---
spec_doc_type: spec
spec_id: input/openai-compatible-surface
status: 구현됨
source_evidence:
  - type: contract
    path: agent-contract/outer/openai-compatible-api.md
    notes: Public compatibility behavior
  - type: contract
    path: agent-contract/outer/anthropic-compatible-api.md
    notes: Anthropic Messages behavior
  - type: code
    path: apps/edge/internal/openai/routes.go
    notes: HTTP routes and authentication
  - type: code
    path: apps/edge/internal/openai/chat_handler.go
    notes: Chat validation, routing, and tools
  - type: code
    path: apps/edge/internal/openai/responses_handler.go
    notes: Responses validation and normalized execution
  - type: test
    path: apps/edge/internal/openai/chat_handler_test.go
    notes: Standard inference and routing regressions
---

# OpenAI-Compatible Input Surface

## Current implementation

Edge exposes model discovery, Chat Completions, Responses, legacy Completions, Anthropic Messages, and token-counting compatibility routes. Public model ids resolve to provider-pool candidates or direct provider routes.

Standard streaming, sampling, reasoning, tool definitions, tool choice, structured output, provider passthrough extensions, authentication, and compatible error envelopes are preserved. Provider-native tool calls remain structured; text fallback occurs only after an explicit native-tool rejection.

Caller metadata is a bounded string map. It is never treated as local process control or a filesystem execution context. Correlation values do not resume provider state.

## Execution paths

- OpenAI-compatible providers use the raw provider tunnel when selected by the pool.
- Normalized adapters use `RunRequest` and ordered run events.
- Disconnect and timeout cancellation use the active run id.
- Usage attribution follows the selected provider or configured model group.

## Verification

- `go test -count=1 ./apps/edge/internal/openai ./apps/node/internal/adapters/...`
- `make test-e2e`