feat(llm): add OpenCode Go as a managed provider (chat-only, no Responses/CountTokens)

[mkurkar]

Summary

Adds OpenCode Go as a first-class managed provider behind the existing OpenAI-compatible plumbing, registering it with a custom-provider slot whose AllowedRequests gate refuses every endpoint the gateway does not implement.
Default model is kimi-k3.

Problem

OpenCode Go is an OpenAI-compatible inference gateway that exposes only /v1/chat/completions under https://opencode.ai/zen/go/v1 (Bearer auth).
Routing the current openaicompatible config through Bifrost produced two
hard failures, both of which crashed the parser or the gateway:

1. 404 HTML on token preflight. CountTokens always posts a Responses-API request body to /v1/responses/input_tokens, which the gateway does not serve. The 404 HTML response page is not JSON; Bifrost's parser crashes.
2. 400 on injected OpenAI metadata. When CountTokens succeeded (it did not, see (1)), Bifrost would inject reasoning.effort into the body for reasoning-model names, which the gateway rejects as invalid_request_error per its standard chat-completions contract.

The chat stream itself works once both issues are bypassed. The gateway's 16 model IDs are lowercase-hyphenated (kimi-k3, grok-4.5, glm-5.2,qwen3.7-max, deepseek-v4-pro, etc.); not the upstream-native names.

Approach

Mirror the existing chat-only fallback wrap, but apply it on the primary path so /v1/responses* is blocked there too:

1. Register opencodego as a new ServiceType in the openAICompatibleProviders registry. Default model: kimi-k3.
2. Custom-provider wrap on primary (bifrost/bifrost.go::New): when cfg.Provider == schemas.OpenAI && !cfg.UseResponsesAPI, the primary is registered under a custom name (openai::primary) with chatOnly = true.
   GetConfigForProvider then attaches AllowedRequests{ChatCompletion: true, ChatCompletionStream: true} —
   which (per schemas.AllowedRequests semantics) implicitly refuses every other operation including CountTokens. Bifrost returns unsupported_operation synchronously; the plugin maps that to
   llm.ErrUnsupportedTokenCount, which falls through to llm.EstimateTokens.
3. Route through the wrapped slot. b.provider now stores the registered name (primaryEntry.registeredName()) instead of the bare base type. As a bonus this lands on Bifrost's default: arm in vendored.../providers/openai/chat.go (which skips normalizeReasoningEffort), doubling the chat-path protection.
4. One source of truth for managed URLs. Extracted
   applyProviderURLDefaults so the primary path (new) and the serviceConfigToFallbackEntry path both inject
   https://opencode.ai/zen/go/v1 for opencodego, https://api.cohere.ai/compatibility/v1 for cohere, and
   https://api.mistral.ai/v1 for mistral. normalizeOpenAIBaseURL then strips the trailing /v1 so Bifrost can prepend /v1/chat/completions.
5. Native tools excluded. SupportsNativeTools(ServiceTypeOpenCodeGo)
   returns false explicitly even though the base type is schemas.OpenAI, the gateway does not implement native tools. Request-time filtering
   already short-circuits, but the public predicate stays accurate for built-in-tool fallback gating.
6. Transcriber switch is left alone in bots/bots.go OpenCode Go
   cannot serve /v1/audio/transcriptions, so it cannot be the transcript generator. Existing allowlist (OpenAI / OpenAI-Compatible / Azure) is intentionally narrower than the chat provider list.

Files changed

File	Change
llm/service_types.go	New ServiceTypeOpenCodeGo = "opencodego".
llm/providers.go	Registry entry with DefaultModel: "kimi-k3".
llm/configuration.go	IsValidService requires APIKey only; URL defaults.
bifrost/config.go	MapServiceTypeToProvider→schemas.OpenAI, IsSupported adds opencodego, applyProviderURLDefaults helper, SupportsNativeTools excludes opencodego.
bifrost/bifrost.go	Primary-wrap block; LLM.account field added for tests; provider carries registeredName().
webapp/src/components/system_console/service.tsx	Dropdown option, display-name map, isOpenAIType; API URL and Org ID fields stay hidden (matches mistral UX).

Tests

• bifrost/bifrost_test.go
  • New TestNew_OpenCodeGoPrimaryIsWrappedAsChatOnlyCustom — asserts
    b.provider == "openai::primary", the bare schemas.OpenAI slot is
    not registered, and AllowedRequests rejects
    Responses/ResponsesStream/CountTokens while allowing
    ChatCompletion/ChatCompletionStream.
  • New TestNew_DirectOpenAIAndAzurePrimariesRemainStandard — guards
    the inverse: real OpenAI/Azure primaries keep the bare base-type slot,
    not a custom slot.
  • Extended TestProviderAccount_ChatOnlyCustomConfig with
    AllowedRequests.CountTokens assertion.
• bifrost/config_test.go — opencodego entries for SupportsNativeTools
  and FilterNativeToolsForServiceType.
• llm/providers_test.go, llm/configuration_test.go — registry and
  IsValidService cases.
• bifrost/bifrost_test.go::TestCountTokensOmitsMaxOutputTokens updated:
  synthetic Provider: schemas.OpenAI, UseResponsesAPI: false configs are
  no longer reachable from NewFromServiceConfig (which forces true),
  so the test now uses the production value to keep exercising the
  body-shape assertion instead of the chat-only gate.

All ./llm/... ./bifrost/... ./bots/... ./config/... ./api/... tests pass.

Risk & follow-ups

• Behavior expansion. The wrap condition is cfg.Provider == schemas.OpenAI && !cfg.UseResponsesAPI, which also covers openaicompatible with the operator toggle off. That is arguably the right behavior (same gateway pitfall) and is consistent with how the fallback path already worked, but it is a behavior change for any openaicompatible bot created before this PR. Worth a CHANGELOG entry.
• No model auto-discovery. supportsModelFetching in webapp/src/components/system_console/service.tsx does not include
  opencodego, so admins must type the model name. OpenCode Go does serve a /v1/models listing — adding it would be a one-liner behind a feature flag if desired.
• Documentation. docs/providers.md should gain a short OpenCode Go section. Out of scope here; happy to follow up.
• Transcriber parity. If/when OpenCode Go adds /v1/audio/transcriptions, the switch in bots/bots.go::GetTranscribe can grow a case.

Verification

• go test ./llm/... ./bifrost/... ./bots/... ./config/... ./api/... —
  green on Go 1.26.5.
• Bundle built: make dist produces a 113 MB tar.gz with five
  cross-compiled server binaries plus the webapp bundle that contains the
  new "OpenCode Go" dropdown entry.

Screenshot

Summary by CodeRabbit

• New Features

  • Added support for configuring and using the OpenCode Go service.
  • Added OpenCode Go to the service selection menu with appropriate defaults, validation, and display labels.
  • OpenCode Go supports chat completions and streaming while excluding unsupported native tools and request types.
  • Improved provider diagnostics and standardized service URL handling.

• Bug Fixes

  • Prevented unsupported requests from being sent to chat-only providers.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI (base), Organization UI (inherited)

Review profile: CHILL

Plan: Pro Plus

Run ID: 2187b2e0-4871-4f26-ab5e-8984c45bd03f

📥 Commits

Reviewing files that changed from the base of the PR and between 775bdb8 and a7c3150.

📒 Files selected for processing (2)

• webapp/src/components/system_console/service.tsx
• webapp/src/i18n/en.json

🚧 Files skipped from review as they are similar to previous changes (1)

• webapp/src/components/system_console/service.tsx

---

📝 Walkthrough

Walkthrough

Adds OpenCode Go as a chat-only OpenAI-compatible service, including validation, provider registration, URL defaults, native-tool filtering, Bifrost wrapping, tests, and system console support.

Changes

OpenCode Go integration

Layer / File(s)	Summary
Service type and provider contract llm/service_types.go, llm/configuration.go, llm/providers.go, llm/*_test.go	Defines ServiceTypeOpenCodeGo, requires an API key, registers the kimi-k3 default model, and validates provider configuration.
Configuration routing and capability rules bifrost/config.go, bifrost/config_test.go	Maps OpenCode Go to OpenAI, applies its default base URL, marks it unsupported for native tools, and includes it among supported services.
Chat-only Bifrost registration bifrost/bifrost.go, bifrost/bifrost_test.go	Wraps non-Responses OpenAI-base primaries as chat-only custom providers, records the resolved account, and verifies fallback and direct-provider behaviour.
Service console configuration webapp/src/components/system_console/service.tsx, webapp/src/i18n/en.json	Adds OpenCode Go to the service selector and display mapping, treats it as OpenAI-like, and hides organization or account identifier fields.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Admin
  participant ServiceConsole
  participant ServiceConfig
  participant Bifrost
  participant OpenCodeGo
  Admin->>ServiceConsole: select OpenCode Go and enter API key
  ServiceConsole->>ServiceConfig: submit service configuration
  ServiceConfig->>Bifrost: map provider and apply default URL
  Bifrost->>Bifrost: register chat-only custom provider
  Bifrost->>OpenCodeGo: send chat completion request

Loading

Suggested labels: Setup Cloud Test Server

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately captures the main change: adding OpenCode Go as a managed provider with chat-only restrictions and no Responses or CountTokens support.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@webapp/src/components/system_console/service.tsx`:
- Line 52: Update the service display-map entry and dropdown fallback for the
“opencodego” service to use intl.formatMessage(...) with an extracted message
instead of the hardcoded “OpenCode Go” string. Ensure both user-facing paths
reuse the i18n-derived label and remove the hardcoded values.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI (base), Organization UI (inherited)

Review profile: CHILL

Plan: Pro Plus

Run ID: 05d3ae02-8ec2-4213-a886-12e70e203d13

📥 Commits

Reviewing files that changed from the base of the PR and between aaa3e4f and 775bdb8.

📒 Files selected for processing (10)

• bifrost/bifrost.go
• bifrost/bifrost_test.go
• bifrost/config.go
• bifrost/config_test.go
• llm/configuration.go
• llm/configuration_test.go
• llm/providers.go
• llm/providers_test.go
• llm/service_types.go
• webapp/src/components/system_console/service.tsx