[Feature] Add Command Code (commandcode.ai) as a provider — enables the $1/mo Go plan

[minpeter]

Summary

Request to add Command Code (a coding agent + model gateway by Langbase) as a provider. The interesting part is the Go plan: $1/month with $10 of included credits and deal multipliers that push effective usage to ~$40 on selected OSS models (pricing, pricing limits). Roster includes DeepSeek V4 Pro, Kimi K2.7 Code, GLM-5.2, MiniMax M3, Qwen 3.7 Max, MiMo V2.5 Pro, Nemotron 3 Ultra, etc. (21 open-weight models, several with 1M context).

The Go plan is genuinely the best $1 in the space right now, and CLIProxyAPI is the obvious home for it — but there's a wrinkle that's worth flagging up front.

The wrinkle: Go plan works via /alpha/generate, not /provider/v1/*

The gateway exposes three generation endpoints:

Endpoint	Shape	Plan
POST /provider/v1/messages	Anthropic Messages	Pro+ only
POST /provider/v1/chat/completions	OpenAI Chat Completions	Pro+ only
POST /alpha/generate	CLI's own custom envelope	Any plan, incl. Go

On the Go plan, both /provider/v1/* paths return:

{ "error": { "message": "Your Go plan doesn't include API access. Upgrade to Pro or higher…", "code": "upgrade_required" } }

/alpha/generate is the endpoint the cmd CLI itself uses on every turn, so it is not plan-gated. The catch: it speaks a custom envelope, not OpenAI or Anthropic shape. So serving Go-plan users means a translator, exactly the kind of thing CLIProxyAPI does well.

Auth is a plain user_... API key (same key as the CLI, from https://commandcode.ai/settings/billing) — no OAuth needed, unlike Codex/Claude/Gemini.

The protocol is already reverse-engineered and battle-tested

There's a mature reference implementation with a full wire-protocol spec:

• Reference: safzanpirani/pi-commandcode-provider — working OpenAI→envelope + NDJSON→SSE translator.
• Wire protocol doc: docs/wire-protocol.md — strict config envelope, Vercel-AI-SDK ModelMessage[] schema, tool format, every NDJSON event type, usage shape.

Request

POST https://api.commandcode.ai/alpha/generate
Authorization: Bearer user_…
x-cli-environment: production
x-command-code-version: 0.40.0   # must track the installed CLI; stale → rejected
x-session-id: <uuid>

{
  "config": { /* schema-strict; all fields required, neutral defaults OK */ },
  "memory": "", "taste": null, "skills": null, "permissionMode": "standard",
  "params": {
    "model": "deepseek/deepseek-v4-pro",
    "system": "…",
    "messages": [ /* Vercel AI SDK ModelMessage[] — NOT OpenAI, NOT Anthropic */ ],
    "tools": [ /* name + input_schema (JSON Schema) */ ],
    "max_tokens": 32000, "stream": true
  }
}

Response = newline-delimited JSON, NOT SSE

{"type":"text-delta","id":"txt-0","text":"…"}
{"type":"tool-input-delta","id":"call_…","delta":"<json chunk>"}
{"type":"finish-step","finishReason":"stop|length|tool-calls","usage":{...}}

Known sharp edges (verified from open issues in other routers)

• stream:false is rejected (400/ECONNRESET) — must force stream:true and buffer server-side. decolua/9router#1840, fixed in #1843.
• NDJSON lines on large tool outputs overflow bufio.Scanner's default limit — needs an unbounded reader.
• The config envelope is schema-strict; the gateway's 400 usefully lists the exact missing field paths.

Demand is real (this is not hypothetical)

A code/issue search of this repo for commandcode returns zero matches today. But the ecosystem is actively building it independently:

Project	Stars	Status
decolua/9router	18k	Active router; fixing Command Code stream bugs this week
can1357/oh-my-pi	13.9k	PR #1360 "add Command Code provider" (this week)
safzanpirani/pi-commandcode-provider	—	Reference protocol impl, resynced to CLI 0.39.1 four days ago
HaeMeto/Command-Code-AI-Proxy	10	Go proxy on /alpha/generate
dev2k6/command-code-proxy-server	8	Go proxy, but Pro-only (/provider/v1/*) so it can't serve Go plan

Same feature request also open elsewhere: anomalyco/opencode#26338. The knowledge and the demand both exist; what's missing is a canonical home, which CLIProxyAPI is the natural fit for.

Stability note (being candid)

/alpha/generate is an undocumented CLI-internal endpoint, and the command-code npm package is at 0.40.0 and moves frequently — the protocol drifts and the x-command-code-version header must track it. That's the same risk class as this project's existing OAuth/CLI-wrapping providers (Claude Code, Codex, Gemini CLI), and centralizing it here with CI is exactly what would keep consumers current instead of each fork re-deriving on every CLI bump.

Proposed contribution

I've gathered a full evidence bundle (protocol spec, reference impl source, model roster, credit multipliers, stability assessment) and would like to open a PR. Two routes, and I'd like your steer on which you prefer:

1. Example provider under examples/ — a Go Command Code provider built on the v7 SDK pattern already shown in examples/custom-provider/main.go (sdktr.Register(openai.chat → commandcode, reqXform, respXform) + an Executor hitting /alpha/generate + model registration in OnAfterStart). This path is open to contributors.

2. Built-in provider — I noticed pr-path-guard.yml blocks internal/translator/** changes in PRs, so a built-in translator would need your team to land it. If you'd prefer it built-in, I'm happy to hand over a ready diff + tests and let you commit, or however you prefer to work.

My default plan is route 1 (example provider) unless you'd rather it be built-in.

---

Question for maintainers: preferred route (1 example provider vs 2 built-in), and any provider naming convention you'd like me to follow (e.g. commandcode, command-code, cmd)?

I'll open the PR against dev per the retarget workflow once you confirm direction.

cc maintainers

[peach0x33a]

in my opinion, it would be better to use a plugin system to provide support for commandcode.

[nasrulhadi]

Because of this limitation, I ended up building my own proxy. Feel free to check it out nasrulhadi/proxy-commandcode