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

Support adaptive thinking for Anthropic-compatible aliases

Browse Source 
closes #4790
This commit is contained in:
Mario Zechner
2026-05-22 18:30:04 +02:00
parent 7002c68f8b
commit d801d88a11
11 changed files with 239 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
packages/ai/src/models.generated.ts
View File

@@ -1723,6 +1723,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "anthropic",
baseUrl: "https://api.anthropic.com",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"max"},
input: ["text", "image"],
@@ -1741,6 +1742,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "anthropic",
baseUrl: "https://api.anthropic.com",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"xhigh"},
input: ["text", "image"],
@@ -1827,6 +1829,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "anthropic",
baseUrl: "https://api.anthropic.com",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
input: ["text", "image"],
cost: {
@@ -2826,6 +2829,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "cloudflare-ai-gateway",
baseUrl: "https://gateway.ai.cloudflare.com/v1/{CLOUDFLARE_ACCOUNT_ID}/{CLOUDFLARE_GATEWAY_ID}/anthropic",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"max"},
input: ["text", "image"],
@@ -2844,6 +2848,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "cloudflare-ai-gateway",
baseUrl: "https://gateway.ai.cloudflare.com/v1/{CLOUDFLARE_ACCOUNT_ID}/{CLOUDFLARE_GATEWAY_ID}/anthropic",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"xhigh"},
input: ["text", "image"],
@@ -2896,6 +2901,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "cloudflare-ai-gateway",
baseUrl: "https://gateway.ai.cloudflare.com/v1/{CLOUDFLARE_ACCOUNT_ID}/{CLOUDFLARE_GATEWAY_ID}/anthropic",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
input: ["text", "image"],
cost: {
@@ -3924,6 +3930,7 @@ export const MODELS = {
provider: "github-copilot",
baseUrl: "https://api.individual.githubcopilot.com",
headers: {"User-Agent":"GitHubCopilotChat/0.35.0","Editor-Version":"vscode/1.107.0","Editor-Plugin-Version":"copilot-chat/0.35.0","Copilot-Integration-Id":"vscode-chat"},
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"max"},
input: ["text", "image"],
@@ -3943,6 +3950,7 @@ export const MODELS = {
provider: "github-copilot",
baseUrl: "https://api.individual.githubcopilot.com",
headers: {"User-Agent":"GitHubCopilotChat/0.35.0","Editor-Version":"vscode/1.107.0","Editor-Plugin-Version":"copilot-chat/0.35.0","Copilot-Integration-Id":"vscode-chat"},
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"xhigh"},
input: ["text", "image"],
@@ -3981,6 +3989,7 @@ export const MODELS = {
provider: "github-copilot",
baseUrl: "https://api.individual.githubcopilot.com",
headers: {"User-Agent":"GitHubCopilotChat/0.35.0","Editor-Version":"vscode/1.107.0","Editor-Plugin-Version":"copilot-chat/0.35.0","Copilot-Integration-Id":"vscode-chat"},
compat: {"forceAdaptiveThinking":true},
reasoning: true,
input: ["text", "image"],
cost: {
@@ -7226,6 +7235,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "opencode",
baseUrl: "https://opencode.ai/zen",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"max"},
input: ["text", "image"],
@@ -7244,6 +7254,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "opencode",
baseUrl: "https://opencode.ai/zen",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"xhigh"},
input: ["text", "image"],
@@ -7296,6 +7307,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "opencode",
baseUrl: "https://opencode.ai/zen",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
input: ["text", "image"],
cost: {
@@ -13369,6 +13381,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "vercel-ai-gateway",
baseUrl: "https://ai-gateway.vercel.sh",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"max"},
input: ["text", "image"],
@@ -13387,6 +13400,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "vercel-ai-gateway",
baseUrl: "https://ai-gateway.vercel.sh",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
thinkingLevelMap: {"xhigh":"xhigh"},
input: ["text", "image"],
@@ -13439,6 +13453,7 @@ export const MODELS = {
api: "anthropic-messages",
provider: "vercel-ai-gateway",
baseUrl: "https://ai-gateway.vercel.sh",
compat: {"forceAdaptiveThinking":true},
reasoning: true,
input: ["text", "image"],
cost: {

62
packages/ai/src/providers/anthropic.ts
View File

@@ -164,7 +164,9 @@ export type AnthropicThinkingDisplay = "summarized" | "omitted";
const FINE_GRAINED_TOOL_STREAMING_BETA = "fine-grained-tool-streaming-2025-05-14";
const INTERLEAVED_THINKING_BETA = "interleaved-thinking-2025-05-14";
function getAnthropicCompat(model: Model<"anthropic-messages">): Required<AnthropicMessagesCompat> {
function getAnthropicCompat(
model: Model<"anthropic-messages">,
): Required<Omit<AnthropicMessagesCompat, "forceAdaptiveThinking">> {
// Auto-detect session affinity and cache control support from provider
const isFireworks = model.provider === "fireworks";
const isCloudflareAiGatewayAnthropic =
@@ -181,29 +183,34 @@ function getAnthropicCompat(model: Model<"anthropic-messages">): Required<Anthro
export interface AnthropicOptions extends StreamOptions {
/**
* Enable extended thinking.
* For Opus 4.6 and Sonnet 4.6: uses adaptive thinking (model decides when/how much to think).
* For adaptive thinking models: the model decides when/how much to think.
* For older models: uses budget-based thinking with thinkingBudgetTokens.
* Default: undefined (thinking is omitted unless `streamSimpleAnthropic()` maps
* a simple reasoning level to this option, or callers set it explicitly).
*/
thinkingEnabled?: boolean;
/**
* Token budget for extended thinking (older models only).
* Ignored for Opus 4.6 and Sonnet 4.6, which use adaptive thinking.
* Ignored for adaptive thinking models.
* Default: 1024 when `thinkingEnabled` is true and no budget is provided.
*/
thinkingBudgetTokens?: number;
/**
* Effort level for adaptive thinking (Opus 4.6+ and Sonnet 4.6).
* Effort level for adaptive thinking models.
* Controls how much thinking Claude allocates:
* - "max": Always thinks with no constraints (Opus 4.6 only)
* - "xhigh": Highest reasoning level (Opus 4.7)
* - "high": Always thinks, deep reasoning (default)
* - "high": Always thinks, deep reasoning
* - "medium": Moderate thinking, may skip for simple queries
* - "low": Minimal thinking, skips for simple tasks
* Ignored for older models.
* Default: omitted unless `streamSimpleAnthropic()` maps a simple reasoning
* level to this option.
*/
effort?: AnthropicEffort;
/**
* Controls how thinking content is returned in API responses.
* - "summarized": Thinking blocks contain summarized thinking text (default here).
* - "summarized": Thinking blocks contain summarized thinking text.
* - "omitted": Thinking blocks return an empty thinking field; the encrypted
* signature still travels back for multi-turn continuity. Use for faster
* time-to-first-text-token when your UI does not surface thinking.
@@ -211,9 +218,21 @@ export interface AnthropicOptions extends StreamOptions {
* Note: Anthropic's API default for Claude Opus 4.7 and Claude Mythos Preview
* is "omitted". We default to "summarized" here to keep behavior consistent
* with older Claude 4 models. Set this explicitly to "omitted" to opt in.
* Default: "summarized" when thinking is enabled.
*/
thinkingDisplay?: AnthropicThinkingDisplay;
/**
* Whether to request the interleaved thinking beta header for non-adaptive
* thinking models. Adaptive thinking models have interleaved thinking built in,
* so the header is skipped for them regardless of this setting.
* Default: true.
*/
interleavedThinking?: boolean;
/**
* Anthropic tool choice behavior. String values map to Anthropic's built-in
* choices; `{ type: "tool", name }` forces a specific tool.
* Default: omitted (Anthropic default behavior, currently equivalent to auto).
*/
toolChoice?: "auto" | "any" | "none" | { type: "tool"; name: string };
/**
* Pre-built Anthropic client instance. When provided, skips internal client
@@ -686,21 +705,6 @@ export const streamAnthropic: StreamFunction<"anthropic-messages", AnthropicOpti
return stream;
};
/**
* Check if a model supports adaptive thinking (Opus 4.6+, Sonnet 4.6)
*/
function supportsAdaptiveThinking(modelId: string): boolean {
// Adaptive-thinking model IDs (with or without date suffix)
return (
modelId.includes("opus-4-6") ||
modelId.includes("opus-4.6") ||
modelId.includes("opus-4-7") ||
modelId.includes("opus-4.7") ||
modelId.includes("sonnet-4-6") ||
modelId.includes("sonnet-4.6")
);
}
/**
* Map ThinkingLevel to Anthropic effort levels for adaptive thinking.
* Note: effort "max" is only valid on Opus 4.6, while Opus 4.7 supports "xhigh".
@@ -740,9 +744,9 @@ export const streamSimpleAnthropic: StreamFunction<"anthropic-messages", SimpleS
return streamAnthropic(model, context, { ...base, thinkingEnabled: false } satisfies AnthropicOptions);
}
// For Opus 4.6 and Sonnet 4.6: use adaptive thinking with effort level
// For older models: use budget-based thinking
if (supportsAdaptiveThinking(model.id)) {
// For models with adaptive thinking: use an effort level.
// For older models: use budget-based thinking.
if (model.compat?.forceAdaptiveThinking === true) {
const effort = mapThinkingLevelToEffort(model, options.reasoning);
return streamAnthropic(model, context, {
...base,
@@ -781,9 +785,8 @@ function createClient(
dynamicHeaders?: Record<string, string>,
sessionId?: string,
): { client: Anthropic; isOAuthToken: boolean } {
// Adaptive thinking models (Opus 4.6, Sonnet 4.6) have interleaved thinking built-in.
// The beta header is deprecated on Opus 4.6 and redundant on Sonnet 4.6, so skip it.
const needsInterleavedBeta = interleavedThinking && !supportsAdaptiveThinking(model.id);
// Adaptive thinking models have interleaved thinking built in, so skip the beta header.
const needsInterleavedBeta = interleavedThinking && model.compat?.forceAdaptiveThinking !== true;
const betaFeatures: string[] = [];
if (useFineGrainedToolStreamingBeta) {
betaFeatures.push(FINE_GRAINED_TOOL_STREAMING_BETA);
@@ -939,14 +942,13 @@ function buildParams(
);
}
// Configure thinking mode: adaptive (Opus 4.6+ and Sonnet 4.6),
// budget-based (older models), or explicitly disabled.
// Configure thinking mode: adaptive, budget-based, or explicitly disabled.
if (model.reasoning) {
if (options?.thinkingEnabled) {
// Default to "summarized" so Opus 4.7 and Mythos Preview behave like
// older Claude 4 models (whose API default is also "summarized").
const display: AnthropicThinkingDisplay = options.thinkingDisplay ?? "summarized";
if (supportsAdaptiveThinking(model.id)) {
if (model.compat?.forceAdaptiveThinking === true) {
// Adaptive thinking: Claude decides when and how much to think.
params.thinking = { type: "adaptive", display };
if (options.effort) {

10
packages/ai/src/types.ts
View File

@@ -435,6 +435,16 @@ export interface AnthropicMessagesCompat {
* Default: true.
*/
supportsCacheControlOnTools?: boolean;
/**
* Whether to force adaptive thinking (`thinking.type: "adaptive"` plus
* `output_config.effort`) regardless of the model id. Built-in models that
* require adaptive thinking set this in generated metadata. Custom
* Anthropic-compatible providers can set this to `true` for any model whose
* upstream requires the adaptive format. Set to `false` to
* opt out on overridden built-in models.
* Default: false.
*/
forceAdaptiveThinking?: boolean;
}
/**