开源AI工作流

OpenCode Go Proxy

Use your OpenCode Go and OpenCode Zen access in the Codex app.

Codex expects a Responses API (/v1/responses). OpenCode Go exposes an OpenAI-compatible Chat Completions API (/v1/chat/completions), and OpenCode Zen serves GPT, Claude, Gemini, Grok, DeepSeek, GLM, Kimi, and Qwen models over four different API surfaces. This proxy bridges both in one local process:

Codex app
    │
    │  POST /v1/responses (Responses API)
    ▼
opencode-go-proxy  ←── localhost:8787, one runtime dep (zstandard)
    │
    │  POST per family: chat/completions · responses · messages · models/<id>
    ▼
OpenCode Go / OpenCode Zen  ── 13 open Go models · GPT, Claude, Gemini, Grok, DeepSeek, GLM, Kimi, Qwen

Features

• OpenCode Zen support: auto-discovered zen/<id> catalog merged with models.dev metadata, per-family wire translation (Anthropic Messages / Gemini / Responses / chat-completions) with per-family auth header mapping, free models, and provider="zen" metering
• Responses input to chat messages translation
• instructions and developer roles mapped to system messages
• Function tool schema passthrough
• Custom/freeform tool adaptation (Codex apply_patch works)
• Reasoning content replay across tool-call turns
• Real-time SSE streaming (not synthesized)
• Cached image captioning when tools are present: cheapest catalog vision engine (or a probed local runtime) by default, detail: low input, 30s no-retry budget, MiMo V2.5 fallback, reads metered with kind=vision
• SSRF protection on image URLs (data:image/ and https:// only)
• Configurable body cap, bind address guard, keychain credential resolution
• Local health and model-list endpoints
• Prefix caching: byte-stable request prefixes plus include_usage, with per-model hit ratio on /cache
• Honest usage meter: append-only usage-events.jsonl in the state dir (truncated or empty responses never count as success)
• Upstream retry with bounded exponential backoff on transient failures (429/5xx/network/timeout)
• Ops CLI: doctor (reference-style checks with --fix for safe repairs), smoke-test (marker prompt through the local proxy), support-bundle (JSON schema v1, mode 0600), install (points at the macOS menu bar app; no launchd agent), install-skills, refresh-runtime, and status
• Spawned threads inherit the parent session's model (create_thread; chatgptWorkCloud targets are skipped)
• Correctness contract: empty upstream completions are retried once (a second empty stream answers an empty_completion error), zero-input-token reports are estimated for compaction (OPENCODE_GO_PROXY_ESTIMATE_ZERO_INPUT=0 disables), and keepalive comments run until the stream truly ends without interleaving into data frames
• Auth transport guard (zero config): missing Host answers 400, non-loopback Host answers 403 (unless OPENCODE_GO_PROXY_ALLOW_REMOTE=1), browser-originated requests (Origin / Referer / Sec-Fetch-Site) answer 403, non-JSON POSTs answer 415, and OPTIONS preflight stays blocked
• Verbatim /v1/chat/completions passthrough (stream and non-stream): the upstream status and body are relayed byte-for-byte, including the upstream's own error body, and /v1/messages answers an explicit 400
• Rate-limit harvesting (plan 011): upstream x-ratelimit-* and anthropic-ratelimit-* headers are parsed into per-provider quota snapshots, the latest snapshot per provider is kept, and GET /quota exposes quota-state.json
• Menu bar state contract (plan 013): GET /state returns one JSON document (status, port, upstream, latest quota snapshot, today's turns/tokens, last-7-day token bars, current model) computed from the meter file and quota state
• WebSocket upgrade requests answered with 426 Upgrade Required (desktop app falls back to HTTP streaming)
• zstd-compressed request bodies decompressed (Content-Encoding: zstd; the desktop app sends them)
• Single-port guard in the menu bar app: refuses Start when 8787 is already owned