API reference
The complete, always-current surface is machine-readable:
/openapi.json— every endpoint, request and response shapes/llms.txt— one-line orientation for agents/agents.md— operating instructions and the default loop/agent-setup/prompt.md— required harness install, then the onboarding runbook in execution order/onboard.md— backwards-compatible alias of the agent-setup prompt/skill.md— the loop procedure as an installable skill
The endpoints you'll actually touch
| Endpoint | Auth | What |
|---|---|---|
POST /v1/provision | none / own key | Idempotent project + keys (agents prefer this) |
POST /v1/signup | none | Deprecated non-idempotent alias of provision (same terms) |
POST /v1/recover | recovery code | Rotate lost keys |
POST /v1/traces · POST /v1/logs · POST /v1/metrics | producer key | OTLP ingest, all three pillars (JSON or protobuf) |
POST /v1/feedback · /v1/feedback/audio | feedback key | Typed / spoken user feedback |
GET /v1/ingest/health | any project key | Setup status: last accepted / last rejected + reason |
GET /v1/telemetry | agent key | Bounded record window; resolve ids from packets |
POST /v1/runs/diagnose | agent key | The improvement packet over HTTP |
POST /v1/observations/open · /close | agent key | Claim the window, then admit cited Issues with recommendations |
POST /v1/issues/decide | agent key | Record an attributed Issue disposition; actor is self-reported on this key-only lane |
GET /v1/observations | agent key | The observation ledger (open and decided) |
POST /v1/runs/replay · GET ?replay_id= | agent key | Prove a prompt/model fix against captured failures pre-ship |
POST /v1/runs/eval-spec | agent key | Executable eval spec your coding agent runs against real code |
POST /v1/evals | producer or agent key | Record eval results — failed evals cluster like production failures |
GET /v1/loops · GET /v1/loops/<run_id> | agent key | Loop history: runs, stages, verification |
POST /v1/loops/events | agent key | Record grounded progress; verification outcomes are server-owned |
POST /v1/runs/verify | agent key | Window-over-window verification |
GET/POST /v1/scoring | agent key | Session-scorecard weights (reason required to change) |
GET/POST /v1/redaction | agent key | Data-scrubbing config (reason required) |
POST /v1/projects/<id>/archive | agent key | Stop ingest, downgrade keys — confirm with the project id |
Schemas: /schema/handoff-packet.json · /schema/improvement-packet.json
Everything here is also reachable as MCP tools from a connected agent — same operations, same bounds, plus the loop's state machine.