@code-yeongyu/senpi 2026.9.4 → 2026.9.5-2

package/node_modules/@earendil-works/pi-ai/dist/api/anthropic-messages.js

@@ -179,6 +179,11 @@ const UNSUPPORTED_NATIVE_COMPUTER_TOOL_MODEL_MARKERS = [
     "opus-4.8",
 ];
 const SERVER_SIDE_FALLBACK_BETA = "server-side-fallback-2026-07-01";
+function shouldUseServerSideFallbackBeta(model) {
+    return (model.compat?.allowedFallbackModels?.length ?? 0) > 0;
+}
+const MID_CONVERSATION_OUTPUT_CONFIG_BETA = "mid-conversation-output-config-2026-07-01";
+const THINKING_BINDING_CONTROLS_BETA = "thinking-binding-controls-2026-08-01";
 // A provider can reject its own empty signatures. Learn that capability for one
 // conversation without changing the shared model definition used by other sessions.
 const unsignedThinkingTextReplayFallbacks = new Set();
@@ -953,12 +958,14 @@ async function* iterateAnthropicEvents(response, signal) {
 export const stream = (model, context, options) => {
     const stream = new AssistantMessageEventStream();
     (async () => {
+        const providerThinkingLevel = model.compat?.supportsMidConvoEffort ? (options?.effort ?? "high") : undefined;
         const output = {
             role: "assistant",
             content: [],
             api: model.api,
             provider: model.provider,
             model: model.id,
+            ...(providerThinkingLevel === undefined ? {} : { providerThinkingLevel }),
             usage: {
                 input: 0,
                 output: 0,
@@ -978,6 +985,7 @@ export const stream = (model, context, options) => {
             let client;
             let isOAuth;
             let usageModel = model;
+            let inputTransformations;
             if (options?.client) {
                 client = options.client;
                 isOAuth = false;
@@ -1022,13 +1030,15 @@ export const stream = (model, context, options) => {
                     ...(payloadRequestMetadata.headers ? { headers: payloadRequestMetadata.headers } : {}),
                 };
                 try {
-                    const response = await client.messages.create({ ...params, stream: true }, requestOptions).asResponse();
+                    const response = await client.beta.messages
+                        .create({ ...params, stream: true }, requestOptions)
+                        .asResponse();
                     return { params, response };
                 }
                 catch (error) {
                     if (isForcedToolChoiceUnsupportedError(error, isForcedAnthropicToolChoice(params.tool_choice))) {
                         params = omitToolChoiceParam(params);
-                        const response = await client.messages
+                        const response = await client.beta.messages
                             .create({ ...params, stream: true }, requestOptions)
                             .asResponse();
                         return { params, response };
@@ -1060,6 +1070,9 @@ export const stream = (model, context, options) => {
             for await (const event of iterateAnthropicEvents(response, requestSignal)) {
                 if (event.type === "message_start") {
                     output.responseId = event.message.id;
+                    const transformations = event.message.input_transformations;
+                    if (Array.isArray(transformations))
+                        inputTransformations = transformations;
                     output.model = event.message.model;
                     const fallback = output.model === model.id
                         ? undefined
@@ -1089,6 +1102,9 @@ export const stream = (model, context, options) => {
                     }
                 }
                 else if (event.type === "content_block_start") {
+                    if (event.content_block.type === "fallback" && output.content.length > 0) {
+                        throw new Error("Anthropic performed an unsupported mid-output model fallback");
+                    }
                     const receipt = options?.abortServerSideFallback === true
                         ? parseServerFallbackReceipt(event.content_block)
                         : undefined;
@@ -1251,6 +1267,9 @@ export const stream = (model, context, options) => {
                     }
                 }
                 else if (event.type === "message_delta") {
+                    const transformations = event.input_transformations;
+                    if (Array.isArray(transformations))
+                        inputTransformations = transformations;
                     if (event.delta.stop_reason) {
                         output.rawStopReason = event.delta.stop_reason;
                         const stopReasonResult = mapStopReason(event.delta.stop_reason, event.delta.stop_details);
@@ -1277,11 +1296,8 @@ export const stream = (model, context, options) => {
                         if (event.usage.cache_creation_input_tokens != null) {
                             output.usage.cacheWrite = event.usage.cache_creation_input_tokens;
                         }
-                        // Anthropic reports reasoning tokens in `output_tokens_details.thinking_tokens` on the
-                        // final message_delta usage (a subset of output_tokens). SDK 0.91.1 omits the field from
-                        // its Usage type, so read it through a narrow cast. Verified against the live API.
-                        const thinkingTokens = event.usage
-                            .output_tokens_details?.thinking_tokens;
+                        // Anthropic reports reasoning tokens as a subset of output tokens.
+                        const thinkingTokens = event.usage.output_tokens_details?.thinking_tokens;
                         if (thinkingTokens != null) {
                             output.usage.reasoning = thinkingTokens;
                         }
@@ -1304,6 +1320,19 @@ export const stream = (model, context, options) => {
             if (output.stopReason === "aborted" || output.stopReason === "error") {
                 throw new Error(output.errorMessage || "An unknown error occurred");
             }
+            if (inputTransformations && inputTransformations.length > 0) {
+                appendAssistantMessageDiagnostic(output, {
+                    type: "anthropic_input_transformations",
+                    timestamp: Date.now(),
+                    details: {
+                        transformations: inputTransformations.map((transformation) => ({
+                            type: transformation.type ?? undefined,
+                            path: transformation.path ?? undefined,
+                            reason: transformation.reason ?? undefined,
+                        })),
+                    },
+                });
+            }
             combinedAbort.cleanup();
             stream.push({ type: "done", reason: output.stopReason, message: output });
             stream.end();
@@ -1380,8 +1409,6 @@ function supportsNativeXhighEffort(model) {
 function cannotDisableThinking(model, compat) {
     if (!compat.supportsDisabledThinking)
         return true;
-    if (model.thinkingLevelMap?.off === null)
-        return true;
     return matchesModelMarker(model, DISABLED_THINKING_REJECTING_MODEL_MARKERS);
 }
 function disableThinkingForRequest(params, model, compat) {
@@ -1560,6 +1587,42 @@ export function buildAnthropicWarmPromptCacheParams(model, context, options) {
         max_tokens: 0,
     });
 }
+function getBetaFeatures(model, context, isOAuthToken, options) {
+    let configuredFeatures;
+    for (const headers of [model.headers, options?.headers]) {
+        for (const [name, value] of Object.entries(headers ?? {})) {
+            if (name.toLowerCase() === "anthropic-beta")
+                configuredFeatures = value;
+        }
+    }
+    if (configuredFeatures === null)
+        return [];
+    if (configuredFeatures !== undefined) {
+        return [
+            ...new Set(configuredFeatures
+                .split(",")
+                .map((feature) => feature.trim())
+                .filter(Boolean)),
+        ];
+    }
+    const features = [];
+    if (isOAuthToken)
+        features.push("claude-code-20250219", "oauth-2025-04-20");
+    if (shouldUseFineGrainedToolStreamingBeta(model, context))
+        features.push(FINE_GRAINED_TOOL_STREAMING_BETA);
+    if (model.reasoning &&
+        options?.thinkingEnabled === true &&
+        (options.interleavedThinking ?? true) &&
+        !supportsAdaptiveThinking(model)) {
+        features.push(INTERLEAVED_THINKING_BETA);
+    }
+    if (shouldUseServerSideFallbackBeta(model))
+        features.push(SERVER_SIDE_FALLBACK_BETA);
+    if (model.compat?.supportsMidConvoEffort === true) {
+        features.push(MID_CONVERSATION_OUTPUT_CONFIG_BETA, THINKING_BINDING_CONTROLS_BETA);
+    }
+    return [...new Set(features)];
+}
 function buildParams(model, context, isOAuthToken, options, unsignedThinkingReplay = getAnthropicCompat(model).unsignedThinkingReplay) {
     const compat = getAnthropicCompat(model);
     const { cacheControl } = getCacheControl(model, options?.cacheRetention ?? model.cacheRetention, options?.env);
@@ -1581,11 +1644,28 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
         deferredTools = [];
     }
     const deferredToolNames = new Set(deferredTools.map((tool) => normalizeToolName(tool.name)));
+    const activeEffort = options?.effort ?? "high";
+    const betaFeatures = getBetaFeatures(model, context, isOAuthToken, options);
+    const convertedMessages = convertMessages(transformedMessages, model, isOAuthToken, cacheControl, unsignedThinkingReplay, deferredToolNames, normalizeToolName, model.compat?.supportsMidConvoEffort === true ? model.provider : undefined);
+    const messages = model.compat?.supportsMidConvoEffort === true
+        ? insertThinkingLevelMessages(convertedMessages, activeEffort)
+        : convertedMessages.messages;
+    // Effort markers are appended after history; re-run the fork's final-user cache checkpoint
+    // against the actual last user/tool-result message so the marker cannot displace cache_control.
+    if (model.compat?.supportsMidConvoEffort === true && cacheControl) {
+        for (let index = messages.length - 1; index >= 0; index--) {
+            if (messages[index].role === "user") {
+                markUserMessageCacheCheckpoint(messages[index], cacheControl);
+                break;
+            }
+        }
+    }
     const params = {
         model: model.id,
-        messages: convertMessages(transformedMessages, model, isOAuthToken, cacheControl, unsignedThinkingReplay, deferredToolNames, normalizeToolName),
+        messages,
         max_tokens: options?.maxTokens ?? model.maxTokens,
         stream: true,
+        ...(betaFeatures.length > 0 ? { betas: betaFeatures } : {}),
     };
     // For OAuth tokens, we MUST include Claude Code identity
     if (isOAuthToken) {
@@ -1629,8 +1709,18 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
             ...convertTools(deferredTools, isOAuthToken, compat.supportsEagerToolInputStreaming, compat.supportsStrictTools, undefined, true),
         ];
     }
-    // Configure thinking mode: adaptive, budget-based, or explicitly disabled.
-    if (model.reasoning) {
+    // Managed effort models always use adaptive thinking so prefix mismatches can
+    // be dropped instead of surfacing as persistent 400 responses. Thinking-off is
+    // handled before this managed branch because these models accept `disabled`.
+    if (model.compat?.supportsMidConvoEffort === true && options?.thinkingEnabled !== false) {
+        params.thinking = {
+            type: "adaptive",
+            display: options?.thinkingDisplay ?? "summarized",
+            block_binding: { prefix_mismatch_behavior: "drop_block" },
+        };
+        params.output_config = { effort: "high" };
+    }
+    else 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").
@@ -1684,7 +1774,9 @@ function buildParams(model, context, isOAuthToken, options, unsignedThinkingRepl
     }
     applyExtraBodyToAnthropicParams(params, options?.extraBody);
     if (options?.refusalFallbacks !== undefined) {
-        params.fallbacks = options.refusalFallbacks;
+        // AnthropicRefusalFallback models a readonly list; the SDK's BetaFallbacksParam
+        // is mutable, so copy rather than widen the fork's public type.
+        params.fallbacks = options.refusalFallbacks === "default" ? "default" : [...options.refusalFallbacks];
     }
     return params;
 }
@@ -1746,8 +1838,9 @@ function convertToolResult(msg, isOAuthToken, deferredToolNames, loadedToolNames
                 : convertedContent,
     };
 }
-function convertMessages(transformedMessages, model, isOAuthToken, cacheControl, unsignedThinkingReplay = "text", deferredToolNames = new Set(), normalizeToolName = (name) => name) {
+function convertMessages(transformedMessages, model, isOAuthToken, cacheControl, unsignedThinkingReplay = "text", deferredToolNames = new Set(), normalizeToolName = (name) => name, managedProvider) {
     const params = [];
+    const assistantLevels = new Map();
     const loadedToolNames = new Set();
     // Tool calls from a declined pre-fallback attempt are dropped from their
     // assistant turn below; drop their tool_results in lockstep so none dangle.
@@ -1896,10 +1989,17 @@ function convertMessages(transformedMessages, model, isOAuthToken, cacheControl,
             }
             if (blocks.length === 0)
                 continue;
+            const messageIndex = params.length;
             params.push({
                 role: "assistant",
                 content: blocks,
             });
+            if (managedProvider !== undefined &&
+                msg.api === "anthropic-messages" &&
+                msg.provider === managedProvider &&
+                isAnthropicEffort(msg.providerThinkingLevel)) {
+                assistantLevels.set(messageIndex, msg.providerThinkingLevel);
+            }
         }
         else if (msg.role === "toolResult") {
             // Collect all consecutive toolResult messages, needed for z.ai Anthropic endpoint.
@@ -1933,7 +2033,22 @@ function convertMessages(transformedMessages, model, isOAuthToken, cacheControl,
             }
         }
     }
-    return params;
+    return { messages: params, assistantLevels };
+}
+function isAnthropicEffort(value) {
+    return value === "low" || value === "medium" || value === "high" || value === "xhigh" || value === "max";
+}
+function insertThinkingLevelMessages(converted, activeEffort) {
+    const messages = [];
+    for (let index = 0; index < converted.messages.length; index++) {
+        const historicalEffort = converted.assistantLevels.get(index);
+        if (historicalEffort !== undefined) {
+            messages.push({ role: "system", content: [], output_config: { effort: historicalEffort } });
+        }
+        messages.push(converted.messages[index]);
+    }
+    messages.push({ role: "system", content: [], output_config: { effort: activeEffort } });
+    return messages;
 }
 function shouldUseFineGrainedToolStreamingBeta(model, context) {
     return !!context.tools?.length && !getAnthropicCompat(model).supportsEagerToolInputStreaming;