Cloudflare Anthropic provider ignores forceAdaptiveThinking, sends incompatible thinking config

[zeke]

Describe the Bug

streamCloudflareAnthropicAi in @flue/runtime's dist/cloudflare/workers-ai-provider.mjs always sets:

thinkingEnabled: Boolean(options?.reasoning)

That sends Anthropic's older budget-based thinking.type.enabled shape. Some models need Anthropic's newer adaptive-thinking API instead (thinking.type.adaptive plus output_config.effort), and for those, this fails every time.

claude-sonnet-5 is one of them. It's registered in @earendil-works/pi-ai's model catalog (dist/providers/data/anthropic.json) with "compat":{"forceAdaptiveThinking":true}. The other two code paths in the same file, the direct Anthropic API path and streamCloudflareResponsesAi/applyResponsesReasoning, both check that flag correctly. Only the Cloudflare AI Gateway path for Anthropic models skips it.

Here's what actually comes back, calling useModel('cloudflare/anthropic/claude-sonnet-5') through the Cloudflare env.AI binding:

Cloudflare AI binding request failed with 400 Bad Request: {"success":false,"result":[],"messages":[],"error":[{"code":7003,"message":"Model execution failed (User Input Error): \"thinking.type.enabled\" is not supported for this model. Use \"thinking.type.adaptive\" and \"output_config.effort\" to control thinking behavior."}] ...}

Expected Behavior

streamCloudflareAnthropicAi should check compat.forceAdaptiveThinking the same way the other two paths in this file already do, and send thinking.type.adaptive plus output_config.effort for models that need it.

Steps to Reproduce

1. In a Cloudflare-target Flue app with an ai binding in wrangler.jsonc, define an agent:

   'use agent';
   import { useModel } from '@flue/runtime';
   
   export function Extract() {
     useModel('cloudflare/anthropic/claude-sonnet-5');
     return 'Say hello.';
   }

2. Dispatch it against a deployed Worker with Workers AI Billing on the AI Gateway set to Unified billing, so a 402 billing error isn't muddying the picture.
3. Watch it 400 with the error above. It never completes.

Workaround

Turning off thinking entirely dodges the bad payload, at the cost of losing extended thinking for this model:

useModel('cloudflare/anthropic/claude-sonnet-5', { thinkingLevel: 'off' });

[rikbrown]

We hit the same failure on claude-opus-5 and filed #618 before spotting this issue — closing ours as a duplicate, but porting the root-cause detail here because it moves the fix location.

streamCloudflareAnthropicAi does not need its own flag check: pi-ai's Anthropic serializer already reads model.compat.forceAdaptiveThinking. The flag just never survives to the model object, because workers-ai-provider.ts drops the catalog compat twice before the serializer runs:

1. gatewayCatalogModels sets compat: undefined when it re-tags the gateway catalog for the binding (its comment defers to toAnthropicGatewayModel).
2. toAnthropicGatewayModel then builds a fresh compat of four binding overrides instead of layering them over the model's own.

Restoring the flow fixes it — keep the catalog compat in gatewayCatalogModels, and spread it under the overrides:

compat: {
  ...model.compat,
  supportsCacheControlOnTools: false,
  supportsEagerToolInputStreaming: false,
  supportsLongCacheRetention: false,
  sendSessionAffinityHeaders: false,
},

We run 2.0.3 in production with exactly this change as a pnpm patch: requests go out as thinking: {type: "adaptive"} and complete. It also restores supportsTemperature: false from the catalog, which prevents the equivalent temperature rejection on Opus 4.7+.

[rahulpatle-sol]

The root cause is in packages/runtime/src/cloudflare/workers-ai-provider.ts. The toAnthropicGatewayModel() function overwrites the model's compat field, discarding forceAdaptiveThinking:

  function toAnthropicGatewayModel(model: Model<Api>): Model<'anthropic-messages'> {
  	return {
  		...model,
  		api: 'anthropic-messages',
  		baseUrl: '',
  		compat: {
+ 			...(model.compat as Record<string, unknown> | undefined),
  			supportsCacheControlOnTools: false,
  			supportsEagerToolInputStreaming: false,
  			supportsLongCacheRetention: false,
  			sendSessionAffinityHeaders: false,
  		},
  	};
  }

By spreading model.compat first, the forceAdaptiveThinking: true flag from the pi-ai catalog (e.g. claude-sonnet-5) is preserved. pi-ai's anthropicMessagesApi().stream() already reads this flag and sends thinking.type.adaptive + output_config.effort when it's set — the Cloudflare adapter just wasn't passing it through.

---