Skip to main content

Day-in-the-life workflows

FleetForge’s “Hello Fleet” storyline covers both the individual contributor (IC) loop on a laptop and the ChangeOps gate that runs on every pull request. Use this guide when you need a concrete, reproducible path for fleetforge-ctl receipts, verification, replay, and CI evidence.

IC workflow (local dev)​

  1. Run the canonical demo once
    Bring up the demo runtime (just demo or the quickstart Compose stack), submit the Hello Fleet DAG, and capture the printed RUN_ID. Keep that ID handy and reuse it everywhere (CLI, /demo, receipts, replay).

    # export a signed receipt (choose any manifest path under build/ or tmp/)
    fleetforge-ctl receipt --run-id "$RUN_ID" --manifest "build/receipts/${RUN_ID}.c2pa.json"

    # verify an artifact + manifest locally
    fleetforge-ctl verify tmp/hello-fleet/artifacts/demo-output.bin \
      --manifest tmp/hello-fleet/artifacts/demo-output.bin.c2pa.json

    # prove determinism
    fleetforge-ctl replay "$RUN_ID"

  2. Tighten a guardrail and re-run
    Flip the budget preset or run the agent_team_openai variant so you intentionally trip a policy denial. This yields the policy_decisions artifact, surfaces the signed TrustDecision, and gives you a concrete “bad path” that CI can catch later.

  3. Understand the evidence you just generated

    • Receipts are C2PA Content Credentials bound to each artifact, so verifiers can audit “who did what/when” without your runtime.
    • fleetforge-ctl verify walks the manifest + capability-token chain locally; pair it with fleetforge-ctl replay to surface drift (tokens, spend, or tool I/O) immediately.
    • Telemetry follows the OpenTelemetry GenAI semantic conventions (with trust.* attributes) so downstream collectors label runs with attestation IDs, capability tokens, and policy verdicts.

  4. Optional – sign like prod
    Local dev defaults to the env-ed25519 signer. When you need parity with production, point the CLI at AWS/GCP/Azure KMS via the existing shims (FLEETFORGE_TRUST_SIGNING_KEY, FLEETFORGE_SCITT_*, etc.) without changing the workflow above.

CI gate (pull requests)​

Treat every pull request as two checks that run back-to-back:

  1. ChangeOps decision (policy/evidence gate) – Pipelines build a JSON bundle (diff summary, novelty, eval coverage, replay/canary telemetry, budget deltas) and request a ruling:

    fleetforge-ctl gates check --input change.json --json decision.json

    The gate returns allow, follow_up, or deny. Missing replay evidence or thin coverage yields structured reasons (for example “coverage below minimum” or “missing replay telemetry”) so CI can fail fast. Gate inputs should include:

    • Replay telemetry with attestation_match=true, tool_io_match=true, and ≤1% token drift.
    • Canary/promote telemetry listing the attestation IDs it observed so reviewers can trace what passed or failed.

  2. Receipts-as-gates (artifact/replay verifier) – A GitHub Action (or equivalent) fetches fleetforge-ctl receipt --run-id … plus replay stats, enforces drift thresholds (defaults: 1% tokens, 0.5% cost), and posts an attestation diff on the PR if anything moved. Pair it with the step above (details in Receipts-as-gates).

Operational knobs you’ll set once​

  • Apply core/storage/migrations/0014_changeops_gates.sql and rotate the CI service token that calls gates check.
  • Export gate decisions to your observability stack so reviewers can correlate incidents to specific gate_ids.

Transparency hooks​

  • Emit SCITT statements or store receipts in a transparency log by enabling the FLEETFORGE_TRANSPARENCY_* env vars; local runs can stub these while production points at HTTP-backed logs.

Putting it together​

On a laptop: run Hello Fleet → export receipt → verify → replay (no drift) and capture one intentional guardrail denial. On every PR: assemble change.json → gates check returns allow/follow_up/deny → the receipts/verifier action enforces replay + receipt quality → PR blocks when evidence is missing or drift exceeds thresholds.

Why this matters to Security/Risk​

The gate enforces controls mapped to OWASP’s LLM Top-10 (prompt-injection guardrails, tool ACLs, output handling) and surfaces receipts plus replay evidence as proof—not just logs—so Risk/Security reviewers inherit deterministic, auditable workflows.