Feature: add Freebuff provider preset for free DeepSeek (blocked on official OpenAI-compatible endpoint) #1919

Summary

Request: add a Freebuff provider preset to opencodex so Codex / Claude Code / OpenCode users can route free DeepSeek (and Freebuff's other included models) through the proxy.

This is not a config-only change today, because Freebuff does not expose a public OpenAI-compatible endpoint. The blocker is upstream. This issue documents the analysis and asks Freebuff for the one thing that would make a clean integration possible: an official OpenAI-compatible API surface.

Why this is worth doing

Freebuff (freebuff.com, GitHub CodebuffAI/freebuff, Apache-2.0, ~9.7k stars) is a genuinely free, ad-supported coding agent with DeepSeek V4 Pro / V4 Flash, GPT-5.6 Luna, MiniMax M3, and MiMo 2.5 in its model picker. For users who want free DeepSeek through opencodex, it is the most model-rich free option after the existing built-in opencode-free provider (which already serves deepseek-v4-flash-free from https://opencode.ai/zen/v1, no key needed — that path already works today).

A Freebuff preset would extend that free-DeeSeek story with a wider model catalog, so it is a real value-add rather than a duplicate.

Blocker: no public OpenAI-compatible endpoint

Freebuff's CLI/Desktop/Web products talk to the Codebuff backend (codebuff.com / www.codebuff.com) through a session + agent-run flow (/api/v1/agent-runs, instance/session metadata, codebuff_metadata). There is no /v1/chat/completions or /v1/responses surface to point a provider preset at.

The only bridges that exist are third-party community proxies (Quorinex/Freebuff2API, ferdiunal/freebuff-proxy, pingmike2/freebuff2api-wokers). All three have problems that make them unsuitable to bundle into opencodex:

1. No official endpoint — a preset needs a stable baseUrl. None exists.
2. ToS risk — Freebuff2API advertises "Stealth Request Handling: dynamic randomized client fingerprints that mimic official Freebuff SDK behavior" (i.e. ad-stripping by impersonation). Bundling that into opencodex would likely violate Freebuff's terms and get users banned.
3. Responses API gap — opencodex speaks the OpenAI Responses API (/v1/responses). ferdiunal/freebuff-proxy explicitly notes it does not yet serve /v1/responses, only Chat Completions, so it would not work end-to-end with Codex as-is.
4. Auth extraction — every bridge needs a Freebuff auth token pulled from a logged-in session plus token rotation. Fragile and account-bound.

A literal id: "freebuff" preset pointing at a community proxy would be experimental, ToS-violating, and Responses-incompatible. I am deliberately not raising that as a PR.

The clean path (what this issue asks for)

Ask Freebuff to ship an official OpenAI-compatible endpoint (Chat Completions and/or Responses). If/when that exists, the opencodex preset is a one-entry addition to src/providers/registry.ts, modeled on the existing keyless free tier:

{
  id: "freebuff",
  label: "Freebuff",
  adapter: "openai-chat",          // or openai-responses once upstream ships it
  baseUrl: "https://<freebuff-openai-compat-endpoint>/v1",
  authKind: "key",                 // Freebuff account token
  keyOptional: false,
  freeTier: true,
  defaultModel: "deepseek/deepseek-v4-flash",
  models: ["deepseek/deepseek-v4-pro", "deepseek/deepseek-v4-flash", "gpt-5.6-luna", "minimax-m3", "mimo-2.5"],
  note: "Free, ad-supported models (DeepSeek V4 Pro/Flash, GPT-5.6 Luna, MiniMax M3, MiMo 2.5). Requires a Freebuff account token. Ad-supported; see Freebuff terms.",
}

Plus the matching deepseekThinkingEffortsFor / preserveReasoningContentModels wiring already used for opencode-free/deepseek-v4-flash-free so DeepSeek thinking-mode tool-call continuations do not 400.

What can land without upstream

Until Freebuff ships an endpoint, the only safe, in-repo action is documentation: a note in the providers guide that free DeepSeek is already available via opencode-free (opencode-free/deepseek-v4-flash-free, keyless), and that Freebuff is tracked pending an official OpenAI-compatible API. I can open a docs PR for that if maintainers want it.

References

• opencodex provider registry: src/providers/registry.ts (opencode-free entry at ~line 2386 already serves free DeepSeek)
• Freebuff: https://freebuff.com , https://github.com/CodebuffAI/freebuff
• Community bridges (not recommended for bundling): Quorinex/Freebuff2API, ferdiunal/freebuff-proxy, pingmike2/freebuff2api-wokers