Agent requests

GET /agent

Tell whether an agent is configured and which provider and model it uses.

Clients check this before offering the chat composer. Without a configured LLM provider available is false and provider and model are null.

Responses

GET /observe

List the caller’s own agent requests, newest first.

Only requests the caller made are listed; a user sees their own and a service account its own. source narrows the list to new-dashboard requests (observe), follow-up prompts (prompt) or comment revisions (comment). A new-dashboard request that failed has no dashboard and lives only in this list. Progress steps are not included; read GET /observe/{id} for them.

Parameters

Responses

POST /observe

Ask the agent to build a new dashboard, or to revise an existing one with a follow-up prompt.

The request is queued and answers 202 at once; poll GET /observe/{id} until the status is succeeded or failed. Without dashboard_id the agent builds a new dashboard (source is observe); the finished request carries the new dashboard_id and url. With dashboard_id the agent revises that dashboard (source is prompt), exactly like POST /dashboards/{id}/revisions; base_version defaults to the dashboard’s latest version. The prompt is trimmed first, and the limit of 1 to 4000 characters applies to the trimmed text. Control characters other than newline, carriage return and tab, and the bidirectional override and isolate characters (U+202A to U+202E and U+2066 to U+2069), are refused. A caller may have at most 3 new-dashboard requests queued or running (429 too_many_requests), the instance at most 20 (429 agent_busy), and a dashboard at most 5 revisions (429 too_many_revisions). Without a configured LLM provider the endpoint answers 503 agent_unavailable.

Request body

application/json, required.

Responses

GET /observe/{id}

Read one agent request with its progress steps and usage.

Works for any agent request, including comment and prompt revisions. A request that has no dashboard yet, or never got one, is readable only by the principal that made it; anyone else gets 404. Clients poll it once a second until the status is succeeded or failed. After a shutdown release the re-claimed run reuses its attempt number with a fresh step list, so a client detects a restarted list by each step’s at, not by attempt alone.

Parameters

Responses

Generated from api/openapi.yaml at bac0fe39. Report a problem.