OpenCode Go gateway¶

Use this guide to connect the OpenCode Go gateway to snodo and check that its models are reachable. The gateway serves model families through different API formats, so Claude and MiniMax need an explicit route rule.

1. What you need¶

You need an OpenCode Go API key and a snodo project where you can edit the provider configuration. This guide uses ocgo as the provider name and OCGO_API_KEY as the environment variable; choose other names if they suit your setup.

2. Store the key¶

Set the environment variable you chose, for example OCGO_API_KEY, in your shell or secret manager. You can instead register the key with snodo config add ocgo <key>, or save a credential reference with snodo config add ocgo --ref 'env:OCGO_API_KEY'.

The provider's api_key_ref can point to an environment variable or command. A literal api_key (including an encrypted provider key) takes precedence over api_key_ref, which takes precedence over api_key_env; a configured key therefore wins over the environment variable.

3. Configure the provider¶

Add this block to your snodo configuration:

providers:
  ocgo:
    api_key_env: OCGO_API_KEY
    base_url: https://opencode.ai/zen/go/v1
    litellm_provider: openai
    model_routes:
      - match: claude-*
        litellm_provider: anthropic
      - match: minimax-*
        litellm_provider: anthropic

api_key_env names the environment variable containing the gateway key.
base_url points snodo at the OpenCode Go API.
litellm_provider selects the default OpenAI-compatible API format.
model_routes overrides that format for matching model IDs.
Each match sends Claude or MiniMax IDs to the Anthropic Messages format.

Rules are checked in order and match the model ID after the local ocgo/ prefix is removed. Other families, such as ocgo/qwen3.6-plus, keep the default OpenAI-compatible route. Without the Claude route rule, a Claude model is sent in the wrong format and its calls fail.

You can optionally set one gateway session per task under the same ocgo provider block:

providers:
  ocgo:
    extra_headers:
      x-opencode-session: "{task_id}"

Snodo substitutes {task_id} with the current task ID; when there is no task ID, it substitutes unknown.

If you have two gateway accounts, configure separate provider blocks, such as ocgo1 and ocgo2, with their own keys to spread requests across accounts.

4. Choose model IDs¶

List the models available through this provider:

snodo models --provider ocgo

Use the model IDs from that list, with the ocgo/ provider prefix, for each role. For example, you can choose:

Role	Example model ID
Coder	`ocgo/claude-haiku-5-5`
Validator	`ocgo/claude-haiku-5-5`
Classifier	`ocgo/qwen3.6-plus`
Recon	`ocgo/qwen3.6-plus`

Claude IDs need the Anthropic route rule above; Qwen uses the default route. Confirm that each example is present in your provider's model listing before using it.

5. Verify the connection¶

Run the live check against your configured models:

snodo models --check

The check makes a cheap live call for each configured model. A Claude model should pass with its Anthropic route rule in place.

6. Troubleshooting¶

Symptom	What to check
Claude model calls fail	Add the `claude-*` route with `litellm_provider: anthropic`; without it, Claude requests use the default OpenAI-compatible format.
MiniMax model calls fail	Add the `minimax-*` route with `litellm_provider: anthropic`.
`snodo models --check` reports HTTP 413	The request is too large for the gateway. Try a smaller request or a model with a lower context requirement.
The key is not picked up	Check the exact `api_key_env` name and that the variable is available to the snodo process. Also check whether a literal or encrypted `api_key`, or `api_key_ref`, is taking precedence.
A model is listed but fails using the wrong API format	Check its ID against `model_routes`. Unmatched families use the block's default `openai` route.