius — if this dataset was wrong, what else is suspect?
Replay chain — what must be re-run to regenerate this?

SQLite by default. Kuzu, Postgres, AGE, and JanusGraph backends exist for cluster scale and are experimental.

06/

Export evidence anyone can verify

works today

The payoff. A signed bundle an auditor verifies offline, with no NovaFabric installed — only sha256sum and an ed25519 verifier. Evidence only its own tool can check is not evidence; it is a database row.

nova — bash

$ nova export-evidence 01KR5SQZPDGTKE3MDP3ZRX8WP1
✓ evidence-01KR5SQZPD.zip  (in-toto DSSE + RFC 3161 timestamp)

$ nova seal verify 01KR5SQZPDGTKE3MDP3ZRX8WP1
✓ signature valid    ✓ timestamp valid    ✓ redaction proof intact

07/───THE REST OF THE SURFACE

Beyond the core loop. Everything below is implemented; the experimental label means the interface may still move, not that it is vapour.

Asset registry

works today

Versioned models, prompts, datasets and agent configs as name@version, pinned to a git SHA, with eval-gated promotion.

HPC / SLURM capture

works today

No daemon, no root, no privileged access. Compute-node hot paths write to a local spool, never a database.

Compliance exports

experimental

EU AI Act Annex IV, NIST AI RMF, GDPR Art. 30, NIS2, HIPAA, CycloneDX AI-BOM. Evidence workflows — not legal certification.

Cost & energy attribution

experimental

Per-run LLM cost from a local pricing catalog, and energy-anchored action receipts.

Local dashboard

experimental

nova serve --experimental. Read-only, binds to 127.0.0.1 only.

Server mode

experimental

Multi-user REST API, Postgres, object store, tenancy with row-level security.

Cluster-scale collector

experimental

Node spools, parent/child capsules, object capsule store, four lineage backends.

Knowledge graph & diagnosis

experimental

Capsule knowledge graph, causal-graph attribution, counterfactual root-cause search.

Run the whole thing yourself

nova — bash

$ pip install novafabric
$ nova capture python my_agent.py
$ nova validate <run-id>
$ nova replay <run-id> --mode forensic
$ nova diff <run-a> <run-b>
$ nova export-evidence <run-id>

If any of that does not work for you, that is a bug and we want to hear about it — not a sign you did it wrong.

full walkthrough →documentation →source →

See it work

Capture any command

Inspect the capsule — and try to break it

Replay it

Diff two runs

Trace lineage

Export evidence anyone can verify

Asset registry

HPC / SLURM capture

Compliance exports

Cost & energy attribution

Local dashboard

Server mode

Cluster-scale collector

Knowledge graph & diagnosis

Run the whole thing yourself