[origin-review] anomalyco/opencode#42660 — CHECK #35

Why this matters

Custom OpenAI-compatible providers force manual model ID entry. Server-side discovery from /v1/models (with safe errors) unblocks LiteLLM/LM Studio style setups.

Evidence

Tip 6247b448fa9c791c21c2373c222623871cdd6a0b (base dev):

• Live: cd packages/opencode && bun test ./src/provider/discover.test.ts → 9/9 pass (auth, path preserve, env keys, 401, timeout, invalid URL/format)
• Evidence class: BEHAVIORAL

Invariant: discovery never replaces an existing /api/v1 path with a root /v1/models; failures stay typed rather than 500ing.

Ask (design, light) — prospective; non-blocking

1. Should daily background re-discovery be opt-in for privacy-sensitive air-gapped hosts?
2. Prefer config-hardcoded models always winning over discovered ids — any UI affordance when they diverge?

Fork mirror of a sealed OSS review draft. Not posted upstream.

• origin: feat(provider): add dynamic model discovery for custom providers anomalyco/opencode#42660
• tip: 6247b448fa9c791c21c2373c222623871cdd6a0b
• evidence: BEHAVIORAL — tip packages/opencode bun test ./src/provider/discover.test.ts → 9/9
• pack: 04 draft 010
• dedupe_key: anomalyco/opencode#42660
• github_writes_origin: 0