feat(provider): add OpenCode Zen and OpenCode Go providers

[DevAgarwal2]

Summary

Added support for two new model providers from OpenCode.

OpenCode Zen

Pay-as-you-go access to 35+ curated models:

• GPT 5.x series (5.4-pro, 5.4, 5.3-codex, etc.)
• Claude Opus/Sonnet/Haiku (4.6, 4.5, 4, etc.)
• Gemini 3.x series (3.1-pro, 3-pro, 3-flash)
• MiniMax M2.5, GLM 5/4.7/4.6, Kimi K2.5/K2, Qwen3 Coder, Big Pickle
• API endpoint: https://opencode.ai/zen/v1

OpenCode Go

$10/month subscription for open models:

• GLM-5, Kimi K2.5, MiniMax M2.5
• API endpoint: https://opencode.ai/zen/go/v1

Usage

hermes model opencode-zen:gpt-5.3-codex
hermes model zen:claude-opus-4-6
hermes model go:kimi-k2.5

Files Changed

• hermes_cli/auth.py - Provider registration with API key auth
• hermes_cli/config.py - OPENCODE_ZEN_API_KEY, OPENCODE_GO_API_KEY env vars
• hermes_cli/models.py - Model catalogs and provider aliases
• agent/model_metadata.py - Context lengths for all models
• .env.example - Documentation

API Keys

Get keys at https://opencode.ai/auth

[Allmight97]

Thanks for putting this together. The provider presets and model list are useful prior art, but I think the provider model here is still incomplete for how OpenCode actually works.

Three concrete issues:

1. OpenCode Go and OpenCode Zen should not be modeled as separate credentials. They are separate provider products, but they sit behind the same OpenCode account/auth surface, so Hermes should use one canonical OPENCODE_API_KEY and let the user choose OpenCode Go vs OpenCode Zen explicitly.

2. Onboarding needs to expose both products directly. Right now this adds provider IDs and model aliases, but a new Hermes user still cannot choose OpenCode Go or OpenCode Zen during setup in the same first-class way as the other providers.

3. Hermes currently resolves one runtime transport per provider, but OpenCode needs transport selection by provider + model family, not just provider. Zen is not one OpenAI-compatible endpoint: OpenAI-family traffic goes through a Responses path, Anthropic-family traffic goes through a Messages path, and Gemini-family traffic uses a different path again. Go also differs from Zen in model discovery behavior.

Because of that, I think the clean path is:

• first, a small runtime abstraction PR so Hermes can resolve transport from provider + selected model
• then, a focused provider/onboarding PR for opencode-go and opencode-zen using one canonical OPENCODE_API_KEY

I’m working on that approach from a fresh branch off current upstream main so the follow-up is reviewable and based on the latest code.

[teknium1]

Thanks for this! The code looks clean and follows our provider patterns correctly.

Before we move forward, could you share screenshots showing it working end-to-end? Specifically:

1. The hermes model command showing OpenCode Zen/Go as available providers
2. The model selection step (after picking the provider) showing the models list populating correctly
3. A conversation with the banner visible showing the provider:model string, confirming a response comes back

We don't have OpenCode subscriptions on our end so we can't test this ourselves. Once we can see it working we'll reconsider — thanks!

[DevAgarwal2]

Thanks for this! The code looks clean and follows our provider patterns correctly.

Before we move forward, could you share screenshots showing it working end-to-end? Specifically:

1. The hermes model command showing OpenCode Zen/Go as available providers
2. The model selection step (after picking the provider) showing the models list populating correctly
3. A conversation with the banner visible showing the provider:model string, confirming a response comes back

We don't have OpenCode subscriptions on our end so we can't test this ourselves. Once we can see it working we'll reconsider — thanks!

yeah sure bro

[Allmight97]

I got it working fine as well from the onboarding page. Works fine! Highly recommend adding opencode go/zen support as they are starting to become a serious alternative to Openrouter.

[xamey]

I agree, their plan is highly competitive

[teknium1]

Merged via PR #1650. Your provider registrations, model catalogs, and context lengths were preserved with authorship. I completed the integration by adding the missing setup wizard branches (hermes_cli/setup.py), model flow dispatch (hermes_cli/main.py), and aux model defaults (agent/auxiliary_client.py) that are needed on current main. Thanks @DevAgarwal2!