Interface IModelAdapter

ReadonlyproviderId

ReadonlymodelId

Readonlycapabilities

complete

• complete(
      request: CompletionRequest,
  ): Promise<Result<CompletionResponse, ModelError>>

  Send a completion request.

  Parameters

  • request: CompletionRequest

    The completion request

  Returns Promise<Result<CompletionResponse, ModelError>>

  Result with response or ModelError

stream

• stream(request: CompletionRequest): AsyncIterable<StreamChunk>

  Stream a completion request.

  Parameters

  • request: CompletionRequest

    The completion request

  Returns AsyncIterable<StreamChunk>

  Yields

  StreamChunk objects as they arrive

countTokens

• countTokens(text: string): Promise<number>

  Count tokens in text.

  Parameters

  • text: string

    Text to count tokens for

  Returns Promise<number>

  Approximate token count

validateConfig

• validateConfig(): Result<void, ConfigError>

  Validate adapter configuration.

  Returns Result<void, ConfigError>

  Ok if valid, ConfigError if invalid

OptionallistModels

• listModels?(): Promise<readonly ModelMetadata[]>

  (Optional, #2529) List models served by this adapter's endpoint.

  Implemented by adapters facing OpenAI-compatible endpoints (the upstream OpenAI API, OpenRouter, vLLM, custom gateways, etc.) — usually wraps GET /v1/models. Result is the harness-side identity resolver's most-trusted signal for "what model is actually being served behind this adapter."

  Subprocess-CLI adapters (claude / codex / gemini / opencode) leave this undefined; identity for those falls back to modelId parse.

  Implementations should cache the result for ~5 minutes — operators shouldn't pay round-trip latency on every resolve. Failures (network error, endpoint unsupported, auth missing) should throw so the caller can fall back; do NOT silently return an empty list.

  Returns Promise<readonly ModelMetadata[]>