[kelimator1]

Why

Refs #23. OpenCode Go cannot be reached through the existing OpenAI-Compatible API provider: every Go request needs an x-opencode-session header, and kwwk v0.1.100 neither sends it nor registers opencode-go as a provider.

Probed against the live gateway on 2026-10-30:

• POST https://opencode.ai/inference/go/v1/chat/completions without x-opencode-session → 400 MissingSessionID, before auth.
• The same for /go/openai/v1/responses, /go/anthropic/v1/messages and /go/google/....
• With x-opencode-session: test-session, routing passes and the request reaches auth/model errors.
• GET https://opencode.ai/inference/go/v1/models → 200, 28 models with an api family per model.

kwwk carries the opencode-go catalog and display name, but has no providerDirectory entry, no registerStored case, and no occurrence of x-opencode-session (checked the pinned revision and origin/main). This PR adds the provider the same way DeepSeek was added.

What

• New AgentProvider.openCodeGo (opencode-go), API-key sign-in only, oc_sk_… placeholder, in the API Keys group after DeepSeek.
• OpenCodeGoProvider registers each model on the wire its catalog entry names:
  • Chat Completions and Responses → https://opencode.ai/inference/go/openai/v1
  • Anthropic Messages → https://opencode.ai/inference/go/anthropic
  • Gemini → https://opencode.ai/inference/go/google
• OpenCodeGoSessionProvider stamps x-opencode-session from the conversation's session id (a process id for verification and other requests with no conversation), following CopilotRequestHeadersProvider; it forwards session cleanup and native compaction.
• Model discovery uses https://opencode.ai/inference/go/v1/models; known catalog ids keep their compat, reasoning levels, image input and context. Unknown ids fall back to Chat Completions, which Console translates, with a conservative text-only, non-reasoning model.
• Display name, official favicon mark, 12 translations, AGENTS.md decision text and README provider list.
• Tests: per-model wire/base mapping, fallback, auth resolver, header stamping and caller precedence, catalog list/reasoning, key-only sign-in, verification through the existing local-server test, and a live test for GLM-5.2 (Chat) and Claude Haiku 5.5 (Anthropic).

Testing

• make check — passed locally.
• swift build in Packages/ApelinAgent — passed.
• swift test in Packages/ApelinAgent — 528 tests passed, including the new OpenCodeGoProviderTests, ProviderModelsTests, AccountDeskTests and ProviderKeyVerificationTests additions.
• swift test --filter OpenCodeGoLiveTests — compiles and skips without APELIN_LIVE_OAUTH_PATH; not run against a real key here.
• No installable .deb was built locally: this machine has Xcode 26.3 / iOS 26.2, while the repo builds with Xcode 27.1 / iOS 27.1. CI's build job produces apelin-<sha> for device testing.

To run the live test, add an opencode-go entry to a kwwk-format credential file and run:

cd Packages/ApelinAgent
APELIN_LIVE_OAUTH_PATH=~/.kwwk/auth.json swift test --filter OpenCodeGoLiveTests

Not included

• OpenCode Zen (pay-as-you-go): it does not require x-opencode-session and is already reachable through OpenAI-Compatible at https://opencode.ai/inference/v1. A named .openCodeZen can follow separately.
• OAuth: Go is API-key only.
• A kwwk bump: the pinned catalog lacks four live ids and carries a few stale ones. Known ids work; live discovery offers the rest; unknown ids use the fallback. The adapter is deliberately local and can be removed when kwwk registers opencode-go itself.

Follow-ups

• File an upstream kwwk issue/PR for opencode-go registration and the session header.
• Refresh the bundled model catalog with a deliberate kwwk bump.
• Device test the CI .deb: sign in, enable one Chat model and one Anthropic model, run a tool-call round trip on each.