Govern

Fixture replay

YAML cases against a live plane. Connector policy still runs. Spend and daily run caps skip. Not a sandbox.

What it is

runkite sim -f fixtures.yaml creates a fresh thread per case, starts a run with X-Runkite-Simulation, polls until terminal, and optionally matches Admin audit deny/pending rows. Admin lists those runs with a sim badge.

Why it is here

Predicate and HITL checks need a real MCP path. Replay is how CI proves the plane still pending a large transfer without filling the day cap or the spend ledger.

How to implement

Write YAML: agent, cases[].input, optional expect.policy_effects.
Use an admin API key when auth is on. A write-only key gets 403.
Run runkite sim -f … --url … --api-key …. Exit 0 means every case passed.
Expecting pending does not approve the HITL row. Approve stays in Admin.

runkite sim -f examples/sim/predicates.yaml \
  --url http://127.0.0.1:2026 \
  --api-key "$RUNKITE_API_KEY" \
  --timeout 2m

What skips

Skip: FinOps holds, usage ingest, daily max_runs_per_day, LLM cache.
Keep: concurrent admission, rate limits, kill, connector policy, HITL.
SQL required for policy asserts. Mongo fails those cases with a clear error.

What to expect

Allow cases call the real connector. Downstream is not stubbed.
In-graph tools that never hit connector MCP stay ungoverned.
The admin key is a real Admin credential (kill, grants, spend), not a sim-only permission.
Stdout is PASS/FAIL lines. No JSON output in this release.

Reference: docs/sim.md · Grants & HITL · Runs & threads