wursor/docs/decisions/0010-openrouter-live-golden.md

10. Golden harness scores live runs through a provider-agnostic LLM client (OpenRouter first)

• Status: Accepted
• Date: 2026-08-15

Context

Phase 0's golden-task spike was the last gate item, stuck at "partial" because no XAI_API_KEY was set and the harness hard-coded api.x.ai. The developer holds an OpenRouter key, and OpenRouter speaks the same OpenAI-compatible chat/completions shape (including tool-calling), so the live run could be unblocked without a Grok-specific key.

Decision

Replace the hard-coded grok-client.ts with a provider-agnostic llm-client.ts that supports grok (x.ai) and openrouter, selected by LLM_PROVIDER. The OpenRouter default model is x-ai/grok-4.6. The harness now passes the site's page slugs into the prompt.

Options considered

• Wait for an XAI_API_KEY and keep the hard-coded x.ai client.
• Hard-code OpenRouter, dropping the x.ai path.
• Provider-agnostic client with grok + openrouter (chosen).

Rejected

• Wait for x.ai — blocks the gate on a key we don't have, for no technical reason.
• OpenRouter-only — the plan (IMPLEMENTATION §3) still names Grok the default adapter; keeping both providers matches that plan and costs one env switch.

Consequences

• The live gate run is scored and passing (gb-01 passed on x-ai/grok-4.6); the Phase 0 golden spike flips to done.
• x-ai/grok-latest is not a callable OpenRouter ID (returns 400); the pinned x-ai/grok-4.6 works and supports tool_choice: required.
• A real finding: without the page-slug list in the prompt, the model guessed page: "home" and failed. Tool-call prompts must always carry site/page context. Sprint 3 must inherit this when it builds api/src/agents/llm-client.ts.
• llm-client.ts here is a spike; Sprint 3 will generalize it (streaming, circuit breaker, fallback) rather than promote this file verbatim.