Integrations

Use loopexec run as the Go loop governor. Supply an external check and an optional work command. A check exit of 0 means the oracle passed; loopexec exits 10 when all guards also pass. Halt reasons are computed from observed state.

mkdir -p build/integration-demo
printf 'red\n' > build/integration-demo/status.txt
set +e
./build/loopexec run --json --workdir build/integration-demo \
  --run-id integration --max-iterations 3 --timeout 10s --command-timeout 2s \
  --exec "printf 'green\\n' > status.txt" --check "grep -qx green status.txt"
code=$?
set -e
test "$code" -eq 10
./build/loopexec replay --json --workdir build/integration-demo --run-id integration
./build/loopexec report --workdir build/integration-demo --run-id integration

Machine output is one JSON object with tool, version, status, errors[] and, where applicable, run ID, iteration, halt reason, check exit, and receipt path. Workflow output adds failure cause and the best verified candidate path. Use the halt reason for precise decisions and the exit class for CI branching. The complete map is in SPEC.md.

replay re-runs only the read-only saved check and compares its verdict fingerprint. It never runs an agent. report reads state/receipts without executing anything. reexecute --confirm is a live rerun of legacy configs; workflow receipts require explicit initialization of a new adapter workspace.

For unattended agentic work, use the workflow contract. Each adapter executes one bounded work/check/optional-review sequence. The governor owns retries, subprocess bounds, exact accounting, candidate promotion/restoration, and numeric patience. The creative example is one integration, not the product boundary; its fixture needs no Python runner, network, model key, Blender, or Godot.

A nonzero --budget-usd now requires workflow metering. Strict caps require enforced per-call reservations; observed metering only stops subsequent work. Local/subscription usage may report unknown dollars while time, iteration, call, and token bounds remain available. Model/sampling flags record metadata; the adapter’s actual argv/configuration chooses and limits the invoked model.

The unreleased model assignment contract adds explicit argv-bound profiles and role mappings, plus a read-only models plan preview. Adapters must consume the selected settings and report observed invocation identity; updating LoopExec alone does not migrate an existing consumer.

Legacy --failures-cmd and --integrity-cmd collect newline identities or {"ids":[]}. Collector failure or stderr cannot become valid evidence. The integrity baseline must retain all members. Numeric mode accepts a separate candidate-bound score and does not classify stable visual issue IDs as oscillation.

CI runs formatting, vet, tests, the existing deterministic dogfood scenarios, and the creative proof script. Hidden receipt files are included in proof artifacts; private signing keys and diagnostic tails are excluded. See CLI reference for the full command surface.

For SMALL CLI v1.1.2 and collaborative v2 sessions, see the SMALL integration guide. The optional real-CLI proof keeps task acceptance separate from a command exit and preserves session attribution without introducing a second loop controller.