fix(runtime): bound Anthropic-compatible output limits

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000

[Astro-Han]

Summary

• scope catalog-derived output limits to requests that actually use Anthropic Messages, where SDK-unknown models otherwise fall back to max_tokens: 4096
• keep per-model GitHub Copilot routing and output-limit selection on the same account-advertised protocol fact
• reserve fixed Anthropic thinking tokens inside each model's advertised wire output limit, so kimi-for-coding sends max_tokens: 32768 instead of 33792
• leave OpenAI-compatible adapters such as Mistral on their provider-owned defaults, while preserving connection-advertised limits over catalog metadata

This is a focused follow-up to #1209. Applying catalog maximums to every provider caused models such as mistral-large-latest, whose advertised output maximum equals its full context window, to receive an unsafe per-request generation limit. Gating at the final request protocol keeps the explicit limit where the Anthropic SDK requires one, including Kimi K3, SDK-unknown MiniMax models, and account-routed GitHub Copilot Anthropic models.

Verification

• npm run build && node --test dist/__tests__/ai-sdk-backend.test.js dist/__tests__/computer-use-provider-protocol.test.js dist/__tests__/model-factory-thinking.test.js — 197 passed, 0 failed
• npm test in packages/runtime — 2149 passed, 7 skipped, 0 failed
• scripted HTTP coverage confirms K3 sends max_tokens: 131072, kimi-for-coding sends max_tokens: 32768 after the SDK accounts for its 1024-token fixed thinking budget, MiniMax-M3 preserves max_tokens: 128000, and an SDK-unknown GitHub Copilot Anthropic model honors its account-advertised max_tokens: 128000