provider-opencode-go

OpenCode Go Chat Completions provider worker behind llm-router. Implements the provider protocol from tech-specs/2026-06-agentic/llm-router.md: provider::opencode_go::stream (SSE chunks → AssistantMessageEvent frames into a router-owned channel), provider::opencode_go::refresh_models (live GET /v1/models id list enriched from a hardcoded curated metadata table → router::models::reconcile), provider::opencode_go::count_tokens (local tiktoken prompt-token estimate, the same surface the other providers expose behind router::count_tokens), and provider::opencode_go::abort (cancels an in-flight upstream request). There is no embedding surface — the OpenCode Go API is Chat Completions only.

Install with iii trigger compose::add worker=provider-opencode-go; the worker takes no per-worker config — credentials and endpoint live in the llm-router configuration entry (providers.opencode_go.api_key, default endpoint https://opencode.ai/zen/go/v1/chat/completions), exactly like provider-openai.

Behavior

• Registration: self-declares via router::provider::register with backoff until acked, and re-declares on the router::ready trigger type. The model slice is populated from live discovery and the declaration carries credential_env_var: OPENCODE_GO_API_KEY.
• Transport: the upstream endpoint is https://opencode.ai/zen/go/v1/chat/completions (overridable via api_url); Chat Completions wire format only.
• Identity binding: the router returns a registration_token on first registration; it is persisted in state (scope provider-opencode-go, key registration_token) and presented on every later register/resolve/reconcile. If that state is lost the router rejects re-registration — the operator must clear the binding on the router side.
• Credentials: resolved per request via router::provider::resolve (config slice → OPENCODE_GO_API_KEY env on the router → none). The key is sent as Authorization: Bearer.
• Liveness: ping at least every 30s of upstream silence; a failed channel write (caller gone / router::abort) drops the SSE receiver and aborts the in-flight HTTP request.
• Errors: 401/403 → auth_expired, 429 → rate_limited, context_length_exceeded → context_overflow, 5xx/network → transient, other 4xx → permanent. No transport retries here — the router owns retry policy.
• Model metadata: discovery fetches the live id list from GET /v1/models (the API carries no capability data) and enriches each id from a hardcoded curated table (src/curated.rs) covering the maintainer's OpenCode Go subscription catalog — the 24 opencode-go entries on models.dev (fetched 2026-08-03) plus hy3-preview, which models.dev does not list and which keeps conservative defaults — context windows, reasoning support/effort levels, tool-call and structured-output capability. Ids the table does not know keep conservative defaults (128K context, no thinking, tools on). Same pattern as provider-openai.
• Reasoning: thinking_level maps to the upstream reasoning_effort when the model's curated effort list accepts the level (e.g. grok-4.5 accepts low/medium/high, deepseek-v4-flash accepts high/max); models that reason without published effort levels, and unknown ids, stream without the field. When the upstream emits reasoning_content deltas they are relayed as thinking blocks; models that never emit them stream text only.
• Structured output: a response_format with a schema maps to strict json_schema mode; without one, json_object mode (the caller must mention "JSON" in the prompt per OpenAI-compatible API rules).
• Prompt caching: upstream-managed; prompt_tokens_details.cached_tokens lands on usage.cache_read when the API reports it.

Tests

cargo test --lib --test schemas                        # unit (pure modules + TCP stubs) + golden schema checks
III_ENGINE_BIN=$(which iii) cargo test --test integration -- --test-threads=1

The integration suite spawns a real engine, the real router (path dep), this provider, and a local stub upstream — no external API calls anywhere.

Running

The binary takes the standard worker CLI flags: --url (engine WebSocket, default ws://127.0.0.1:49134, falls back to the III_URL environment variable), --manifest (print the registry manifest and exit), and --config (accepted but ignored with a warning — provider config comes from the llm-router configuration entry).

Credits

Originally contributed by @faramirezs.