Sessions, snapshots, and evidence

An Interceptor host process runs a target session and owns the local API used by Operator. Keep that process alive through campaign finalization and collection.

Start and verify

With a sealed and built blueprint and host credentials configured:

interceptor run --environment ./interceptor-environment \
  --model-provider openai-chat --provider-base-url https://api.openai.com/v1 \
  --provider-auth api-key-env --model-api-key-env TARGET_MODEL_KEY \
  --allow-model YOUR_TARGET_MODEL

Add separate embedding options from setup when the application needs them. Record the session ID and printed application URL. Non-HTTP interfaces expose POST /operations/<operation-id> at that application URL; use the actual printed port, not the separate Operator API address.

curl --fail-with-body -H 'Content-Type: application/json' \
  --data '{"message":"What is the status of order-456?"}' \
  http://PRINTED_APPLICATION_HOST_AND_PORT/operations/invoke

This request matches the support recipe's input contract; other agents have their own operation/input shapes. Check application content and dependency observations, not only HTTP success. Do not attach a probe campaign to the instance reserved for Operator; restart after standalone verification to begin from fresh seed state.

Inspect and export

From another terminal with the same data-directory setting:

interceptor list
interceptor status SESSION_ID --json
interceptor logs verify SESSION_ID
interceptor logs export SESSION_ID --output ./interceptor-evidence.tar

Use a new export path. Interceptor retains protected events, state journals, blobs, checkpoints, operation receipts, and application stdout/stderr. Export each relevant session in a restore lineage. Verification establishes integrity of available data, not completeness or a successful security outcome.

For operational simulations, the same distinction applies to functional effectiveness. Intact evidence can record an incorrect answer or an unfinished task. Compare the recorded output and state with your task criteria, and retain the starting conditions needed to interpret repeated or varied requests.

Collect native application logs

The support recipe writes /state/support/logs/agent.log in a mutable root. Collect that file separately before purging. As the authorized local Docker administrator, find the exact session container ID from Interceptor's host binding/status and confirm it against the session and Docker daemon; do not select an unrelated container by a guessed name. While that container remains available:

mkdir -p collected-native-logs
docker cp EXACT_SESSION_CONTAINER_ID:/state/support/logs/agent.log ./collected-native-logs/agent.log

Read the copied file and check that it contains the completed invocation's log records. This is explicit administrator collection, not automatic campaign evidence. Application stdout/stderr also remain under the session's native logs directory.

Live mutable roots can reside in session-owned Docker volumes; host staging folders are not necessarily their current contents. If the container has been removed, preserve the session and its volumes and follow the release's native state collection procedure rather than guessing a volume or changing ownership. A verified snapshot can retain declared file state when snapshots are supported. Do not assume the ordinary logs export includes every arbitrary mutable file.

Snapshots and restoration

For an unbound administrative session with snapshot support:

interceptor snapshot create SESSION_ID --name baseline --json
interceptor snapshot list SESSION_ID --json
interceptor snapshot inspect SESSION_ID CHECKPOINT_ID

Once bound to a campaign, use snapshot operations through Operator or the native API. Restore is a lifecycle API operation, not an interceptor restore CLI command. It returns a fresh native session ID and binding; use those for subsequent work.

Snapshots include declared logical state, managed services, and application context explicitly serialized by configured hooks. They do not save process memory, open connections, remote state, or Attack Harness's conversation. Check actual restored behavior before advertising support. The basic support recipe leaves snapshots disabled because its conversation state lives in memory.

Stop without deleting evidence

interceptor stop SESSION_ID

Stopping retains evidence and leaves the API available until the owning run process exits. Ctrl+C ends that host service and any remaining target. Confirm Operator has finished first. Purge is a separate destructive action; export and collect required files before using a narrowly selected purge command.

If evidence storage fails, preserve the diagnostic and original retained records. Further turns or snapshots may be blocked. A response-recording failure can follow a committed side effect; absence of a record does not prove rollback. Fix the storage problem and prepare a fresh session rather than replaying uncertain work.