Prepare and run a campaign
Keep reusable inputs separate from each retained run. A campaign combines a target, your bundle, an Attack Harness image, model access, and limits.
Prepare the target profile
Complete setup and prepare the target's Interceptor blueprint. Copy
the selected Operator release's release/templates/interceptor-target-profile.json
to a private operator-environment/target-profile.json. Configure the target ID,
permitted operation IDs, caller identities, routes, feedback, timeout, and target-stop
policy from the actual target and your authorized objectives.
The template uses interceptor/v1 at 127.0.0.1:8080, allows invoke and
black-box target output, and allows no injection routes. Do not invent route IDs
or add permissions merely to make a bundle pass. Optional client endpoint fields
do not create a remote Interceptor listener or authentication system.
Export capabilities
From the campaign workspace, with a complete profile and sealed blueprint:
umask 077
mkdir -p operator-environment runs exports
interceptor capabilities --environment ./interceptor-environment > ./native-capabilities.json
operatorctl capabilities export --config /absolute/private/config.yaml \
--target-profile "$PWD/operator-environment/target-profile.json" \
--native ./native-capabilities.json \
--output ./operator-environment/capabilities.json
Use new output paths. Preserve capabilities.json and its exact hash-named companion
unchanged. Do not recreate the native digest by hashing the file yourself. Capabilities
describe available operations; the target profile grants permissions.
This directory now contains the profile and generated export pair. It is different from the Interceptor blueprint and executes no environment scripts.
Prepare without starting
Author a scenario bundle against that export:
operatorctl submit --config /absolute/private/config.yaml \
--bundle ./scenario-bundle.json --environment ./operator-environment \
--output ./runs/support-review
operatorctl validate --config /absolute/private/config.yaml --run ./runs/support-review
Add --artifacts ./bundle-artifacts only for declared supporting files. The output
parent must exist and the submission directory must be new. A successful receipt
says validated-offline. These commands do not contact Docker, the target, or model
providers and do not start a campaign.
Prepare the real Interceptor instance before execution. It binds to the first attached campaign during that host process's lifetime. Use a separate verification instance, then restart cleanly instead of attaching a throwaway probe campaign.
Execute and inspect
Keep Interceptor running in one terminal and start the submitted run in another:
operatorctl campaign start --config /absolute/private/config.yaml --run ./runs/support-review
Alternatively, for new unsubmitted inputs:
operatorctl run --config /absolute/private/config.yaml \
--bundle ./scenario-bundle.json --environment ./operator-environment \
--output ./runs/another-review
Foreground execution stays attached through finalization. Keep the terminal and target running while final evidence is collected. Live admission verifies contracts, image, inputs, policy, and current target binding.
operatorctl inspect --config /absolute/private/config.yaml --run ./runs/support-review
To stop deliberately, use the campaign ID from inspection or its receipt:
operatorctl campaign terminate --config /absolute/private/config.yaml --campaign CAMPAIGN_ID
Interrupted work does not automatically resume. Inspect uncertain operations and cleanup before preparing a replacement campaign; do not replay side effects blindly.
Limits and completion
Bundle limits cannot expand installed ceilings or target permissions. Align the target turn timeout, complete-operation allowance, and request waits; the remaining campaign deadline may still end work sooner.
Completed execution, cleanup, complete evidence, and a supported security conclusion are different outcomes. Continue with results and reports after execution closes. Retain data until deliberately choosing to purge it.
For operational simulations, functional effectiveness is another outcome to assess from evidence. Campaign completion does not establish that the agent completed the intended task correctly. Account for requested cases, actual task outcomes, and failures; report consistency or reliability only within the conditions exercised.