feat(one-reason): support configuring model reasoning effort

[yaonyan]

What

Adds support for configuring a model's reasoning/thinking effort through environment variables and reason.json, without code changes:

• ONE_REASON_REASONING_EFFORT (fallbacks: ONE_REASONING_EFFORT, then REASONING_EFFORT in reason.json)
• Valid values: off, low, medium, high (unset = provider default)
• reason auth can now persist REASONING_EFFORT, and --help/README/.env.example document it

Why

Most OpenAI-compatible reasoning APIs expose reasoning_effort (plus vendor-specific fields), but the package previously had no way to set it. The fields are not universal, so the mapping is provider-aware and deliberately best-effort:

Effort	OpenAI-compatible	OpenAI (native)	Anthropic
low / medium / high	reasoning_effort: low|medium|high	reasoningEffort: low|medium|high	Claude 4.6+ / Claude 5: adaptive thinking + effort; older Claude models: enabled + fixed budgetTokens (2048/8192/16384)
off	reasoning_effort: "none"	reasoningEffort: "minimal"	thinking: { type: "disabled" }

Anthropic handling is split by model generation because the two thinking modes are not interchangeable: type: "enabled" + budgetTokens is rejected on Claude Opus 4.7+ / Claude 5, while Claude 4.6 and earlier have no effort parameter, so each effort level maps to a fixed thinking budget. off also requires @ai-sdk/anthropic >= 3.0.93 — earlier versions silently dropped thinking: { type: "disabled" } from the request, so thinking-stopped-by-default models like Sonnet 5 would keep thinking.

Caveats documented in the README: some providers ignore reasoning_effort (e.g., Zhipu GLM gates thinking via its own thinking.type), and always-reasoning models (e.g., MiniMax M2) may reject or ignore the parameter. Use a non-thinking model when thinking must be guaranteed off.

Verification

• New request-body regression tests stub fetch and assert what the AI SDK actually sends for anthropic, openai-compatible, and openai, covering off/low/medium/high, adaptive vs legacy Anthropic models, and gateway-prefixed model ids. The off Anthropic test fails against the previously locked @ai-sdk/anthropic@3.0.68.
• Local OpenAI-compatible captures: unset → no field; low/medium/high → forwarded verbatim; off → "none"; generic ONE_REASONING_EFFORT fallback works; uppercase values normalized; invalid values fail fast with a clear error.
• Config precedence confirmed: reason.json REASONING_EFFORT: "high" sends reasoning_effort: "high"; setting ONE_REASON_REASONING_EFFORT=off overrides it to "none".
• Live against the internal OpenAI-compatible gateway with kimi-k2.6: off ≈ 2s, low/high ≈ 21–25s, all returning schema-valid JSON.
• Vendor probes showed GLM silently keeps thinking under reasoning_effort=none, while Kimi K2.6 honors it; native OpenAI mapping verified against the pinned AI SDK schemas.
• pnpm --recursive typecheck, vitest run, oxfmt --check, and the one-reason build all pass.

[yaonyan]

Addressed both P2 findings from the review.

1. high/medium/low now actually enable Anthropic thinking

• Claude 4.6+ / Claude 5 models (Sonnet 4.6/5, Opus 4.6/4.7/4.8/5, Fable 5): anthropic: { thinking: { type: "adaptive" }, effort } → wire body thinking: { type: "adaptive" } + output_config: { effort: ... }.
• Claude 4.6 and earlier (4/4.1/4.5, Haiku 4.5, 3.x): anthropic: { thinking: { type: "enabled", budgetTokens } } (2048/8192/16384 for low/medium/high). Effort is an adaptive-only parameter, so it is never sent to legacy models.
• Selection is by model generation, not a blanket adaptive mode.

2. off now reaches the wire

• Bumped @ai-sdk/anthropic to ^3.0.116 (>= 3.0.93); earlier versions only serialized thinking for enabled/adaptive and dropped disabled.
• Captured request body for off is now thinking: { type: "disabled" } for both adaptive and legacy models.

Regression tests

• Added packages/one-reason/tests/reasoning-effort.test.ts (vitest) that stubs fetch and asserts the actual HTTP body sent by the pinned AI SDK providers (anthropic, openai-compatible, openai) across off/low/medium/high, adaptive vs legacy Anthropic models, and gateway-prefixed ids. The off Anthropic case fails against @ai-sdk/anthropic@3.0.68.
• pnpm --recursive typecheck, vitest run, oxfmt --check, and the one-reason build all pass.