feat(openai): add GPT-5.6 Sol/Terra/Luna models

[zephg]

What

Adds first-class support for the GPT-5.6 family (GA 2026-07-09) to the native OpenAI provider — Sol (flagship), Terra (balanced), Luna (fast/cheap) — and, transitively, OpenRouter. gpt-5.6-sol becomes the default OpenAI model.

Model	API id	Context / Output	In / Out ($/1M)	Cache read	Reasoning effort
Sol	gpt-5.6-sol	1.05M / 128K	5 / 30	0.5	none…max (default high)
Terra	gpt-5.6-terra	1.05M / 128K	2.5 / 15	0.25	none…max (default medium)
Luna	gpt-5.6-luna	1.05M / 128K	1 / 6	0.1	none…max (default low)

How

• packages/types/src/providers/openai.ts — three openAiNativeModels entries templated on the existing gpt-5.5; adds the new max effort level, verbosity, and per-model pricing. Default → gpt-5.6-sol.
• packages/types/src/providers/index.ts — getProviderDefaultModelId("openai-native") now returns openAiNativeDefaultModelId instead of a stale hardcoded "gpt-4o", so the loading-time fallback also honors Sol.
• OpenRouter — no code change: parseOpenRouterModel already mirrors the static effort array for any openai/* id, so openai/gpt-5.6-* surfaces max automatically; caching auto-detects from API pricing.

Everything downstream (handler selection, cost, UI picker, reasoning dropdown) is data-driven off ModelInfo, so no handler/UI code was needed. max is already plumbed end-to-end (precedent: Anthropic Opus 4.8).

Scope / deferred (deliberate)

• Bedrock deferred — GPT-5.6 is not GA on Bedrock and AWS hasn't published openai.gpt-5-6-* IDs; our existing openai.gpt-oss-* Bedrock entries also expose no reasoning. Revisit when AWS ships it.
• Verify against OpenAI's official pricing page before relying on cost at scale: the >272K long-context surcharge is mirrored from gpt-5.5 (drop if removed for this family); flex/priority service tiers, the new 1.25× cache-write price, and 30-min cache retention are omitted for now (base in/out/cache-read pricing is high-confidence).
• Out of scope: Programmatic Tool Calling (separate Responses-API feature), the openai-codex provider, Requesty.

Test plan

• pnpm --filter @roo-code/types check-types → pass (catalog conforms to ModelInfo).
• Targeted specs 96/96 pass, incl. new: Sol getModel(), "effort":"max" request-body passthrough (openai-native.spec.ts), and openai/gpt-5.6-sol effort-mirror (openrouter fetcher spec).
• Pre-existing 22-failure reasoning/model-params baseline unchanged (no regressions).
• Pre-commit lint/prettier hooks passed.

Remaining runtime check (needs a live key): send a max-effort request to gpt-5.6-sol end-to-end and confirm a 200 — the unit test confirms we emit "effort":"max"; only live API acceptance is unverified.

🤖 Generated with Claude Code