[Fix] GitHub mentions fail with LiteLLM helper models

[roomote-roomote]

Opened on behalf of Matt Rubens. View the task or mention @roomote for follow-up asks.

What changed

• Reused one shared OpenAI-compatible provider builder for task workers and restricted non-task helpers.
• Materialized LiteLLM's adapter, configured endpoint, environment-referenced API key, and selected model catalog for helper inference.
• Kept model/provider selection in resolveEffectiveModelRuntimeEnv() and limited the helper runtime to materializing provider config for already-qualified model IDs.
• Added focused shared-provider and restricted-runtime regression coverage, including a guard against inferring model ownership from endpoint availability.

Why this change was made

Non-task helper inference selected LiteLLM model IDs without registering the endpoint-backed provider, causing GitHub mention routing and similar control-plane calls to target undefined/chat/completions.

Impact

Self-hosted deployments can use LiteLLM models for GitHub routing and other non-task helper calls without supplying a manual OPENCODE_CONFIG_CONTENT workaround. Task-worker provider behavior and the restricted helper permission boundary remain unchanged.

Targeted tests and package type checks pass. The full @roomote/cloud-agents lint command remains blocked by a pre-existing unused eslint-disable warning in src/__tests__/utils.test.ts unrelated to this diff; the repository pre-push checks pass.

Linked work items

Closes #963

[roomote-roomote]

No new code issues found. See task

_{Reviewed 9e614e8}

[tomny-dev]

@roomote-roomote

The shared OpenAI-compatible provider extraction looks good, especially reusing the same provider builder for task workers and restricted helpers.

Could the restricted helper path also reuse the existing model-resolution logic rather than applying LiteLLM inference again inside opencode-runtime.ts?

resolveEffectiveModelRuntimeEnv() already resolves the persisted/runtime model configuration and applies applyImplicitLiteLlmModelPrefix() before returning R_MODEL and the other role models. The new code then independently repeats that decision based only on whether LITELLM_BASE_URL is present:

const isLiteLlmConfigured = isConfiguredEnvValue(env.LITELLM_BASE_URL);
const rawModel = applyImplicitLiteLlmModelPrefix(
  env.R_MODEL?.trim() ?? '',
  isLiteLlmConfigured,
);

Would it be safer to treat the environment passed into buildOpenCodeCliEnv() as already resolved and provider-qualified, then only use the shared mergeOpenAiCompatibleProviderConfig() utility to materialize the selected provider?

That would preserve one source of truth for both concerns:

• resolveEffectiveModelRuntimeEnv() resolves model/provider selection
• mergeOpenAiCompatibleProviderConfig() builds the matching OpenCode provider block

It would also avoid a lower-level runtime inferring provider selection from provider availability. If buildOpenCodeCliEnv() must support callers that bypass the model-runtime resolver, perhaps that fallback should be explicit or separately tested rather than inferred solely from LITELLM_BASE_URL.

[roomote-roomote]

I saw the mention, but I need your GitHub account linked to Roomote before issue and PR comments can start work here. Link it from Settings -> Linked Accounts and then mention me again.

[mrubens]

@roomote go

[roomote-roomote]

I'm on it. I routed this request into the existing PR task so follow-up work stays on one Roomote thread for this PR, and I'll keep updates here. See task

[roomote-roomote]

Addressed the model-resolution feedback in 9e614e8. resolveEffectiveModelRuntimeEnv() remains the single owner of model qualification; the lower-level OpenCode runtime now only materializes provider config and explicitly tests that endpoint availability does not infer model ownership. Helper runtime tests, provider-usage tests, type checks, and pre-push checks pass. See task

[mrubens]

@tomny-dev hows it look?

[tomny-dev]

@mrubens LGTM but not able to test it end to end yet. any plans to allow pr image pushes to ghcr? would make testing changes like this easier

[tomny-dev]

LGTM

built a custom image and tested without the workaround. github pr mentions are working as expected with litellm model

[mrubens]

Thank you @tomny-dev! Btw if you're on Discord can you send me a message? I just got the community Roomote set up and would love to add you so you can talk to it directly.

[roomote-community]

Follow-up is in draft at #966. It adds a maintainer-triggered /publish-images command that publishes amd64 roomote-app and roomote-worker images with immutable PR/SHA tags plus a movable PR tag, then comments the exact references back on the PR. Fork code builds without package-write access; a separate publisher job handles the GHCR push.