Skip to content

console-api.md

Latest commit

 

History

History
93 lines (75 loc) · 4.9 KB

console-api.md

File metadata and controls

93 lines (75 loc) · 4.9 KB
Raw

Redrob Console API (verified 2026-08-21)

Office's hosted inference talks to one backend: the Redrob Console. This file records what that API actually accepts, read from its published docs and OpenAPI document and checked against the live API.

Human reference: https://console.redrob.ai/docs/api-reference. Machine spec: https://console.redrob.ai/api/backend/docs-json (Swagger UI at /api/backend/docs). The server reports itself as "Redrob Console API 1.0".

Shape

  • Base URL: https://console.redrob.ai/api/backend/v1
  • Auth: Authorization: Bearer <key>. The spec names two schemes: jwt for console sessions and api-key for keys issued from the console. Inference endpoints take api-key only.
  • Key format: rrk_{prefix}_{secret}, returned once at creation. The prefix is the public half, meant to be displayed as rrk_{prefix}_••••.
  • Errors from OpenAI-compatible endpoints use OpenAI's error envelope. An unauthenticated request may be rejected before that handler with the gateway's standard 401 body.

Endpoints Office uses

  • GET /v1/models — key verification and model list. Returns { object: "list", data: [{ id, object, created, owned_by }] }.
  • POST /v1/chat/completions — OpenAI-compatible completion. stream: true produces data:-prefixed SSE chunks terminated by data: [DONE].

Model ids on the OpenAI-compatible chat path use the wire id auto (Console route redrob/auto). Older docs listed redrob-ai / redrob-translate; live Console chat/completions rejects redrob-ai with model_not_found, so Office wires auto only. model in settings may be empty because the engine ignores it.

There are also Redrob-specific inference endpoints Office does not use: POST /v1/complete, POST /v1/complete/stream, POST /v1/translate. The rest of the API (auth, workspaces, members, keys, billing, usage, logs) is console management behind the jwt scheme, not something the desktop app calls.

What POST /v1/chat/completions accepts

ChatCompletionDto accepts OpenAI request fields and relays supported sampling fields to the serving vendor. Important fields for Office are:

field notes
model wire id auto (route redrob/auto)
messages required; role is system | developer | user | assistant | tool; content is a string, text-part array, or null on an assistant tool-call turn
stream boolean
temperature 0 to 2
max_tokens, max_completion_tokens equivalent answer-length bounds
tools, tool_choice, parallel_tool_calls OpenAI function-calling shape; Console relays tools and does not execute them
functions, function_call legacy function-calling aliases
sampling and compatibility fields top_p, n, stop, penalties, seed, response_format, logprobs, stream options, and accepted metadata fields
indicAssist Redrob extension: answer a non-English conversation in its own language
detectLanguage Redrob extension
sourceLanguage Redrob extension; enum of 15 language codes including ko

The response is { id, object, created, model, choices, usage, redrob }, where an assistant message may carry tool_calls, finish_reason may be tool_calls, and redrob carries { requestId, detectedLanguage, translationUsed, latencyMs }.

Function calling: verified live

The function-calling loop was tested against the live API on 2026-08-21:

  1. GET /v1/models returned redrob-ai and redrob-translate.
  2. Plain chat using max_completion_tokens returned the requested text.
  3. A request with text-part content, one function, tool_choice: "required", and parallel_tool_calls: false returned finish_reason: "tool_calls" and a valid get_weather({"city":"Seoul"}) call.
  4. Sending the assistant tool_calls message followed by a role: "tool" result returned the final natural-language answer.

The exact OpenClaw path was tested too. A Gateway configured as Office configures it (api: "openai-completions", provider redrob, model redrob-ai) sent its normal content-block request and tool catalog. Console returned an exec call, OpenClaw ran printf openclaw-console-ok, sent the tool result back, and Console returned openclaw-console-ok.

That run is kept as evidence about Console, not as a description of Office. The gateway it used is gone: Office runs every turn on the Redrob Code engine, which reads the same Console key from REDROB_API_KEY and speaks the same endpoints recorded above. What the run still proves is that Console relays a tool catalog, returns a tool call, and accepts a role: "tool" result, which is the part of the contract the engine depends on.

Verifying the wire is not the same as verifying the app. The engine contract test (office/src/main/redrob-code/engine-contract.test.ts) is what exercises Office's own tools and policy prompts against a real engine, and it is what to run when that integration changes.