shumengya/sproutclaw
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(ai): add thinkingDisplay option for Anthropic and Bedrock Claude

Browse Source 
Exposes the new ThinkingConfig.display field on Anthropic and Bedrock
Claude providers. Defaults to 'summarized' so Claude Opus 4.7 and Mythos
Preview keep returning thinking text despite Anthropic's silent default
change to 'omitted'. Set to 'omitted' explicitly to skip thinking
streaming for faster time-to-first-text-token.
This commit is contained in:
Mario Zechner
2026-04-16 22:12:06 +02:00
parent 7b45c52807
commit acbf8eca06
4 changed files with 45 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
packages/ai/src/providers/amazon-bedrock.ts
View File

@@ -46,6 +46,8 @@ import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
import { adjustMaxTokensForThinking, buildBaseOptions, clampReasoning } from "./simple-options.js";
import { transformMessages } from "./transform-messages.js";
export type BedrockThinkingDisplay = "summarized" | "omitted";
export interface BedrockOptions extends StreamOptions {
region?: string;
profile?: string;
@@ -56,6 +58,17 @@ export interface BedrockOptions extends StreamOptions {
thinkingBudgets?: ThinkingBudgets;
/* Only supported by Claude 4.x models, see https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-extended-thinking.html#claude-messages-extended-thinking-tool-use-interleaved */
interleavedThinking?: boolean;
/**
* Controls how Claude's thinking content is returned in responses.
* - "summarized": Thinking blocks contain summarized thinking text (default here).
* - "omitted": Thinking content is redacted but the signature still travels back
* for multi-turn continuity, reducing time-to-first-text-token.
*
* Note: Anthropic's API default for Claude Opus 4.7 and Mythos Preview is
* "omitted". We default to "summarized" here to keep behavior consistent with
* older Claude 4 models. Only applies to Claude models on Bedrock.
*/
thinkingDisplay?: BedrockThinkingDisplay;
/** Key-value pairs attached to the inference request for cost allocation tagging.
* Keys: max 64 chars, no `aws:` prefix. Values: max 256 chars. Max 50 pairs.
* Tags appear in AWS Cost Explorer split cost allocation data.
@@ -759,9 +772,12 @@ function buildAdditionalModelRequestFields(
}
if (model.id.includes("anthropic.claude") || model.id.includes("anthropic/claude")) {
// Default to "summarized" so Opus 4.7 and Mythos Preview behave like
// older Claude 4 models (whose API default is also "summarized").
const display: BedrockThinkingDisplay = options.thinkingDisplay ?? "summarized";
const result: Record<string, any> = supportsAdaptiveThinking(model.id)
? {
thinking: { type: "adaptive" },
thinking: { type: "adaptive", display },
output_config: { effort: mapThinkingLevelToEffort(options.reasoning, model.id) },
}
: (() => {
@@ -781,6 +797,7 @@ function buildAdditionalModelRequestFields(
thinking: {
type: "enabled",
budget_tokens: budget,
display,
},
};
})();