Review each tool call before execution.
Send the proposed action, context, policy, trajectory, and model metadata to the runtime API, then enforce the returned intervention.
Runtime checks
A runtime check is one pre-tool-call review. LatentOps is not only a log destination; the call must happen before the tool runs.
curl -X POST https://api.latentops.space/v1/runtime/review \
-H "Content-Type: application/json" \
-H "X-API-Key: lo_your_tenant_key_here" \
-d '{
"prompt": "Apply infrastructure changes",
"tool_name": "run_shell",
"tool_args": {
"command": "terraform apply -auto-approve"
},
"repo_context": {
"branch": "main",
"production_environment": true,
"protected_paths": ["infra", ".github", ".env"]
},
"policy": "balanced",
"log_event": true,
"model_provider": "openai",
"model_name": "gpt-5-mini",
"agent_framework": "github-actions",
"environment": "production"
}'Request fields
Start with prompt, tool_name, tool_args, policy, and log_event. Add workspace context, trajectory, and model metadata for production.
Task, plan, user instruction, issue text, or model message that led to the proposed tool call.
Tool identifier and arguments. For shell commands, include tool_args.command.
Repository, deployment, and domain context: production environment, sensitive data, approval requirements, changed files, protected resources, and operating domain.
Recent agent steps, tool results, retries, failed tests, prior warnings, or signs of drift.
Policy profile such as balanced. Policies determine whether risk becomes warn, block, or escalate.
model_provider, model_name, agent_framework, integration, environment, agent_id, and agent_name.
Response shape
Store the event for traceability, but enforce the intervention object. The exact evidence varies by policy and action context.
{
"event": {
"decision": "block",
"risk_score": 0.93,
"tool_name": "run_shell",
"policy": "balanced",
"metadata": {
"model_metadata": {
"model_provider": "openai",
"model_name": "gpt-5-mini",
"agent_framework": "github-actions",
"environment": "production"
}
}
},
"intervention": {
"decision": "block",
"risk_score": 0.93,
"recommended_control": "Do not execute. Ask a human reviewer.",
"evidence": [
"Destructive shell command",
"Production environment",
"Protected path"
]
},
"incident": {}
}Enforcement
Decide fail behavior before production. Use fail closed for deployment, secrets, production config, migrations, and destructive shell actions.
Execute the proposed tool call.
Execute only with a visible warning and audit trail.
Do not execute the tool call.
Require human approval before the action can proceed.
Endpoint reference
Use runtime review for production integration. The other endpoints support dashboard data, incidents, audits, models, and policies.
Review a proposed tool action before the AI system executes it.
List action categories, sensitivity levels, and default controls.
Fetch recent runtime checks, incidents, summaries, and model counts.
List incidents created from blocked or escalated checks.
Approve an escalated action and close the pending approval.
Reject an escalated action and retain the audit evidence.
Resolve or update an incident after operator review.
Export markdown audit evidence for a workspace.
List or store encrypted OpenAI and Anthropic keys for managed live calls.
List or run provider model calls and record them as traces.
View policies, create drafts, test rules, publish, and approve.
Import YAML or JSON policy text as a workspace draft, with optional publish.
Export a saved policy version as YAML or JSON for source control.
Fetch built-in templates for engineering, support, finance, cloud, and external messaging workflows.
Create a workspace draft policy from a built-in template.
List available benchmark datasets that can be selected in the dashboard.
Run a bounded benchmark comparison using dataset, limit, and policy query parameters.
Run the same comparison with a JSON body for API clients and internal tooling.