@bastani/pi-ai 0.9.19 → 0.9.20-alpha.10

package/dist/api/anthropic-messages.js

@@ -1,6 +1,5 @@
 import Anthropic from "@anthropic-ai/sdk";
 import { calculateCost } from "../models.js";
-import { splitDeferredTools } from "../utils/deferred-tools.js";
 import { appendAssistantMessageDiagnostic } from "../utils/diagnostics.js";
 import { assertSupportedDocumentMimeType } from "../utils/document-input.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
@@ -11,7 +10,9 @@ import { getProviderEnvValue } from "../utils/provider-env.js";
 import { retryProviderRequest } from "../utils/provider-retry.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
 import { createStreamDeadline, withStreamDeadline } from "../utils/stream-deadline.js";
-import { getJsonSchemaToolParameters, resolveJsonSchemaStrictSampling } from "./constrained-sampling.js";
+import { getSystemMessageText, renderSystemMessageUpdate } from "../utils/text.js";
+import { getCurrentTools, getDeclaredTools, getInitialSystemMessage, hasToolRedefinitions, resolveTranscript, } from "../utils/transcript.js";
+import { ANTHROPIC_STRICT_JSON_SCHEMA_PROFILE, getJsonSchemaToolParameters, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
 import { buildCopilotDynamicHeaders, hasCopilotVisionInput, preserveCopilotIntegrationHeader, } from "./github-copilot-headers.js";
 import { adjustMaxTokensForThinking, buildBaseOptions, clampMaxTokensToContext } from "./simple-options.js";
 import { transformMessages } from "./transform-messages.js";
@@ -42,13 +43,13 @@ function getCacheControl(model, cacheRetention, env) {
 // Stealth mode: Mimic Claude Code's tool naming exactly
 //
 // Anthropic gates newer models on the `claude-cli/<version>` user agent alone and rejects an
-// older one with `claude_code_version_too_old`. `claude-fable-5-1` requires >= 2.1.251, bisected
-// against the live API: 2.1.250 -> 400, 2.1.251 -> 200. This is pinned to that exact published
-// minimum rather than the newest release, and it is a strict superset of the previous 2.1.75 --
-// every model this provider ships answers 200 at 2.1.251. Raise it only when a model rejects
-// this value; a caller can override it for one client through the `headers` option, whose
-// lowercase `user-agent` key is merged last.
-const claudeCodeVersion = "2.1.251";
+// older one with `claude_code_version_too_old`. `claude-fable-5-1` requires >= 2.1.251 (bisected
+// against the live API: 2.1.250 -> 400, 2.1.251 -> 200) and Claude Opus 5.5 requires >= 2.1.280,
+// the floor the API names in its rejection. This is pinned to the highest published minimum
+// among shipped models rather than the newest release; every model this provider ships answers
+// 200 at 2.1.280. Raise it only when a model rejects this value; a caller can override it for
+// one client through the `headers` option, whose lowercase `user-agent` key is merged last.
+const claudeCodeVersion = "2.1.280";
 // Claude Code 2.x tool names (canonical casing)
 // Source: https://cchistory.mariozechner.at/data/prompts-2.1.11.md
 // To update: https://github.com/badlogic/cchistory
@@ -127,6 +128,75 @@ const INTERLEAVED_THINKING_BETA = "interleaved-thinking-2025-05-14";
 const SERVER_SIDE_FALLBACK_BETA = "server-side-fallback-2026-07-01";
 const THINKING_BINDING_CONTROLS_BETA = "thinking-binding-controls-2026-08-01";
 const MID_CONVERSATION_OUTPUT_CONFIG_BETA = "mid-conversation-output-config-2026-07-01";
+const MID_CONVERSATION_TOOL_CHANGES_BETA = "mid-conversation-tool-changes-2026-07-01";
+const FAST_MODE_BETA = "fast-mode-2026-02-01";
+/**
+ * Fast mode bills every token category at twice the standard rate for each supported model, across
+ * the full context window, and prompt-caching multipliers stack on top.
+ * https://platform.claude.com/docs/en/build-with-claude/fast-mode#pricing
+ */
+const FAST_MODE_PRICE_MULTIPLIER = 2;
+function requestedSpeed(model) {
+    return model.fastRoute?.speed;
+}
+function upstreamModelId(model) {
+    return model.fastRoute?.upstreamModelId ?? model.id;
+}
+function reportedSpeed(usage) {
+    const speed = usage?.speed;
+    return speed === "fast" || speed === "standard" ? speed : undefined;
+}
+function applyFastModePricing(usage) {
+    usage.cost.input *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.output *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.cacheRead *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.cacheWrite *= FAST_MODE_PRICE_MULTIPLIER;
+    usage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;
+}
+/**
+ * Reject a payload hook that rewrote a field the model's fast route owns.
+ *
+ * A fast variant is recorded, persisted, and billed under its own `-fast` identity, so a hook that
+ * changes the upstream model or drops `speed: "fast"` would send a different request under that
+ * identity. Refusing loudly surfaces the misconfigured hook instead of silently restoring the route.
+ */
+function assertPayloadPreservesFastRoute(model, payload) {
+    const fastRoute = model.fastRoute;
+    if (fastRoute?.speed === undefined)
+        return;
+    const remedy = `A fast model variant must send its route's upstream model and speed, because it is recorded, persisted, ` +
+        `and billed under its own identity. Select the normal model "${model.provider}/${fastRoute.baseModelId}" ` +
+        `if the request needs different routing`;
+    if (typeof payload !== "object" || payload === null || Array.isArray(payload)) {
+        throw new Error(`A request hook changed payload for fast model "${model.provider}/${model.id}" to a value that cannot carry ` +
+            `the route's upstream model and speed. ${remedy}, or return an object that preserves those fields.`);
+    }
+    const record = payload;
+    const conflicts = [];
+    if (record.model !== fastRoute.upstreamModelId) {
+        conflicts.push(`model (expected "${fastRoute.upstreamModelId}", got ${JSON.stringify(record.model)})`);
+    }
+    if (record.speed !== fastRoute.speed) {
+        conflicts.push(`speed (expected ${JSON.stringify(fastRoute.speed)}, got ${JSON.stringify(record.speed)})`);
+    }
+    if (conflicts.length === 0)
+        return;
+    throw new Error(`A request hook changed ${conflicts.join(" and ")} for fast model "${model.provider}/${model.id}". ` +
+        `${remedy}, or stop rewriting those fields.`);
+}
+/**
+ * Stable deferred tool declared whenever native tool changes are in use. Anthropic adds
+ * hidden prompt scaffolding as soon as any tool has `defer_loading`; declaring this
+ * placeholder from the first request keeps that scaffolding in the cached prefix, so the
+ * first real late tool does not invalidate the cache (measured: full miss without it).
+ * It is never activated and the model cannot see it.
+ */
+const DEFERRED_TOOL_PLACEHOLDER = {
+    name: "__pi_deferred_placeholder__",
+    description: "Reserved placeholder. Never available. Never call this.",
+    input_schema: { type: "object", properties: {}, required: [] },
+    defer_loading: true,
+};
 function shouldUseServerSideFallbackBeta(model) {
     return (model.compat?.allowedFallbackModels?.length ?? 0) > 0;
 }
@@ -205,7 +275,7 @@ function addEarlierAttemptCosts(output, model, servingModel, event) {
         if (!iteration.output_tokens)
             continue;
         // Price at the rates of the model that ran the attempt, which is not the serving model.
-        const attemptCost = iteration.model === model.id
+        const attemptCost = iteration.model === upstreamModelId(model)
             ? model.cost
             : iteration.model === servingModel.id
                 ? servingModel.cost
@@ -244,24 +314,10 @@ function getAnthropicCompat(model) {
         supportsTemperature: model.compat?.supportsTemperature ?? true,
         allowEmptySignature: model.compat?.allowEmptySignature ?? false,
         supportsStrictTools: model.compat?.supportsStrictTools ?? false,
-        supportsToolReferences: model.compat?.supportsToolReferences ?? defaultSupportsToolReferences(model),
+        supportsMidConvoSystemMessages: model.compat?.supportsMidConvoSystemMessages ?? false,
+        supportsMidConvoToolChanges: model.compat?.supportsMidConvoToolChanges ?? false,
     };
 }
-/**
- * Default for `supportsToolReferences`: first-party Anthropic models except
- * Haiku (rejects client-side tool_reference blocks) and models that predate
- * tool search (Claude 3.x, Opus/Sonnet 4.0, Opus 4.1).
- */
-function defaultSupportsToolReferences(model) {
-    if (model.provider !== "anthropic" || model.id.includes("haiku"))
-        return false;
-    const version = model.id.match(/^claude-(?:opus|sonnet|fable)-(\d+)(?:-(\d+))?(?:-|$)/);
-    if (!version)
-        return false;
-    const major = Number(version[1]);
-    const minor = version[2] && version[2].length < 8 ? Number(version[2]) : 0;
-    return major > 4 || (major === 4 && minor >= 5);
-}
 function mergeHeaders(...headerSources) {
     const merged = {};
     for (const headers of headerSources) {
@@ -456,6 +512,8 @@ async function* iterateAnthropicEvents(response, signal) {
 }
 export const stream = (model, context, options) => {
     const stream = new AssistantMessageEventStream();
+    const normalizedContext = resolveTranscript(context, getAnthropicCompat(model).supportsMidConvoSystemMessages);
+    const currentTools = getCurrentTools(normalizedContext.messages);
     (async () => {
         const providerThinkingLevel = supportsMidConvoEffort(model) ? (options?.effort ?? "high") : undefined;
         const output = {
@@ -482,6 +540,7 @@ export const stream = (model, context, options) => {
             let client;
             let isOAuth;
             let usageModel = model;
+            let servedSpeed = requestedSpeed(model);
             if (options?.client) {
                 client = options.client;
                 isOAuth = false;
@@ -491,30 +550,34 @@ export const stream = (model, context, options) => {
                 assertRequestAuth(model.provider, apiKey, options?.headers);
                 let copilotDynamicHeaders;
                 if (model.provider === "github-copilot") {
-                    const hasImages = hasCopilotVisionInput(context.messages);
+                    const hasImages = hasCopilotVisionInput(normalizedContext.messages);
                     copilotDynamicHeaders = preserveCopilotIntegrationHeader(model.headers, buildCopilotDynamicHeaders({
-                        messages: context.messages,
+                        messages: normalizedContext.messages,
                         hasImages,
                         apiKey,
                     }));
                 }
                 const cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);
                 const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;
-                const created = createClient(model, apiKey, model.reasoning && options?.thinkingEnabled === true && (options.interleavedThinking ?? true), shouldUseFineGrainedToolStreamingBeta(model, context), shouldUseServerSideFallbackBeta(model), options?.headers, options?.fetch, copilotDynamicHeaders, cacheSessionId);
+                const created = createClient(model, apiKey, model.reasoning && options?.thinkingEnabled === true && (options.interleavedThinking ?? true), shouldUseFineGrainedToolStreamingBeta(model, context), shouldUseServerSideFallbackBeta(model), options?.headers, options?.fetch, copilotDynamicHeaders, cacheSessionId, getAnthropicCompat(model).supportsMidConvoToolChanges &&
+                    getAnthropicCompat(model).supportsMidConvoSystemMessages &&
+                    (getInitialSystemMessage(normalizedContext.messages)?.toolsAdded?.length ?? 0) > 0 &&
+                    !hasToolRedefinitions(normalizedContext.messages));
                 client = created.client;
                 isOAuth = created.isOAuthToken;
             }
-            let params = buildParams(model, context, isOAuth, options);
+            let params = buildParams(model, normalizedContext, isOAuth, options);
             const nextParams = await options?.onPayload?.(params, model);
             if (nextParams !== undefined) {
                 params = nextParams;
             }
+            assertPayloadPreservesFastRoute(model, params);
             const requestOptions = {
                 ...(streamDeadline.signal ? { signal: streamDeadline.signal } : {}),
                 ...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),
                 maxRetries: 0,
             };
-            const response = await retryProviderRequest(() => client.messages.create({ ...params, stream: true }, requestOptions).asResponse(), {
+            const response = await retryProviderRequest(() => client.beta.messages.create({ ...params, stream: true }, requestOptions).asResponse(), {
                 maxRetries: options?.maxRetries,
                 maxRetryDelayMs: options?.maxRetryDelayMs,
                 signal: streamDeadline.signal,
@@ -523,13 +586,16 @@ export const stream = (model, context, options) => {
             stream.push({ type: "start", partial: output });
             const blocks = output.content;
             for await (const event of withStreamDeadline(iterateAnthropicEvents(response, streamDeadline.signal), streamDeadline.deadlineMs, streamDeadline.abort)) {
+                await options?.onProviderStreamEvent?.(event, model);
                 if (event.type === "message_start") {
                     output.responseId = event.message.id;
-                    output.model = event.message.model;
-                    const fallbackCost = output.model === model.id
+                    const responseModel = event.message.model;
+                    if (responseModel !== upstreamModelId(model))
+                        output.responseModel = responseModel;
+                    const fallbackCost = responseModel === upstreamModelId(model)
                         ? undefined
-                        : model.compat?.allowedFallbackModels?.find((fallback) => fallback.provider === model.provider && fallback.model === output.model)?.cost;
-                    usageModel = fallbackCost ? { ...model, id: output.model, cost: fallbackCost } : model;
+                        : model.compat?.allowedFallbackModels?.find((fallback) => fallback.provider === model.provider && fallback.model === responseModel)?.cost;
+                    usageModel = fallbackCost ? { ...model, id: responseModel, cost: fallbackCost } : model;
                     // Capture initial token usage from message_start event
                     // This ensures we have input token counts even if the stream is aborted early
                     output.usage.input = event.message.usage.input_tokens || 0;
@@ -541,6 +607,9 @@ export const stream = (model, context, options) => {
                     output.usage.totalTokens =
                         output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
                     calculateCost(usageModel, output.usage);
+                    servedSpeed = reportedSpeed(event.message.usage) ?? servedSpeed;
+                    if (servedSpeed === "fast")
+                        applyFastModePricing(output.usage);
                     inputTransformations = getInputTransformations(event.message) ?? inputTransformations;
                 }
                 else if (event.type === "content_block_start") {
@@ -579,7 +648,7 @@ export const stream = (model, context, options) => {
                             type: "toolCall",
                             id: event.content_block.id,
                             name: isOAuth
-                                ? fromClaudeCodeName(event.content_block.name, context.tools)
+                                ? fromClaudeCodeName(event.content_block.name, currentTools)
                                 : event.content_block.name,
                             arguments: event.content_block.input ?? {},
                             partialJson: "",
@@ -728,6 +797,11 @@ export const stream = (model, context, options) => {
                         // its Usage type, so read it through a narrow cast. Verified against the live API.
                         const thinkingTokens = event.usage
                             .output_tokens_details?.thinking_tokens;
+                        // Vercel AI Gateway includes the TTL breakdown in deltas, though the SDK only types it on message_start.
+                        const cacheCreation = event.usage.cache_creation;
+                        if (cacheCreation?.ephemeral_1h_input_tokens != null) {
+                            output.usage.cacheWrite1h = cacheCreation.ephemeral_1h_input_tokens;
+                        }
                         if (thinkingTokens != null) {
                             output.usage.reasoning = thinkingTokens;
                         }
@@ -736,6 +810,9 @@ export const stream = (model, context, options) => {
                     output.usage.totalTokens =
                         output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
                     calculateCost(usageModel, output.usage);
+                    servedSpeed = reportedSpeed(event.usage) ?? servedSpeed;
+                    if (servedSpeed === "fast")
+                        applyFastModePricing(output.usage);
                     // A non-empty final fallback report supersedes the request-start report with
                     // the serving model's transformations. An empty report must not erase drops
                     // already reported at message_start.
@@ -839,7 +916,7 @@ export const streamSimple = (model, context, options) => {
 function isOAuthToken(apiKey) {
     return apiKey.includes("sk-ant-oat");
 }
-function createClient(model, apiKey, interleavedThinking, useFineGrainedToolStreamingBeta, useServerSideFallbackBeta, optionsHeaders, fetch, dynamicHeaders, sessionId) {
+function createClient(model, apiKey, interleavedThinking, useFineGrainedToolStreamingBeta, useServerSideFallbackBeta, optionsHeaders, fetch, dynamicHeaders, sessionId, useNativeToolChanges = false) {
     // Adaptive thinking models have interleaved thinking built in, so skip the beta header.
     const needsInterleavedBeta = interleavedThinking && model.compat?.forceAdaptiveThinking !== true;
     const betaFeatures = [];
@@ -858,6 +935,10 @@ function createClient(model, apiKey, interleavedThinking, useFineGrainedToolStre
     if (supportsMidConvoEffort(model)) {
         betaFeatures.push(MID_CONVERSATION_OUTPUT_CONFIG_BETA, THINKING_BINDING_CONTROLS_BETA);
     }
+    if (useNativeToolChanges)
+        betaFeatures.push(MID_CONVERSATION_TOOL_CHANGES_BETA);
+    if (requestedSpeed(model) === "fast")
+        betaFeatures.push(FAST_MODE_BETA);
     const uniqueBetaFeatures = [...new Set(betaFeatures)];
     // Copilot: Bearer auth, selective betas.
     if (model.provider === "github-copilot") {
@@ -915,29 +996,74 @@ function createClient(model, apiKey, interleavedThinking, useFineGrainedToolStre
     });
     return { client, isOAuthToken: false };
 }
+function getNativeToolChangeBetas(model, context, isOAuth, options) {
+    let configured;
+    for (const headers of [model.headers, options?.headers]) {
+        for (const [name, value] of Object.entries(headers ?? {})) {
+            if (name.toLowerCase() === "anthropic-beta")
+                configured = value;
+        }
+    }
+    if (configured === null)
+        return [];
+    if (configured !== undefined)
+        return [
+            ...new Set(configured
+                .split(",")
+                .map((feature) => feature.trim())
+                .filter(Boolean)),
+        ];
+    const features = [MID_CONVERSATION_TOOL_CHANGES_BETA];
+    if (isOAuth)
+        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 &&
+        (options.interleavedThinking ?? true) &&
+        model.compat?.forceAdaptiveThinking !== true) {
+        features.push(INTERLEAVED_THINKING_BETA);
+    }
+    if (shouldUseServerSideFallbackBeta(model))
+        features.push(SERVER_SIDE_FALLBACK_BETA);
+    if (shouldUseThinkingBindingControlsBeta(model))
+        features.push(THINKING_BINDING_CONTROLS_BETA);
+    if (supportsMidConvoEffort(model))
+        features.push(MID_CONVERSATION_OUTPUT_CONFIG_BETA, THINKING_BINDING_CONTROLS_BETA);
+    if (requestedSpeed(model) === "fast")
+        features.push(FAST_MODE_BETA);
+    return [...new Set(features)];
+}
 function buildParams(model, context, isOAuthToken, options) {
     const { cacheControl } = getCacheControl(model, options?.cacheRetention, options?.env);
     const compat = getAnthropicCompat(model);
+    const initialSystemMessage = getInitialSystemMessage(context.messages);
+    const initialSystemText = initialSystemMessage ? getSystemMessageText(initialSystemMessage) : "";
     const transformedMessages = transformMessages(context.messages, model, normalizeToolCallId);
-    const normalizeToolName = isOAuthToken ? toClaudeCodeName : (name) => name;
-    const toolPlacement = splitDeferredTools({ ...context, messages: transformedMessages }, compat.supportsToolReferences, normalizeToolName);
-    let immediateTools = toolPlacement.immediate;
-    let deferredTools = [...toolPlacement.deferred.values()];
-    if (immediateTools.length === 0 && deferredTools.length > 0) {
-        immediateTools = deferredTools;
-        deferredTools = [];
-    }
-    const deferredToolNames = new Set(deferredTools.map((tool) => normalizeToolName(tool.name)));
-    const converted = convertMessages(transformedMessages, isOAuthToken, cacheControl, compat.allowEmptySignature, deferredToolNames, normalizeToolName, supportsMidConvoEffort(model) ? model.provider : undefined);
+    const conversationMessages = initialSystemMessage ? transformedMessages.slice(1) : transformedMessages;
+    // Native tool changes reference tools by name, so a redefined name cannot be expressed,
+    // and Anthropic rejects a tool list where every tool is deferred, so there must be an
+    // initial active tool to anchor the deferred ones. Otherwise the current tool list is sent.
+    const initialTools = initialSystemMessage?.toolsAdded ?? [];
+    const nativeToolChanges = compat.supportsMidConvoSystemMessages &&
+        compat.supportsMidConvoToolChanges &&
+        initialTools.length > 0 &&
+        !hasToolRedefinitions(context.messages);
+    const converted = convertMessages(conversationMessages, isOAuthToken, cacheControl, compat.allowEmptySignature, supportsMidConvoEffort(model) ? model.provider : undefined, nativeToolChanges);
     const activeEffort = options?.effort ?? "high";
     const params = {
-        model: model.id,
+        model: upstreamModelId(model),
         messages: (supportsMidConvoEffort(model)
             ? insertThinkingLevelMessages(converted, activeEffort)
             : converted.messages),
         max_tokens: options?.maxTokens ?? model.maxTokens,
         stream: true,
     };
+    if (nativeToolChanges)
+        params.betas = getNativeToolChangeBetas(model, context, isOAuthToken, options);
+    const speed = requestedSpeed(model);
+    if (speed !== undefined)
+        params.speed = speed;
     // For OAuth tokens, we MUST include Claude Code identity
     if (isOAuthToken) {
         params.system = [
@@ -947,20 +1073,20 @@ function buildParams(model, context, isOAuthToken, options) {
                 ...(cacheControl ? { cache_control: cacheControl } : {}),
             },
         ];
-        if (context.systemPrompt) {
+        if (initialSystemText) {
             params.system.push({
                 type: "text",
-                text: sanitizeSurrogates(context.systemPrompt),
+                text: sanitizeSurrogates(initialSystemText),
                 ...(cacheControl ? { cache_control: cacheControl } : {}),
             });
         }
     }
-    else if (context.systemPrompt) {
+    else if (initialSystemText) {
         // Add cache control to system prompt for non-OAuth tokens
         params.system = [
             {
                 type: "text",
-                text: sanitizeSurrogates(context.systemPrompt),
+                text: sanitizeSurrogates(initialSystemText),
                 ...(cacheControl ? { cache_control: cacheControl } : {}),
             },
         ];
@@ -972,12 +1098,26 @@ function buildParams(model, context, isOAuthToken, options) {
         compat.supportsTemperature) {
         params.temperature = options.temperature;
     }
-    if (immediateTools.length > 0 || deferredTools.length > 0) {
+    const toolCacheControl = compat.supportsCacheControlOnTools ? cacheControl : undefined;
+    if (nativeToolChanges) {
+        // Initial tools stay active with the cache breakpoint on the last one. Every later
+        // declaration is deferred and only surfaced by its `tool_addition` block; removed
+        // tools stay declared and are withdrawn by `tool_removal`. The request-level list
+        // therefore only grows, keeping the cached prefix intact across tool changes.
+        const initialNames = new Set(initialTools.map((tool) => tool.name));
+        const laterTools = getDeclaredTools(context.messages).filter((tool) => !initialNames.has(tool.name));
         params.tools = [
-            ...convertTools(immediateTools, isOAuthToken, compat.supportsEagerToolInputStreaming, compat.supportsStrictTools, compat.supportsCacheControlOnTools ? cacheControl : undefined),
-            ...convertTools(deferredTools, isOAuthToken, compat.supportsEagerToolInputStreaming, compat.supportsStrictTools, undefined, true),
+            ...convertTools(initialTools, isOAuthToken, compat.supportsEagerToolInputStreaming, compat.supportsStrictTools, toolCacheControl),
+            DEFERRED_TOOL_PLACEHOLDER,
+            ...convertTools(laterTools, isOAuthToken, compat.supportsEagerToolInputStreaming, compat.supportsStrictTools).map((tool) => ({ ...tool, defer_loading: true })),
         ];
     }
+    else {
+        const tools = getCurrentTools(context.messages);
+        if (tools.length > 0) {
+            params.tools = convertTools(tools, isOAuthToken, compat.supportsEagerToolInputStreaming, compat.supportsStrictTools, toolCacheControl);
+        }
+    }
     // Managed effort models always use adaptive thinking. Their per-turn markers
     // carry the actual effort; the top-level high value deliberately stays stable.
     if (supportsMidConvoEffort(model)) {
@@ -1082,41 +1222,54 @@ function buildParams(model, context, isOAuthToken, options) {
 function normalizeToolCallId(id) {
     return id.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64);
 }
-function convertToolResult(msg, isOAuthToken, deferredToolNames, loadedToolNames, normalizeToolName) {
-    const references = [];
-    for (const name of msg.addedToolNames ?? []) {
-        const normalizedName = normalizeToolName(name);
-        if (!deferredToolNames.has(normalizedName) || loadedToolNames.has(normalizedName))
-            continue;
-        loadedToolNames.add(normalizedName);
-        references.push({
-            type: "tool_reference",
-            tool_name: isOAuthToken ? toClaudeCodeName(name) : name,
-        });
-    }
-    const convertedContent = convertContentBlocks(msg.content);
-    // Anthropic rejects tool references mixed with ordinary tool-result content.
+function convertToolResult(msg) {
     return {
-        toolResult: {
-            type: "tool_result",
-            tool_use_id: msg.toolCallId,
-            content: references.length > 0 ? references : convertedContent,
-            is_error: msg.isError,
-        },
-        siblingContent: references.length === 0
-            ? []
-            : typeof convertedContent === "string"
-                ? [{ type: "text", text: convertedContent }]
-                : convertedContent,
+        type: "tool_result",
+        tool_use_id: msg.toolCallId,
+        content: convertContentBlocks(msg.content),
+        is_error: msg.isError,
     };
 }
-function convertMessages(transformedMessages, isOAuthToken, cacheControl, allowEmptySignature = false, deferredToolNames = new Set(), normalizeToolName = (name) => name, managedProvider) {
+function convertMessages(transformedMessages, isOAuthToken, cacheControl, allowEmptySignature = false, managedProvider, nativeToolChanges = false) {
     const params = [];
     const assistantLevels = new Map();
-    const loadedToolNames = new Set();
+    // Later system messages are held back and emitted directly before the next assistant
+    // message (or at the end of the transcript). Anthropic requires `tool_result` blocks to
+    // immediately follow their `tool_use`, so a system message between them is rejected; this
+    // also mirrors where the managed-effort system messages are inserted. As a result an
+    // update placed before a user message in the transcript lands after it on the wire.
+    const pendingSystemMessages = [];
+    const flushPendingSystemMessages = () => {
+        params.push(...pendingSystemMessages);
+        pendingSystemMessages.length = 0;
+    };
     for (let i = 0; i < transformedMessages.length; i++) {
         const msg = transformedMessages[i];
-        if (msg.role === "user") {
+        if (msg.role === "system") {
+            // Later system messages only reach this point when the model accepts them natively;
+            // otherwise the transcript was collapsed into the leading message before conversion.
+            const text = renderSystemMessageUpdate(msg);
+            const blocks = [];
+            if (text.length > 0)
+                blocks.push({ type: "text", text: sanitizeSurrogates(text) });
+            if (nativeToolChanges) {
+                for (const tool of msg.toolsRemoved ?? []) {
+                    blocks.push({
+                        type: "tool_removal",
+                        tool: { type: "tool_reference", name: isOAuthToken ? toClaudeCodeName(tool.name) : tool.name },
+                    });
+                }
+                for (const tool of msg.toolsAdded ?? []) {
+                    blocks.push({
+                        type: "tool_addition",
+                        tool: { type: "tool_reference", name: isOAuthToken ? toClaudeCodeName(tool.name) : tool.name },
+                    });
+                }
+            }
+            if (blocks.length > 0)
+                pendingSystemMessages.push({ role: "system", content: blocks });
+        }
+        else if (msg.role === "user") {
             if (typeof msg.content === "string") {
                 if (msg.content.trim().length > 0) {
                     params.push({
@@ -1173,6 +1326,7 @@ function convertMessages(transformedMessages, isOAuthToken, cacheControl, allowE
             }
         }
         else if (msg.role === "assistant") {
+            flushPendingSystemMessages();
             const blocks = [];
             for (const block of msg.content) {
                 if (block.type === "text") {
@@ -1261,31 +1415,32 @@ function convertMessages(transformedMessages, isOAuthToken, cacheControl, allowE
         else if (msg.role === "toolResult") {
             // Collect all consecutive toolResult messages, needed for z.ai Anthropic endpoint.
             const toolResults = [];
-            const siblingContent = [];
             let j = i;
             while (j < transformedMessages.length && transformedMessages[j].role === "toolResult") {
-                const converted = convertToolResult(transformedMessages[j], isOAuthToken, deferredToolNames, loadedToolNames, normalizeToolName);
-                toolResults.push(converted.toolResult);
-                siblingContent.push(...converted.siblingContent);
+                toolResults.push(convertToolResult(transformedMessages[j]));
                 j++;
             }
             // Skip the messages we've already processed.
             i = j - 1;
-            // Displaced reference-bearing results must follow every tool_result block.
             params.push({
                 role: "user",
-                content: [...toolResults, ...siblingContent],
+                content: toolResults,
             });
         }
     }
-    // Add cache_control to the last user message to cache conversation history
+    flushPendingSystemMessages();
+    // Add cache_control to the last user or system message to cache conversation history
     if (cacheControl && params.length > 0) {
         const lastMessage = params[params.length - 1];
-        if (lastMessage.role === "user") {
+        if (lastMessage.role === "user" || lastMessage.role === "system") {
             if (Array.isArray(lastMessage.content)) {
                 const lastBlock = lastMessage.content[lastMessage.content.length - 1];
                 if (lastBlock &&
-                    (lastBlock.type === "text" || lastBlock.type === "image" || lastBlock.type === "tool_result")) {
+                    (lastBlock.type === "text" ||
+                        lastBlock.type === "image" ||
+                        lastBlock.type === "tool_result" ||
+                        lastBlock.type === "tool_addition" ||
+                        lastBlock.type === "tool_removal")) {
                     lastBlock.cache_control = cacheControl;
                 }
             }
@@ -1318,7 +1473,7 @@ function insertThinkingLevelMessages(converted, activeEffort) {
     return messages;
 }
 function shouldUseFineGrainedToolStreamingBeta(model, context) {
-    return !!context.tools?.length && !getAnthropicCompat(model).supportsEagerToolInputStreaming;
+    return getCurrentTools(context.messages).length > 0 && !getAnthropicCompat(model).supportsEagerToolInputStreaming;
 }
 function projectObjectUnionForAnthropic(schema) {
     if (schema.properties !== undefined || !Array.isArray(schema.anyOf) || schema.anyOf.length === 0)
@@ -1354,12 +1509,16 @@ function projectObjectUnionForAnthropic(schema) {
         required: (branches[0]?.required ?? []).filter((key) => branches.every((branch) => (branch.required ?? []).includes(key))),
     };
 }
-function convertTools(tools, isOAuthToken, supportsEagerToolInputStreaming, supportsStrictTools, cacheControl, deferLoading = false) {
+function convertTools(tools, isOAuthToken, supportsEagerToolInputStreaming, supportsStrictTools, cacheControl) {
     if (!tools)
         return [];
+    // OAuth canonicalization can make distinct transcript names collide. Keep the latest
+    // declaration, as the pre-transcript tool serializer did, without sending duplicate names.
+    if (isOAuthToken)
+        tools = [...new Map(tools.map((tool) => [toClaudeCodeName(tool.name), tool])).values()];
     return tools.map((tool, index) => {
-        const strict = resolveJsonSchemaStrictSampling(tool, supportsStrictTools);
-        const parameters = getJsonSchemaToolParameters(tool, strict);
+        const strict = resolveJsonSchemaStrictSampling(tool, supportsStrictTools, ANTHROPIC_STRICT_JSON_SCHEMA_PROFILE);
+        const parameters = getJsonSchemaToolParameters(tool, strict, ANTHROPIC_STRICT_JSON_SCHEMA_PROFILE);
         const schema = parameters;
         // Anthropic rejects top-level combinators, so project object-union fields for advertising only.
         // Local runtime validation still enforces the complete authored union without mutating it.
@@ -1381,7 +1540,6 @@ function convertTools(tools, isOAuthToken, supportsEagerToolInputStreaming, supp
             ...(supportsEagerToolInputStreaming ? { eager_input_streaming: true } : {}),
             ...(strict === true ? { strict: true } : {}),
             input_schema: inputSchema,
-            ...(deferLoading ? { defer_loading: true } : {}),
             ...(cacheControl && index === tools.length - 1 ? { cache_control: cacheControl } : {}),
         };
     });