AI Operations Agent
The operations agent is inactive by default. Its provider is empty, its permission ceiling isread, and system tools are off.
Read execute.Safe, Mutate and Full tools return a permission error before their tool body runs.Log masking
query_logs resolves an app, deployment id or owned container id to a persisted Voxeltron deployment. It reads that deployment's runtime logs unless[monitoring].loki_url is set, in which case it queries Loki by the deployment's container id.
Both providers use that resolver, so they only read containers recorded by a Voxeltron deployment.
Both paths mask that deployment's environment values and protected file contents, then applyredact_pii. The pattern pass masks email addresses, Bearer tokens,xoxb- and xoxp- Slack tokens, sk- API keys of at least 20 characters, hex runs of at least 44 characters, and base64 runs of at least 44 characters.
It preserves IPv4 addresses and 40-character hex commit SHAs by design. It does not detect sensitive text that is neither one of the deployment's configured secret values nor matched by those patterns.
Registered read tools
An active agent registry contains these tools:
query_logs, which uses runtime logs withoutloki_logs_urland Loki with it.query_metrics, which uses Prometheus when its URL is configured and the metrics store otherwise.query_traces, which returnsno tracing backend configuredwithout Jaeger.list_apps,get_project_status,list_deployments,list_alerts, andassess_deployment_feasibility.db_query,kubernetesandcicd_inspection, each of which reports that it is not configured when its backend is absent.compose_listandcompose_validate.investigate_incident,analyze_logs,correlate_events, andsuggest_fix, always registered. They make no model calls of their own.analyze_logscounts levels, common error patterns and the most frequent messages in the logs it is given. The other three run root cause analysis over the service topology, and say which evidence sources were not configured, had no container, or timed out after 3 seconds.
These tools have registration conditions:
- Outside sovereign mode, the registry adds
web_search,http_client, andartifact_inspect. - With a compose deployer, it adds
compose_logs,compose_deploy,compose_update, andcompose_delete. - With
loki_logs_urlset, it addsstructured_logs.
structured_logs searches the daemon's local LogStore, not Loki. The daemon creates that store as in-memory SQLite, and no production code inserts log rows into it. The only insert_log call in server.rs is in a test fixture, so the production tool searches an empty store.
Other registered tools
The registry always adds microvm_status at Safe,microvm_rebuild and microvm_warm at Mutate, andnotify at Safe. Outside sovereign mode it also addsscheduled_job at Mutate.
Docker state changes are registered only when a Docker backend exists. They arerestart_container, stop_container, and start_container, all at Safe.
System tools stay off by default. Setting enable_system_tools = true adds the shell, file, git and related tool family to a non-sovereign daemon.
Permission ceiling
The ceiling comes from ai.max_permission. Accepted values areread, safe, mutate, and full.
| Tool ids | Permission |
|---|---|
restart_container, stop_container,start_container, microvm_status, notify | Safe |
deploy, rotate_certs, microvm_rebuild,microvm_warm, compose_deploy, compose_update,compose_delete, http_request, scheduled_job | Mutate |
rollback | Full |
rotate_certs, http_request, and scheduled_job are not registered in sovereign mode. http_request also requires a non-empty[ai.http_request].allowed_hosts list.
The registry compares each tool's permission with the ceiling and returnsPermissionDenied without executing the tool when the tool is above it.
Direct agent tool calls do not ask a human for approval. Runbooks can pause a step markedRequired and request approval through the RunbookService stream, but approval does not change max_permission. A Mutate orFull step still fails at read after an approval, so the operator must raise the ceiling before the tool can run.
Memory
The daemon builds memory withMemorySystem::new_sqlite_with_deterministic_embeddings and a dimension count of 1536.
That constructor always installs DeterministicEmbeddingGenerator, which hashes the text into a normalized vector.
OpenAiEmbeddingGenerator exists, but the daemon does not construct it and the configuration has no embedding provider setting.
Memory similarity is therefore hash based, not semantic. No text from memory retrieval is sent to an embedding provider.