AI & Agents

The ai service covers model configuration, the agent kernel and the embedded worker: encrypted model keys, full run traces, usage and quotas.

Capabilities

Model management: provider / base_url / model / capabilities; API keys are stored with AES-256-GCM (ciphertext and nonce separated), every list/detail response masks them as ********.
Duplicate-name protection: model names are unique; create/rename check for conflicts and propagate lookup failures (only not found counts as free).
Agent kernel: definitions (agent_key ^[a-z][a-z0-9_-]{1,63}$) → versions (instructions / model / tools / max_steps / timeout) → runs (run / steps / events persisted).
Idempotent runs: Idempotency-Key reuses the original run for the same input and answers 409 for conflicting input.
Embedded worker: the worker runs inside the ai-service process — there is no standalone worker container; run requests are logged as agent-run-<id> with token/cost/latency usage.
Usage & quotas: ai_usage_logs / ai_quotas per model and tenant; platform admins see all tenants, regular users are pinned to their own.

Model and run APIs go through the gateway (ai:agent:* permissions) with strict tenant scoping. Plaintext keys never leave the service.

🧪 External model vendors are user-configured; the repository stores no vendor credentials.
Isolated verification uses a local protocol fixture — not external vendor qualification.