OpenCode의 일반 Chat 요청을 GPT provider가 거부한 뒤 재시도 가능한 오류로 왜곡해 벤치가 장시간 정체됐다. 선택된 protocol profile에 맞춰 출력 토큰 필드를 정규화하고 upstream 400을 비재시도 validation 오류로 유지한다. |
||
|---|---|---|
| .. | ||
| cmd/edge | ||
| internal | ||
| Dockerfile | ||
| README.md | ||
IOP Edge
IOP Edge owns Node registration, dispatch readiness, provider-pool routing, queue admission, leases, reconnect fencing, external inference surfaces, and Control Plane operations.
Provider boundary
- Public model ids resolve to provider adapter/target routes.
- Normalized inference uses ordered run events; OpenAI-compatible passthrough uses provider tunnel frames.
session_idis correlation only and never resumes local execution state.- Cancellation targets an active run id.
- Edge operations are
health.check,node.status, andprovider.command. - Provider commands are limited to capabilities, transport status, and the Ollama API tunnel.
- Caller metadata remains an ordinary bounded string map.
- Edge and Node use a shared tunnel-tolerant liveness profile: a 30-second heartbeat interval and a 45-second response wait. Long provider prefill or streamed-response backpressure does not prematurely fence a healthy connection; a real timeout still closes only the current generation and triggers Node reconnect.
Edge and Node preserve standard OpenAI/Anthropic-compatible inference, streaming, tools, usage, provider lifecycle, capacity, queues, and reconnect behavior. They do not own interactive host programs, persistent conversations, terminal emulation, arbitrary host commands, or local filesystem execution context.
Development
go test -count=1 ./apps/edge/...
go test -race -count=1 ./apps/edge/internal/service ./apps/edge/internal/openai
make test-e2e
Canonical contracts:
agent-contract/outer/openai-compatible-api.mdagent-contract/outer/anthropic-compatible-api.mdagent-contract/inner/edge-node-runtime-wire.mdagent-contract/inner/control-plane-edge-wire.md