Add OpenCode Go provider

[mulfyx]

Add OpenCode Go provider

Closes #46.

Summary

• Add OpenCode Go as an API-key provider behind the existing Anthropic-compatible server.
• Route each model to the OpenCode Go wire protocol it actually supports: OpenAI-compatible chat/completions or Anthropic-compatible messages.
• Support both canonical opencode-go/<model-id> names and unambiguous bare model IDs.
• Reuse the existing OpenAI-compatible translation path for chat models and stream Anthropic-compatible models directly.
• Add configuration, model catalog registration, CLI model listing, documentation, and regression coverage.

Routing and authentication

OpenCode Go splits one subscription catalog across two upstream APIs:

Upstream	Authentication	Models
/chat/completions	Authorization: Bearer ...	glm-5.2, glm-5.1, glm-5, kimi-k2.7-code, kimi-k2.6, kimi-k2.5, deepseek-v4-pro, deepseek-v4-flash, mimo-v2.5, mimo-v2.5-pro
/messages	x-api-key + anthropic-version	minimax-m3, minimax-m2.7, minimax-m2.5, qwen3.7-max, qwen3.7-plus, qwen3.6-plus, qwen3.5-plus

The API key is resolved in this order:

1. CCP_OPENCODE_API_KEY
2. OPENCODE_API_KEY
3. opencode.apiKey in config.json

CCP_OPENCODE_BASE_URL or opencode.baseUrl can override the default https://opencode.ai/zen/go/v1 base URL.

The canonical opencode-go/ prefix can always be used to force this provider. Bare IDs are also registered where they do not conflict with an existing provider. In particular, bare kimi-k2.6 remains assigned to the existing Kimi provider, while opencode-go/kimi-k2.6 selects OpenCode Go.

The current OpenCode documentation lists 14 endpoint assignments. Live /models discovery also returned three older IDs that still completed real requests: glm-5, kimi-k2.5, and qwen3.5-plus. Three other discovered IDs were deliberately not advertised because direct requests failed: mimo-v2-pro, mimo-v2-omni, and hy3-preview.

Protocol handling

• OpenAI-compatible models reuse the Kimi request translation code, with provider-specific Kimi fields kept out of the generic request path.
• Dynamic system-prompt sections that recent Claude Code versions append as role: system messages are merged with the top-level system prompt before translation.
• Their upstream SSE is decoded and translated incrementally. Claude Code receives each completed Anthropic event without waiting for the entire OpenCode response to finish.
• Anthropic-compatible models preserve the incoming Messages shape, replace the provider-qualified model name with the upstream ID, and pass streaming responses through.
• minimax-m3 receives OpenCode's default thinking: {"type":"adaptive"} when the caller does not provide a thinking option; an explicit caller value is preserved unchanged.
• Streaming requests have a connection timeout but no fixed total-response deadline, so an active long-running response is not terminated after 120 seconds.
• /v1/messages/count_tokens is handled locally and does not consume OpenCode Go quota.
• Missing credentials produce a clear Anthropic-style authentication error.
• Upstream 429 responses map to rate_limit_error and preserve Retry-After.
• Buffered upstream responses are capped at 8 MiB.

The provider intentionally keeps a static reviewed model catalog and does not add a runtime models.dev dependency. This change also does not introduce a new output-token truncation policy.

Tool-call streaming fix

End-to-end Claude Code testing exposed a pre-existing bug in the shared OpenAI-compatible SSE reducer. Upstream tool-call argument fragments are keyed by the upstream tool index, but the reducer previously looked up later fragments using the generated Anthropic content-block index. If a thinking or text block appeared before the tool call, Claude Code could receive an empty tool input such as Read {} or Bash {}.

The reducer now tracks both indices independently and matches argument fragments by the upstream tool index. A regression test covers a reasoning block followed by a fragmented tool call. This also hardens the existing Kimi translation path.

Verification

• cargo test --all-targets: 604 passed, 0 failed.
• OpenCode provider unit tests cover endpoint selection, credential precedence, provider-qualified IDs, authentication headers, error mapping, response limits, and token counting.
• A deterministic HTTP integration test holds the upstream terminal event open and proves that the first translated text event reaches the downstream client before upstream completion.
• Incremental decoder tests cover every byte split, LF/CRLF framing, incomplete frames, invalid UTF-8, frame bounds, reasoning, and fragmented tool-call arguments across chunks.
• Request-shape tests cover the qwen3.5-plus Messages route and verify that the MiniMax adaptive-thinking fallback does not overwrite caller-supplied thinking.
• Real Claude Code CLI 2.1.207 smoke tests on the current implementation passed for all 17 supported models.
• Each model completed a multi-turn agent flow with thinking, Read, Write, Bash auto-review, Bash execution, and final result verification.
• The generated implementation used a handwritten stable insertion sort and was validated both by execution and by an AST check that rejected sorted() and .sort().
• Every model trace contained incremental stream_event output. The run and one clean retry after a Claude Code auto-review timeout recorded 115 successful OpenCode Go upstream requests across all 17 models, with no failed or non-2xx requests.
• A final clean-branch Claude Code smoke covered glm-5.2, qwen3.5-plus, and minimax-m3 as representatives of both upstream protocols and the adaptive-thinking path; all three streamed successfully with HTTP 200 and no proxy errors.
• cargo fmt --all -- --check passes.
• git diff --check passes.

References

• Feature request #46
• OpenCode Go documentation
• OpenCode Go endpoint assignments in the documentation source
• OpenCode streams model output through streamText and consumes fullStream
• qwen3.5-plus is declared with the Anthropic provider protocol
• OpenCode's default adaptive-thinking transform for minimax-m3
• OpenCode provider configuration