@deepstrike/sdk 0.2.11 → 0.2.12

package/dist/providers/profiles.js

@@ -35,18 +35,36 @@ export const endpointProfiles = {
         protocol: "openai-chat",
         baseURL: "https://api.minimaxi.com/v1",
     },
+    "deepseek.anthropic": {
+        id: "deepseek.anthropic",
+        providerId: "deepseek",
+        protocol: "anthropic-messages",
+        baseURL: "https://api.deepseek.com/anthropic",
+    },
     "deepseek.openai": {
         id: "deepseek.openai",
         providerId: "deepseek",
         protocol: "openai-chat",
         baseURL: "https://api.deepseek.com",
     },
+    "kimi.anthropic": {
+        id: "kimi.anthropic",
+        providerId: "kimi",
+        protocol: "anthropic-messages",
+        baseURL: "https://api.moonshot.ai/anthropic",
+    },
     "kimi.openai": {
         id: "kimi.openai",
         providerId: "kimi",
         protocol: "openai-chat",
         baseURL: "https://api.moonshot.cn/v1",
     },
+    "qwen.anthropic": {
+        id: "qwen.anthropic",
+        providerId: "qwen",
+        protocol: "anthropic-messages",
+        baseURL: "https://dashscope-intl.aliyuncs.com/apps/anthropic",
+    },
     "qwen.dashscope": {
         id: "qwen.dashscope",
         providerId: "qwen",
@@ -77,6 +95,12 @@ export const endpointProfiles = {
         protocol: "gemini-embeddings",
         baseURL: "https://generativelanguage.googleapis.com",
     },
+    "glm.anthropic": {
+        id: "glm.anthropic",
+        providerId: "glm",
+        protocol: "anthropic-messages",
+        baseURL: "https://api.z.ai/api/anthropic",
+    },
     "glm.openai": {
         id: "glm.openai",
         providerId: "glm",
@@ -375,28 +399,28 @@ export const modelProfiles = {
     },
     // ── DeepSeek ───────────────────────────────────────────────────────────────
     "deepseek/deepseek-chat": {
-        id: "deepseek/deepseek-chat", providerId: "deepseek", defaultEndpointId: "deepseek.openai",
+        id: "deepseek/deepseek-chat", providerId: "deepseek", defaultEndpointId: "deepseek.anthropic",
         contextWindow: 64_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: false, preserveAcrossToolTurns: false },
         policy: { maxTurns: 25 },
     },
     "deepseek/deepseek-reasoner": {
-        id: "deepseek/deepseek-reasoner", providerId: "deepseek", defaultEndpointId: "deepseek.openai",
+        id: "deepseek/deepseek-reasoner", providerId: "deepseek", defaultEndpointId: "deepseek.anthropic",
         contextWindow: 64_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: true },
         policy: { maxTurns: 50 },
     },
     "deepseek/deepseek-v4-flash": {
-        id: "deepseek/deepseek-v4-flash", providerId: "deepseek", defaultEndpointId: "deepseek.openai",
+        id: "deepseek/deepseek-v4-flash", providerId: "deepseek", defaultEndpointId: "deepseek.anthropic",
         contextWindow: 1_000_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: true },
         policy: { maxTurns: 20 },
     },
     "deepseek/deepseek-v4-pro": {
-        id: "deepseek/deepseek-v4-pro", providerId: "deepseek", defaultEndpointId: "deepseek.openai",
+        id: "deepseek/deepseek-v4-pro", providerId: "deepseek", defaultEndpointId: "deepseek.anthropic",
         contextWindow: 1_000_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: true },
@@ -404,48 +428,48 @@ export const modelProfiles = {
     },
     // ── Kimi ───────────────────────────────────────────────────────────────────
     "kimi/moonshot-v1-8k": {
-        id: "kimi/moonshot-v1-8k", providerId: "kimi", defaultEndpointId: "kimi.openai",
+        id: "kimi/moonshot-v1-8k", providerId: "kimi", defaultEndpointId: "kimi.anthropic",
         contextWindow: 8_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: false, preserveAcrossToolTurns: false },
         policy: { maxTurns: 15 },
     },
     "kimi/moonshot-v1-32k": {
-        id: "kimi/moonshot-v1-32k", providerId: "kimi", defaultEndpointId: "kimi.openai",
+        id: "kimi/moonshot-v1-32k", providerId: "kimi", defaultEndpointId: "kimi.anthropic",
         contextWindow: 32_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: false, preserveAcrossToolTurns: false },
         policy: { maxTurns: 20 },
     },
     "kimi/moonshot-v1-128k": {
-        id: "kimi/moonshot-v1-128k", providerId: "kimi", defaultEndpointId: "kimi.openai",
+        id: "kimi/moonshot-v1-128k", providerId: "kimi", defaultEndpointId: "kimi.anthropic",
         contextWindow: 128_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: false, preserveAcrossToolTurns: false },
         policy: { maxTurns: 30 },
     },
     "kimi/kimi-k2.5": {
-        id: "kimi/kimi-k2.5", providerId: "kimi", defaultEndpointId: "kimi.openai",
+        id: "kimi/kimi-k2.5", providerId: "kimi", defaultEndpointId: "kimi.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 30 },
     },
     "kimi/kimi-k2.6": {
-        id: "kimi/kimi-k2.6", providerId: "kimi", defaultEndpointId: "kimi.openai",
+        id: "kimi/kimi-k2.6", providerId: "kimi", defaultEndpointId: "kimi.anthropic",
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 35 },
     },
     "kimi/kimi-k2-thinking": {
-        id: "kimi/kimi-k2-thinking", providerId: "kimi", defaultEndpointId: "kimi.openai",
+        id: "kimi/kimi-k2-thinking", providerId: "kimi", defaultEndpointId: "kimi.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: true },
         policy: { maxTurns: 50 },
     },
     "kimi/kimi-k2-thinking-turbo": {
-        id: "kimi/kimi-k2-thinking-turbo", providerId: "kimi", defaultEndpointId: "kimi.openai",
+        id: "kimi/kimi-k2-thinking-turbo", providerId: "kimi", defaultEndpointId: "kimi.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: true },
@@ -453,91 +477,91 @@ export const modelProfiles = {
     },
     // ── Qwen ───────────────────────────────────────────────────────────────────
     "qwen/qwen3.7-max-preview": {
-        id: "qwen/qwen3.7-max-preview", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.7-max-preview", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 45 },
     },
     "qwen/qwen3.7-plus-preview": {
-        id: "qwen/qwen3.7-plus-preview", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.7-plus-preview", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 1_000_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 40 },
     },
     "qwen/qwen3.6-max-preview": {
-        id: "qwen/qwen3.6-max-preview", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.6-max-preview", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 40 },
     },
     "qwen/qwen3.6-plus": {
-        id: "qwen/qwen3.6-plus", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.6-plus", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 1_000_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 35 },
     },
     "qwen/qwen3.6-flash": {
-        id: "qwen/qwen3.6-flash", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.6-flash", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 1_000_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 20 },
     },
     "qwen/qwen3.6-35b-a3b": {
-        id: "qwen/qwen3.6-35b-a3b", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.6-35b-a3b", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 25 },
     },
     "qwen/qwen3.6-27b": {
-        id: "qwen/qwen3.6-27b", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.6-27b", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 25 },
     },
     "qwen/qwen3.5-plus": {
-        id: "qwen/qwen3.5-plus", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.5-plus", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 1_000_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 35 },
     },
     "qwen/qwen3.5-flash": {
-        id: "qwen/qwen3.5-flash", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.5-flash", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 1_000_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 20 },
     },
     "qwen/qwen3.5-397b-a17b": {
-        id: "qwen/qwen3.5-397b-a17b", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.5-397b-a17b", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 35 },
     },
     "qwen/qwen3.5-122b-a10b": {
-        id: "qwen/qwen3.5-122b-a10b", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.5-122b-a10b", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 25 },
     },
     "qwen/qwen3.5-35b-a3b": {
-        id: "qwen/qwen3.5-35b-a3b", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.5-35b-a3b", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
         policy: { maxTurns: 20 },
     },
     "qwen/qwen3.5-27b": {
-        id: "qwen/qwen3.5-27b", providerId: "qwen", defaultEndpointId: "qwen.dashscope",
+        id: "qwen/qwen3.5-27b", providerId: "qwen", defaultEndpointId: "qwen.anthropic",
         contextWindow: 256_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: false },
@@ -645,28 +669,28 @@ export const modelProfiles = {
     },
     // ── GLM ────────────────────────────────────────────────────────────────────
     "glm/glm-5.1": {
-        id: "glm/glm-5.1", providerId: "glm", defaultEndpointId: "glm.openai",
+        id: "glm/glm-5.1", providerId: "glm", defaultEndpointId: "glm.anthropic",
         contextWindow: 200_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: true, preserveAcrossToolTurns: true },
         policy: { maxTurns: 50 },
     },
     "glm/glm-4-plus": {
-        id: "glm/glm-4-plus", providerId: "glm", defaultEndpointId: "glm.openai",
+        id: "glm/glm-4-plus", providerId: "glm", defaultEndpointId: "glm.anthropic",
         contextWindow: 128_000,
         modalities: { input: ["text", "image"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: false, preserveAcrossToolTurns: false },
         policy: { maxTurns: 35 },
     },
     "glm/glm-4-flash": {
-        id: "glm/glm-4-flash", providerId: "glm", defaultEndpointId: "glm.openai",
+        id: "glm/glm-4-flash", providerId: "glm", defaultEndpointId: "glm.anthropic",
         contextWindow: 128_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: false, preserveAcrossToolTurns: false },
         policy: { maxTurns: 15 },
     },
     "glm/glm-4-air": {
-        id: "glm/glm-4-air", providerId: "glm", defaultEndpointId: "glm.openai",
+        id: "glm/glm-4-air", providerId: "glm", defaultEndpointId: "glm.anthropic",
         contextWindow: 128_000,
         modalities: { input: ["text"], output: ["text"] },
         tools: { supported: true }, reasoning: { supported: false, preserveAcrossToolTurns: false },

package/dist/providers/qwen.d.ts

@@ -1,7 +1,19 @@
 import OpenAI from "openai";
 import type { LLMProvider, Message, ProviderDescriptor, RenderedContext, StreamEvent, ToolSchema, RuntimePolicy, ProviderReplay } from "../types.js";
+import { AnthropicProvider } from "./anthropic.js";
 import { CircuitBreaker } from "./base.js";
 import { OpenAIChatAdapter } from "./openai-chat.js";
+/**
+ * Qwen over its Anthropic-compatible endpoint.
+ */
+export declare class QwenAnthropicProvider extends AnthropicProvider {
+    constructor(apiKey: string, model?: string, retry?: {
+        maxRetries: number;
+        baseDelay: number;
+    }, baseURL?: string);
+    protected providerName(): string;
+    runtimePolicy(): RuntimePolicy;
+}
 export declare class QwenProvider implements LLMProvider {
     protected readonly model: string;
     protected client: OpenAI;

package/dist/providers/qwen.js

@@ -1,9 +1,9 @@
 import OpenAI from "openai";
 import { withServerRuntimeGuard } from "../runtime/server.js";
-import { CircuitBreaker, omitExtensionKeys } from "./base.js";
+import { AnthropicProvider } from "./anthropic.js";
+import { CircuitBreaker, omitExtensionKeys, openAICachedPromptTokens } from "./base.js";
 import { OpenAIChatAdapter } from "./openai-chat.js";
 import { endpointProfiles } from "./profiles.js";
-const QWEN_BASE = endpointProfiles["qwen.dashscope"].baseURL;
 const QWEN_POLICIES = {
     "qwen3.7-max-preview": { maxTurns: 45 },
     "qwen3.7-plus-preview": { maxTurns: 40 },
@@ -19,6 +19,23 @@ const QWEN_POLICIES = {
     "qwen3.5-35b-a3b": { maxTurns: 20 },
     "qwen3.5-27b": { maxTurns: 20 },
 };
+/**
+ * Qwen over its Anthropic-compatible endpoint.
+ */
+export class QwenAnthropicProvider extends AnthropicProvider {
+    constructor(apiKey, model = "qwen3.6-plus", retry, baseURL = endpointProfiles["qwen.anthropic"].baseURL) {
+        super(apiKey, model, retry, {
+            baseURL,
+            authMode: "api-key",
+        });
+    }
+    providerName() {
+        return "qwen";
+    }
+    runtimePolicy() {
+        return QWEN_POLICIES[this.model] ?? {};
+    }
+}
 export class QwenProvider {
     model;
     client;
@@ -26,7 +43,7 @@ export class QwenProvider {
     maxRetries;
     baseDelay;
     chat = new OpenAIChatAdapter();
-    constructor(apiKey, model = "qwen3.6-plus", retry = { maxRetries: 3, baseDelay: 1000 }, baseURL = QWEN_BASE) {
+    constructor(apiKey, model = "qwen3.6-plus", retry = { maxRetries: 3, baseDelay: 1000 }, baseURL = endpointProfiles["qwen.dashscope"].baseURL) {
         this.model = model;
         this.client = withServerRuntimeGuard(() => new OpenAI({ apiKey, baseURL }));
         this.circuit = new CircuitBreaker();
@@ -113,11 +130,13 @@ export class QwenProvider {
         let totalTokens = 0;
         let inputTokens = 0;
         let outputTokens = 0;
+        let cacheReadTokens = 0;
         for await (const chunk of stream) {
             if (chunk.usage) {
                 totalTokens = chunk.usage.total_tokens;
                 inputTokens = chunk.usage.prompt_tokens ?? 0;
                 outputTokens = chunk.usage.completion_tokens ?? 0;
+                cacheReadTokens = openAICachedPromptTokens(chunk.usage);
                 continue;
             }
             const choice = chunk.choices[0];
@@ -184,7 +203,7 @@ export class QwenProvider {
             yield { type: "tool_call", id: tb.id, name: tb.name, arguments: args };
         }
         if (totalTokens > 0)
-            yield { type: "usage", totalTokens, inputTokens, outputTokens };
+            yield { type: "usage", totalTokens, inputTokens, outputTokens, ...(cacheReadTokens > 0 ? { cacheReadInputTokens: cacheReadTokens } : {}) };
     }
     thinkingExtraBody(extensions) {
         const enableThinking = Boolean(extensions?.enableThinking ?? extensions?.enable_thinking);

package/dist/runtime/runner.d.ts

@@ -1,4 +1,4 @@
-import type { LLMProvider, Message, ToolSchema, StreamEvent, ToolSuspendEvent, PermissionRequestEvent, PermissionResponse, AsyncSummarizer, DreamSummarizer } from "../types.js";
+import type { LLMProvider, Message, ContentPart, ToolSchema, StreamEvent, ToolSuspendEvent, PermissionRequestEvent, PermissionResponse, AsyncSummarizer, DreamSummarizer } from "../types.js";
 import type { DreamStore, MemoryEntry, MemoryQuery, MemoryWriteRequest } from "../memory/protocols.js";
 import type { KnowledgeSource } from "../knowledge/source.js";
 import type { SignalSource } from "../signals/types.js";
@@ -206,6 +206,8 @@ export declare class RuntimeRunner {
         sessionId: string;
         goal: string;
         criteria?: string[];
+        /** Multimodal inputs (images / audio) attached to the task as a user message. */
+        attachments?: ContentPart[];
         extensions?: Record<string, unknown>;
         /** Parent transcript to preload (e.g. sub-agent full context inheritance). */
         inheritEvents?: Array<{

package/dist/runtime/runner.js

@@ -458,9 +458,10 @@ export class RuntimeRunner {
                 criteria: req.criteria ?? [],
                 agent_id: this.opts.agentId,
                 system_prompt: this.opts.systemPrompt,
+                ...(req.attachments?.length ? { attachments: req.attachments } : {}),
             });
         }
-        yield* this.execute(req.sessionId, req.goal, req.criteria ?? [], req.extensions, prior.length > 0 ? prior : undefined, midRun);
+        yield* this.execute(req.sessionId, req.goal, req.criteria ?? [], req.extensions, prior.length > 0 ? prior : undefined, midRun, req.attachments);
     }
     async *wake(sessionId, extensions) {
         const events = await this.opts.sessionLog.read(sessionId);
@@ -470,7 +471,7 @@ export class RuntimeRunner {
         if (!startEntry)
             throw new Error(`No run_started event for session: ${sessionId}`);
         const start = startEntry.event;
-        yield* this.execute(sessionId, start.goal, start.criteria, extensions, events, true);
+        yield* this.execute(sessionId, start.goal, start.criteria, extensions, events, true, start.attachments);
     }
     async *dream(agentId, nowMs = Date.now()) {
         if (!this.opts.dreamStore)
@@ -623,7 +624,7 @@ export class RuntimeRunner {
         }
         return { approved, denied, events };
     }
-    async *execute(sessionId, goal, criteria, extensions, priorEvents, resumeMidRun = false) {
+    async *execute(sessionId, goal, criteria, extensions, priorEvents, resumeMidRun = false, attachments) {
         this.interrupted = false;
         this.pendingObservations = [];
         this.pendingSpoolOutputs.clear();
@@ -791,6 +792,16 @@ export class RuntimeRunner {
                 },
             });
         }
+        // Multimodal upload: seed the user's attachments (images/audio) as a history
+        // message before start_run pushes the "[TASK STATE]" anchor. init_task does not
+        // clear history, so order becomes [attachment user msg, "Proceed…"] — both land
+        // in the first render. On resume the message is already in the replayed history.
+        if (!resumeMidRun && attachments?.length) {
+            kernelApply(runtime, this.pendingObservations, {
+                kind: "add_history_message",
+                message: attachmentsToKernelMessage(attachments),
+            });
+        }
         let action = resumeMidRun
             ? kernelAction(runtime, this.pendingObservations, { kind: "resume" })
             : kernelAction(runtime, this.pendingObservations, startPayload);
@@ -1227,6 +1238,31 @@ export class RuntimeRunner {
 function isMidRun(events) {
     return events.length > 0 && !events.some(e => e.event.kind === "run_terminal");
 }
+/**
+ * Build a kernel `add_history_message` payload from user attachments: a `user`
+ * message whose content is the multimodal parts in the kernel's serde shape
+ * (`Content::Parts`; image `media_type`, not `mediaType`). Lets a caller upload
+ * images/audio with the task — the message lands in history before the first render.
+ */
+function attachmentsToKernelMessage(parts) {
+    const content = parts.map(p => {
+        if (p.type === "image") {
+            return {
+                type: "image",
+                ...(p.url ? { url: p.url } : {}),
+                ...(p.data ? { data: p.data } : {}),
+                ...(p.mediaType ? { media_type: p.mediaType } : {}),
+                ...(p.detail ? { detail: p.detail } : {}),
+            };
+        }
+        if (p.type === "audio")
+            return { type: "audio", data: p.data, media_type: p.mediaType };
+        if (p.type === "text")
+            return { type: "text", text: p.text };
+        return { type: "text", text: "" };
+    });
+    return { role: "user", content };
+}
 function compressionAction(action) {
     if (action === "snip_compact" ||
         action === "micro_compact" ||

package/dist/runtime/session-log.d.ts

@@ -1,4 +1,4 @@
-import type { ProviderReplay, ToolCall, ToolErrorKind } from "../types.js";
+import type { ContentPart, ProviderReplay, ToolCall, ToolErrorKind } from "../types.js";
 import type { KernelEventCategory, KernelPrimitive } from "./kernel-event-log.js";
 export type RollbackReason = {
     kind: "fatal_tool_error";
@@ -26,6 +26,7 @@ export type SessionEvent = {
     criteria: string[];
     agent_id?: string;
     system_prompt?: string;
+    attachments?: ContentPart[];
 } | {
     kind: "llm_completed";
     turn: number;

package/dist/types.d.ts

@@ -73,6 +73,18 @@ export interface ToolCallEvent extends StreamEvent {
     name: string;
     arguments: Record<string, unknown>;
 }
+export interface UsageEvent extends StreamEvent {
+    type: "usage";
+    /** Full prompt size + output (the authoritative prompt size for context accounting). */
+    totalTokens: number;
+    /** Full prompt size: uncached input + cache reads + cache writes. */
+    inputTokens?: number;
+    outputTokens?: number;
+    /** Prompt tokens served from cache this request (billed ~0.1x). Subset of inputTokens. */
+    cacheReadInputTokens?: number;
+    /** Prompt tokens written to cache this request (billed ~1.25x). Subset of inputTokens. */
+    cacheCreationInputTokens?: number;
+}
 export type ToolChunk = string | {
     type: "text";
     text: string;
@@ -169,9 +181,14 @@ export interface ToolDeniedEvent extends StreamEvent {
     reason: string;
 }
 export interface TokenUsage {
+    /** Full prompt size: uncached input + cache reads + cache writes. */
     inputTokens: number;
     outputTokens: number;
     totalTokens: number;
+    /** Prompt tokens served from cache (billed ~0.1x). Subset of inputTokens. */
+    cacheReadInputTokens?: number;
+    /** Prompt tokens written to cache (billed ~1.25x). Subset of inputTokens. */
+    cacheCreationInputTokens?: number;
 }
 export interface ProviderToolSpec {
     name: string;
@@ -236,8 +253,24 @@ export interface RenderedContext {
     systemStable?: string;
     /** Knowledge (memory retrievals, skill definitions, artifacts). Anthropic system[1] with cache_control. */
     systemKnowledge?: string;
-    /** Turns: [0] = State (task_state + signals), [1..N] = History. */
+    /** History turns only — the stable, cacheable message prefix. */
     turns: Message[];
+    /**
+     * Volatile State turn (task_state + signals), rebuilt every call. Providers
+     * render it after the cacheable history (Anthropic: after the cache breakpoint;
+     * OpenAI-family: prepended, preserving order). Absent when produced by an
+     * older binding that has not been rebuilt — then the State turn is still inside
+     * `turns[0]` and providers render `turns` as-is.
+     */
+    stateTurn?: Message;
+    /**
+     * P1-E: count of leading `turns` forming the frozen prefix — byte-stable until the next
+     * compaction. The Anthropic provider pins a deep cache breakpoint at this boundary (a long-lived
+     * cache that survives many turns and is immune to the 20-block lookback miss on heavy tool turns)
+     * and rolls the other breakpoint at the tail. Absent (older binding, or no distinct frozen region
+     * yet) ⇒ the provider falls back to the rolling-pair placement.
+     */
+    frozenPrefixLen?: number;
 }
 /**
  * Runtime execution policy advertised by a provider.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@deepstrike/sdk",
-  "version": "0.2.11",
+  "version": "0.2.12",
   "description": "DeepStrike Node.js SDK",
   "type": "module",
   "main": "dist/index.js",
@@ -20,7 +20,7 @@
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.99.0",
-    "@deepstrike/core": "0.2.11",
+    "@deepstrike/core": "0.2.12",
     "@google/generative-ai": "^0.24.1",
     "openai": "^5.23.2"
   },