PENDING: feat(hermes-agent): add authenticated Hermes client, readiness checks, and health collector #672

Parent epic: #668
Depends on: #670, #671

Context

Hermes exposes documented /health, /health/detailed, /v1/models, and OpenAI-compatible API surfaces. AWCMS-Mini needs a bounded server-side adapter and cannot trust arbitrary tenant-provided URLs.

Objective

Implement the authenticated Hermes client, safe endpoint verification, health snapshot collection, and module readiness signals.

Scope

• Add a HermesProvider/client port and Bun-native HTTP adapter.
• Resolve API_SERVER_KEY through a server-side secret-reference resolver; raw keys never enter PostgreSQL, browser JavaScript, logs, or audit.
• Probe:
  • GET /health;
  • authenticated GET /health/detailed;
  • authenticated GET /v1/models where supported.
• Add awcms_mini_hermes_health_snapshots with bounded, allowlisted fields only.
• Add on-demand endpoint:
  • POST /api/v1/hermes/agents/{agentId}/health/check
• Add read endpoints for latest/history health.
• Add scheduled job bun run hermes:health:collect.
• Add module-management provider-health/readiness signal.
• Report configured model vs live-advertised model without silently changing configuration.
• Normalize unreachable, unauthorized, model-mismatch, degraded, timeout, invalid-response, and healthy states.

SSRF and transport requirements

• Parse with URL; reject credentials, fragments, unexpected paths, and unsupported schemes.
• Exact hostname/origin allowlist; no arbitrary redirect following.
• Protect against DNS rebinding and metadata/link-local/loopback targets unless explicitly declared as a trusted internal endpoint by deployment configuration.
• Production external endpoints require HTTPS.
• Internal HTTP is allowed only for an explicitly trusted internal network/origin configured by the operator.
• Enforce connection/read timeout, response-size limit, bounded retry with jitter, circuit breaker, and concurrency/backpressure.
• Do not return upstream bodies, stack traces, or authorization details.

Acceptance criteria

• Healthy, degraded, unreachable, unauthorized, timeout, and model-mismatch states are deterministic.
• Detailed health fields are allowlisted and size-bounded.
• Server key is resolved only server-side.
• SSRF tests cover localhost, metadata IPs, link-local/private policy, redirects, userinfo, DNS changes, and oversized responses.
• Collector isolates failure per agent and is safe to rerun.
• Health snapshots are tenant-scoped and retention-ready.
• Live model check recognizes configured mimo-v2.5-pro when advertised.
• Audit contains safe status/error codes only.
• OpenAPI/AsyncAPI, jobs descriptor, docs, and tests are updated.
• A changeset is added.
• bun run check passes.

Out of scope

• Container start/stop/restart.
• Hermes session-history enumeration.
• Reading Hermes logs or state files.