models: catalog variant settings sent with camelCase keys to OpenAI-compatible providers -> 400 'Extra inputs are not permitted' #41971

Summary

Variant settings from the built-in model catalog are serialized into the request body with their raw camelCase keys (e.g. reasoningEffort). Strict OpenAI-compatible providers reject these as unknown fields, so selecting a standard variant (e.g. max) on models like Fireworks kimi-k3 fails with a 400 before any token is generated. The same body with the snake_case key (reasoning_effort) succeeds.

Environment

• opencode version: 0.0.0-next-17276 (v2, next channel)
• OS: Linux 6.17.0-1022-azure x86_64 (Ubuntu)
• Terminal: xterm-256color
• Shell: /usr/bin/zsh
• Install/channel: npm @opencode-ai/cli@next, background service
• Active plugins: @javargasm/opencode-kiro-auth, @plannotator/opencode@latest, @tarquinen/opencode-dcp, cc-safety-net@latest, envsitter-guard, oc-codex-multi-auth

Reproduction

1. Use a fresh config with no provider overrides (stock catalog).
2. Select a model whose catalog variants carry reasoningEffort, e.g. fireworks-ai/accounts/fireworks/models/kimi-k3 (catalog variant: {"id": "max", "settings": {"reasoningEffort": "max"}}).
3. Send any message (e.g. via a subagent with variant: max).
4. Observe AI.Error: AISDK.doStream: Bad Request from the session runner.

Expected Behavior

Variant settings keys are normalized before serialization (reasoningEffort → reasoning_effort, reasoningSummary → reasoning_summary), matching what OpenAI-compatible endpoints accept, so catalog variants work out of the box on strict providers.

Actual Behavior

The outgoing request contains the camelCase key verbatim. Fireworks returns:

400 invalid_request_error
"Extra inputs are not permitted, field: 'reasoningEffort', value: 'max'"

Confirmed by capturing the exact outgoing body (26 KB capture shows variant settings pass through raw; nested settings like chat_template_kwargs also arrive verbatim) and by direct API calls: identical body with reasoningEffort → 400, with reasoning_effort → 200.

Additional Context

• No user config involved in the failure: the failing kimi-k3 model had zero config overrides; only the shipped catalog metadata was in play.
• OpenAI/Azure/GitHub Copilot families are unaffected — that SDK path recognizes reasoningEffort/reasoningSummary and renders them correctly; the raw pass-through only surfaces on strict OpenAI-compatible endpoints.
• Affected catalogs observed in the live model catalog (all with camelCase reasoningEffort variants): Fireworks (12 models), Together AI (7), Z.AI (1), Cloudflare Workers AI (7), Cloudflare AI Gateway (33), Hugging Face (12), opencode zen (3).
• Workaround: per-model config overlays defining the same variants with snake_case reasoning_effort settings. This requires one overlay per affected model (63 models across 7 providers in our case) and does not remove the camelCase key from the merged catalog.
• Frequency: consistent, not intermittent.
• Verified on v2 next-17276; V1 behavior not verified.