@deepstrike/wasm 0.2.11 → 0.2.13

package/dist/providers/anthropic.js

@@ -1,12 +1,61 @@
 import { assistantReplayKey, collectStreamMessage, toAnthropicMessages } from "./base.js";
-function buildAnthropicTools(tools) {
+/** Anthropic accepts at most this many cache_control breakpoints per request. */
+const MAX_CACHE_BREAKPOINTS = 4;
+/** Rolling cache breakpoints reserved for the message history (system uses ≤2). */
+const MESSAGE_CACHE_BREAKPOINTS = 2;
+function buildAnthropicTools(tools, anchorCache) {
     return tools.map((t, i) => ({
         name: t.name,
         description: t.description,
         input_schema: JSON.parse(t.parameters),
-        ...(i === tools.length - 1 ? { cache_control: { type: "ephemeral" } } : {}),
+        // Anchor a tool breakpoint only when the system blocks won't carry one;
+        // otherwise systemStable already caches the tools prefix (tools render
+        // first), and a redundant tool breakpoint would burn a slot the message
+        // history needs to stay within the 4-breakpoint budget.
+        ...(anchorCache && i === tools.length - 1 ? { cache_control: { type: "ephemeral" } } : {}),
     }));
 }
+/**
+ * Roll cache breakpoints across the conversation tail so the message-history
+ * prefix is written once and re-read on later turns (without this the cached
+ * prefix stops at the end of `system` and the whole tool-result history is
+ * re-billed at full input price every turn). Marks the final message plus the
+ * nearest preceding user turn (read anchor); a bare string body is promoted to
+ * a cache-bearing text block.
+ */
+function applyMessageCacheControl(msgs) {
+    if (!msgs.length)
+        return;
+    const targets = new Set([msgs.length - 1]);
+    for (let i = msgs.length - 2; i >= 0 && targets.size < MESSAGE_CACHE_BREAKPOINTS; i--) {
+        if (msgs[i].role === "user")
+            targets.add(i);
+    }
+    for (const idx of targets)
+        markLastBlockCacheable(msgs[idx]);
+}
+function markLastBlockCacheable(msg) {
+    const cache_control = { type: "ephemeral" };
+    if (typeof msg.content === "string") {
+        if (!msg.content)
+            return;
+        msg.content = [{ type: "text", text: msg.content, cache_control }];
+        return;
+    }
+    if (Array.isArray(msg.content) && msg.content.length) {
+        const last = msg.content[msg.content.length - 1];
+        last.cache_control = cache_control;
+    }
+}
+/** Regression guard: fail loudly before the API would reject the request for
+ *  exceeding the cache_control breakpoint limit. */
+function assertCacheBudget(system, toolCount) {
+    const systemBreakpoints = Array.isArray(system) ? system.length : 0;
+    const toolBreakpoints = toolCount > 0 && !Array.isArray(system) ? 1 : 0;
+    if (systemBreakpoints + toolBreakpoints + MESSAGE_CACHE_BREAKPOINTS > MAX_CACHE_BREAKPOINTS) {
+        throw new Error(`Anthropic cache_control budget exceeded: ${systemBreakpoints} system + ${toolBreakpoints} tool + ${MESSAGE_CACHE_BREAKPOINTS} message > ${MAX_CACHE_BREAKPOINTS}`);
+    }
+}
 export class AnthropicProvider {
     apiKey;
     model;
@@ -61,13 +110,23 @@ export class AnthropicProvider {
         }
         const system = systemBlocks.length ? systemBlocks : (context.systemText || undefined);
         const msgs = toAnthropicMessages(context, message => this.nativeAssistantBlocks.get(assistantReplayKey(message)));
+        applyMessageCacheControl(msgs);
+        // Append the volatile State turn AFTER the cache breakpoints (uncached tail);
+        // absent on un-rebuilt bindings, where the state is already inside `turns`.
+        if (context.stateTurn) {
+            msgs.push({
+                role: context.stateTurn.role === "assistant" ? "assistant" : "user",
+                content: context.stateTurn.content,
+            });
+        }
+        assertCacheBudget(system, tools.length);
         const body = {
             model: this.model,
             max_tokens: this.maxTokens,
             messages: msgs,
             stream: true,
             ...(system ? { system } : {}),
-            ...(tools.length ? { tools: buildAnthropicTools(tools) } : {}),
+            ...(tools.length ? { tools: buildAnthropicTools(tools, !Array.isArray(system)) } : {}),
         };
         if (extensions?.enable_thinking) {
             body.thinking = { type: "enabled", budget_tokens: 8000 };
@@ -91,6 +150,10 @@ export class AnthropicProvider {
         const reader = resp.body.getReader();
         const decoder = new TextDecoder();
         let buf = "";
+        let uncachedInput = 0;
+        let cacheReadTokens = 0;
+        let cacheCreationTokens = 0;
+        let outputTokens = 0;
         while (true) {
             const { done, value } = await reader.read();
             if (done)
@@ -108,15 +171,26 @@ export class AnthropicProvider {
                     const evt = JSON.parse(data);
                     if (evt.type === "message_start" || evt.type === "message_delta") {
                         const usage = (evt.usage ?? evt.message?.usage);
-                        if (usage?.input_tokens != null) {
-                            const inputTokens = usage.input_tokens ?? 0;
-                            const outputTokens = usage.output_tokens ?? 0;
-                            yield {
-                                type: "usage",
-                                totalTokens: inputTokens + outputTokens,
-                                inputTokens,
-                                outputTokens,
-                            };
+                        if (usage) {
+                            // input + cache counts are pinned at message_start; a later
+                            // message_delta may omit them — Math.max prevents zeroing.
+                            uncachedInput = Math.max(uncachedInput, usage.input_tokens ?? 0);
+                            cacheReadTokens = Math.max(cacheReadTokens, usage.cache_read_input_tokens ?? 0);
+                            cacheCreationTokens = Math.max(cacheCreationTokens, usage.cache_creation_input_tokens ?? 0);
+                            outputTokens = Math.max(outputTokens, usage.output_tokens ?? 0);
+                            // inputTokens is the FULL prompt (uncached + cache read + write):
+                            // the kernel reads it as the authoritative context size.
+                            const inputTokens = uncachedInput + cacheReadTokens + cacheCreationTokens;
+                            if (inputTokens > 0 || outputTokens > 0) {
+                                yield {
+                                    type: "usage",
+                                    totalTokens: inputTokens + outputTokens,
+                                    inputTokens,
+                                    outputTokens,
+                                    cacheReadInputTokens: cacheReadTokens,
+                                    cacheCreationInputTokens: cacheCreationTokens,
+                                };
+                            }
                         }
                     }
                     else if (evt.type === "content_block_start") {

package/dist/providers/base.d.ts

@@ -1,5 +1,10 @@
 import type { Message, RenderedContext } from "../types.js";
 import { assistantReplayKey } from "../runtime/provider-replay.js";
+/** History turns with the volatile State turn appended as the latest turn
+ *  (OpenAI), keeping the history a stable cacheable prefix. Anthropic appends it
+ *  after the cache breakpoint. Absent on un-rebuilt bindings — then the state is
+ *  already inside `turns`. */
+export declare function turnsWithStateAppended(context: RenderedContext): Message[];
 /** Build OpenAI-compatible chat messages from a RenderedContext. */
 export declare function toOpenAIMessages(context: RenderedContext): Array<Record<string, unknown>>;
 export declare function toAnthropicMessages(context: RenderedContext, nativeReplay?: (message: Message) => Array<Record<string, unknown>> | undefined): Array<Record<string, unknown>>;

package/dist/providers/base.js

@@ -7,18 +7,50 @@ function parseToolArguments(args) {
         return {};
     }
 }
+/** Multimodal: OpenAI content blocks from contentParts (text + image). */
+function openAIPartsContent(parts) {
+    return parts.map(p => {
+        if (p.type === "image") {
+            const url = p.data ? `data:${p.mediaType ?? "image/png"};base64,${p.data}` : (p.url ?? "");
+            return { type: "image_url", image_url: { url, ...(p.detail ? { detail: p.detail } : {}) } };
+        }
+        return { type: "text", text: p.text ?? p.output ?? "" };
+    });
+}
+/** Multimodal: Anthropic content blocks from contentParts (text + image). */
+function anthropicPartsContent(parts) {
+    return parts.map(p => {
+        if (p.type === "image") {
+            const source = p.data
+                ? { type: "base64", media_type: p.mediaType ?? "image/png", data: p.data }
+                : { type: "url", url: p.url ?? "" };
+            return { type: "image", source };
+        }
+        return { type: "text", text: p.text ?? p.output ?? "" };
+    });
+}
+/** History turns with the volatile State turn appended as the latest turn
+ *  (OpenAI), keeping the history a stable cacheable prefix. Anthropic appends it
+ *  after the cache breakpoint. Absent on un-rebuilt bindings — then the state is
+ *  already inside `turns`. */
+export function turnsWithStateAppended(context) {
+    return context.stateTurn ? [...context.turns, context.stateTurn] : context.turns;
+}
 /** Build OpenAI-compatible chat messages from a RenderedContext. */
 export function toOpenAIMessages(context) {
     const messages = [];
     if (context.systemText) {
         messages.push({ role: "system", content: context.systemText });
     }
-    for (const msg of context.turns) {
+    for (const msg of turnsWithStateAppended(context)) {
         if (msg.role === "tool") {
             messages.push({ role: "tool", content: msg.content });
             continue;
         }
-        const next = { role: msg.role, content: msg.content };
+        const next = {
+            role: msg.role,
+            content: msg.contentParts?.length ? openAIPartsContent(msg.contentParts) : msg.content,
+        };
         if (msg.role === "assistant" && msg.toolCalls?.length) {
             next.tool_calls = msg.toolCalls.map(tc => ({
                 id: tc.id,
@@ -55,7 +87,10 @@ export function toAnthropicMessages(context, nativeReplay) {
             result.push({ role: "assistant", content: blocks });
             continue;
         }
-        result.push({ role: msg.role, content: msg.content });
+        result.push({
+            role: msg.role,
+            content: msg.contentParts?.length ? anthropicPartsContent(msg.contentParts) : msg.content,
+        });
     }
     if (context.systemVolatile && result.length > 0) {
         const last = result[result.length - 1];

package/dist/types.d.ts

@@ -1,9 +1,27 @@
 import type { WorkflowNodeSpec } from "./runtime/types/agent.js";
+export interface ContentPart {
+    type: "text" | "image" | "audio" | "tool_result";
+    text?: string;
+    /** Remote image URL (mutually exclusive with `data`). */
+    url?: string;
+    /** Raw base64-encoded bytes (image/audio). */
+    data?: string;
+    /** MIME type, e.g. `"image/png"`. */
+    mediaType?: string;
+    /** OpenAI vision detail level. */
+    detail?: "auto" | "low" | "high";
+    callId?: string;
+    output?: string;
+    isError?: boolean;
+}
 export interface Message {
     role: "system" | "user" | "assistant" | "tool";
     content: string;
     tokenCount?: number;
     toolCalls?: ToolCall[];
+    /** Multimodal parts (text + image/audio). When present, providers render these
+     *  instead of the plain `content` string. */
+    contentParts?: ContentPart[];
 }
 export interface ToolCall {
     id: string;
@@ -31,6 +49,9 @@ export interface RenderedContext {
     systemKnowledge?: string;
     systemVolatile?: string;
     turns: Message[];
+    /** Volatile State turn (task_state + signals), rendered after the cacheable
+     *  history. Absent on un-rebuilt bindings — then it's still inside turns[0]. */
+    stateTurn?: Message;
 }
 export interface StreamEvent {
     type: string;
@@ -44,6 +65,8 @@ export interface UsageEvent extends StreamEvent {
     totalTokens: number;
     inputTokens?: number;
     outputTokens?: number;
+    cacheReadInputTokens?: number;
+    cacheCreationInputTokens?: number;
 }
 export interface ThinkingDelta extends StreamEvent {
     type: "thinking_delta";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@deepstrike/wasm",
-  "version": "0.2.11",
+  "version": "0.2.13",
   "description": "DeepStrike WASM SDK — browser, Cloudflare Workers, Deno Deploy",
   "type": "module",
   "main": "dist/index.js",
@@ -15,7 +15,7 @@
     "test": "node --experimental-vm-modules node_modules/.bin/jest"
   },
   "dependencies": {
-    "@deepstrike/wasm-kernel": "0.2.11"
+    "@deepstrike/wasm-kernel": "0.2.13"
   },
   "devDependencies": {
     "@types/jest": "^30.0.0",