hammer-ai 0.2.12 → 0.2.14

package/README.md

@@ -0,0 +1,332 @@
+# hammer-ai
+
+Infrastructure for building tool-calling chat agents in TypeScript.
+
+This package gives you:
+- OpenAI-compatible LLM client with streaming support
+- Agent loop runtime for web/chat interfaces
+- Tool base class and tool registry
+- Built-in handling for tool, bash, and background_bash run targets
+- Memory and validation layers for longer multi-step workflows
+
+The examples below are based on a real production-style integration from monoslides.
+
+## Install
+
+```bash
+bun add hammer-ai
+```
+
+## Mental model
+
+```mermaid
+flowchart LR
+  UI[Chat UI] --> Runtime[WebToolLoopAgentRuntime]
+  Runtime --> LLM[LLMClient]
+  Runtime --> Registry[ToolRegistry]
+  Registry --> Local[Local Tool classes]
+  Registry --> Proxy[ProxyTool to server actions]
+  Registry --> Bash[Bash and BackgroundBash run commands]
+  Runtime --> Store[Runtime store]
+  Store --> UI
+```
+
+## 1. Configure providers (Qwen and DeepSeek)
+
+Configure once at startup. You can keep multiple named presets and choose one for your runtime.
+
+```ts
+import { configure } from "hammer-ai"
+
+configure({
+  providers: {
+    "qwen-plus": {
+      apiKey: process.env.DASHSCOPE_API_KEY!,
+      baseUrl: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+      model: "qwen-plus",
+      enableThinking: false,
+    },
+    deepseek: {
+      apiKey: process.env.DEEPSEEK_API_KEY!,
+      baseUrl: "https://api.deepseek.com",
+      model: "deepseek-v4-flash",
+      enableThinking: false,
+    },
+    "deepseek-reasoner": {
+      apiKey: process.env.DEEPSEEK_API_KEY!,
+      baseUrl: "https://api.deepseek.com",
+      model: "deepseek-reasoner",
+      enableThinking: true,
+    },
+  },
+  compactionProvider: "deepseek",
+})
+```
+
+Notes:
+- For Qwen-style endpoints, enableThinking maps to enable_thinking.
+- For DeepSeek endpoints, enableThinking maps to thinking.type (enabled or disabled).
+
+## 2. Create tools
+
+Create local tools by extending Tool.
+
+```ts
+import { Tool, type ToolResult, type ToolSchema } from "hammer-ai"
+
+export class GetSlideFilePath extends Tool {
+  override getName(): string {
+    return "GetSlideFilePath"
+  }
+
+  override getDescription(): string {
+    return "Return the absolute path for the active slide SVG"
+  }
+
+  override getSchema(): ToolSchema {
+    return {
+      slideId: {
+        type: "string",
+        required: true,
+        positional: false,
+        description: "Slide identifier",
+      },
+    }
+  }
+
+  override async execute(params: Record<string, any>): Promise<ToolResult> {
+    const slideId = String(params.slideId ?? "").trim()
+    if (!slideId) return { success: false, error: "slideId is required" }
+
+    return {
+      success: true,
+      path: `/workspace/slide-${slideId}.svg`,
+    }
+  }
+}
+```
+
+## 3. Wire registry + proxy tools + bash
+
+Use createToolRegistry for a mixed tool surface:
+- local tool classes
+- proxy tools that call server actions
+- bash run command bindings
+
+```ts
+import {
+  createToolRegistry,
+  ProxyTool,
+  createRunCommandRuntimeBindings,
+  BackgroundBashRunCommand,
+} from "hammer-ai"
+import { BashRunCommand } from "hammer-ai"
+
+import { GetSlideFilePath } from "./tools/GetSlideFilePath"
+import { CreateOrReplaceSlide } from "./tools/CreateOrReplaceSlide"
+import { executeWebToolAction } from "./actions/web-tool-actions"
+
+class AppBashRunCommand extends BashRunCommand {
+  protected override async executeCommand(command: string) {
+    // Route shell execution to your sandbox/session implementation.
+    return executeInSandbox(command)
+  }
+}
+
+const registry = createToolRegistry(
+  [
+    new ProxyTool(
+      {
+        name: "BraveWebSearch",
+        description: "Search the web",
+        parameters: {
+          query: {
+            type: "string",
+            description: "Search query",
+            required: true,
+          },
+        },
+      },
+      (parameters) => executeWebToolAction({ tool: "BraveWebSearch", input: parameters }),
+    ),
+    new GetSlideFilePath(),
+    new CreateOrReplaceSlide(),
+  ],
+  createRunCommandRuntimeBindings(
+    new AppBashRunCommand(),
+    new BackgroundBashRunCommand(),
+  ),
+)
+
+const executeTool = registry.createExecutor()
+```
+
+## 4. Build a chat runtime
+
+WebToolLoopAgentRuntime is the easiest way to run a multi-step tool loop and stream content into UI state.
+
+```ts
+import {
+  createRuntimeStore,
+  createInitialWebAgentState,
+  WebToolLoopAgentRuntime,
+  LLMClient,
+  getProviderConfig,
+  buildAgentIdentityLine,
+  buildWebRuntimeRules,
+} from "hammer-ai"
+
+const runtimeStore = createRuntimeStore(() => createInitialWebAgentState())
+
+class SlidesRuntime extends WebToolLoopAgentRuntime {
+  constructor() {
+    super({
+      store: runtimeStore,
+      messageIdPrefix: "slides-msg",
+      llmClient: new LLMClient({
+        ...getProviderConfig("deepseek"),
+        enableThinking: false,
+      }),
+      getToolDefinitions: () => registry.getToolDefinitions(),
+      executeTool,
+      memoryPreset: "monoslides",
+      systemIdentity: buildAgentIdentityLine({
+        agentName: "MonoSlides",
+        roleDescription: "Slide design assistant",
+      }),
+      extraRules: buildWebRuntimeRules({
+        includeVerifiedCompletionRule: true,
+      }),
+      temperature: 0.2,
+      maxTokens: 8192,
+      allowedRunTargets: ["tool", "bash"],
+    })
+  }
+
+  async run(userTask: string): Promise<void> {
+    await this.executeDefaultWebRun(userTask, {
+      shouldSurfaceAssistantContent: (_reasoning, rawContent) => Boolean(rawContent),
+    })
+  }
+
+  abort(): void {
+    this.defaultWebAbort()
+  }
+
+  reset(): void {
+    this.defaultWebReset()
+  }
+}
+
+export const runtime = new SlidesRuntime()
+```
+
+## 5. Expose controller actions
+
+A small controller wrapper keeps runtime actions and store in one object.
+
+```ts
+import { defineRuntimeController } from "hammer-ai"
+
+export const chatController = defineRuntimeController({
+  store: runtimeStore,
+  actions: {
+    run: (prompt: string) => runtime.run(prompt),
+    abort: () => runtime.abort(),
+    reset: () => runtime.reset(),
+  },
+  refs: {} as const,
+})
+```
+
+## 6. Render in UI
+
+You can subscribe directly with useSyncExternalStore.
+
+```tsx
+import { useSyncExternalStore, useState } from "react"
+
+function useRuntimeState() {
+  return useSyncExternalStore(
+    chatController.store.subscribe,
+    chatController.store.getSnapshot,
+    chatController.store.getServerSnapshot,
+  )
+}
+
+export function ChatPanel() {
+  const state = useRuntimeState()
+  const [input, setInput] = useState("")
+  const isRunning = state.phase === "thinking" || state.phase === "tool-calling"
+
+  return (
+    <div>
+      <div>
+        {state.messages.map((message) => (
+          <div key={message.id}>{message.content}</div>
+        ))}
+        {state.streamingContent ? <div>{state.streamingContent}</div> : null}
+      </div>
+
+      <form
+        onSubmit={(event) => {
+          event.preventDefault()
+          const next = input.trim()
+          if (!next || isRunning) return
+          chatController.actions.run(next)
+          setInput("")
+        }}
+      >
+        <textarea value={input} onChange={(event) => setInput(event.target.value)} />
+        {isRunning ? (
+          <button type="button" onClick={() => chatController.actions.abort()}>Stop</button>
+        ) : (
+          <button type="submit">Send</button>
+        )}
+      </form>
+    </div>
+  )
+}
+```
+
+## 7. Recommended production patterns
+
+- Keep local, deterministic operations as Tool subclasses.
+- Wrap privileged/server-only capabilities (web calls, secrets, DB) with ProxyTool.
+- Give the model only the run targets you want via allowedRunTargets.
+- Keep strict completion rules (for example: todo list must be completed before exit).
+- Use executeDefaultWebRun for serialization, abort handling, and consistent state transitions.
+- Prefer enableThinking: false for latency-sensitive UX, and enable it only for tasks that need deeper reasoning.
+
+## 8. Minimal server action bridge example
+
+```ts
+import { createWebSearchToolActions } from "hammer-ai"
+
+const actions = createWebSearchToolActions({
+  executeWebSearch: (tool, input) => {
+    if (tool === "BochaWebSearch") return bocha.execute(input)
+    return brave.execute(input)
+  },
+})
+
+export async function executeWebToolAction(input: {
+  tool: "BraveWebSearch" | "BochaWebSearch"
+  input: Record<string, unknown>
+}) {
+  return actions.executeWebTool(input)
+}
+```
+
+## API Surface
+
+Main exports used in this guide:
+- configure, getProviderConfig
+- LLMClient
+- Tool, ProxyTool, SubAgentTool
+- createToolRegistry, createRunCommandRuntimeBindings
+- BashRunCommand, BackgroundBashRunCommand
+- WebToolLoopAgentRuntime
+- createRuntimeStore, defineRuntimeController
+- createInitialWebAgentState
+- createWebSearchToolActions

package/dist/index.d.ts

@@ -32,14 +32,14 @@ interface LLMProviderConfig {
      * Explicitly enable or disable the provider's thinking/reasoning mode.
      *
      * - `false` — disables thinking (e.g. DashScope `enable_thinking: false` for
-     *   Qwen3 models, which have thinking on by default). Prevents the silent
-     *   multi-minute server-side CoT delay before the first token streams out.
+     *   Qwen3 models, or DeepSeek `thinking: { type: "disabled" }`).
      * - `true`  — explicitly enables thinking with the provider's default budget.
      * - `undefined` — no thinking-related field is sent; the provider uses its
      *   own model default.
      *
-     * Currently maps to `enable_thinking` in the request body, which is the
-     * DashScope OpenAI-compatible API parameter for Qwen3 models.
+     * Currently maps to provider-specific request fields:
+     * - Qwen/DashScope: `enable_thinking`
+     * - DeepSeek: `thinking: { type: "enabled" | "disabled" }`
      */
     enableThinking?: boolean;
 }
@@ -176,7 +176,7 @@ interface LLMResponse {
     outcome: LoopOutcome;
     finishReason?: string;
 }
-type ProviderName = "qwen-max" | "qwen-plus" | "openrouter-claude" | "openrouter-gemini" | "minimax" | "minimax-her" | "chatglm" | "kimi" | "doubao";
+type ProviderName = "qwen-max" | "qwen-plus" | "deepseek" | "deepseek-chat" | "deepseek-reasoner" | "openrouter-claude" | "openrouter-gemini" | "minimax" | "minimax-her" | "chatglm" | "kimi" | "doubao";
 
 /**
  * configure.ts — global configuration store for hammer-agent.
@@ -195,6 +195,8 @@ interface HammerAgentProviderPreset {
     model: string;
     /** Extra headers merged into every request (e.g. HTTP-Referer for OpenRouter). */
     extraHeaders?: Record<string, string>;
+    /** Provider-specific thinking/reasoning toggle (Qwen/DeepSeek). */
+    enableThinking?: boolean;
 }
 /**
  * Options accepted by `configure()`.
@@ -875,8 +877,6 @@ interface MemoryLayerConfig {
     protectedContextTokens: number;
     /** Maximum tokens for the rendered compressed state. */
     stateBudgetTokens: number;
-    /** Hard cap on raw history entries (safety net). */
-    maxRawHistory: number;
     /** Minimum turns between compaction attempts. */
     compactionDebounceTurns: number;
     /** Baseline token overhead for system prompt (conservative estimate). */
@@ -899,7 +899,7 @@ interface MemoryLayerConfig {
  * - State rendering (converting TState to human-readable text)
  *
  * The base class provides:
- * - Append-only raw history with hard cap enforcement
+ * - Append-only raw history
  * - Token-budgeted sliding window for recent messages
  * - buildMessages assembly (system → state → recent)
  * - Compaction triggering (debounce, threshold check, prune)
@@ -1152,8 +1152,6 @@ interface AgentMemoryLayerConfig {
     protectedContextTokens: number;
     /** Token budget for the rendered compressed state block. */
     stateBudgetTokens: number;
-    /** Hard cap on raw history entries. */
-    maxRawHistory: number;
     /** Minimum turns between compaction attempts. */
     compactionDebounceTurns: number;
     /** Token estimate for the system prompt. */

package/dist/index.js

@@ -213,7 +213,13 @@ var LLMClient = class {
       delete payload.temperature;
     }
     if (this.config.enableThinking !== void 0) {
-      payload.enable_thinking = this.config.enableThinking;
+      if (isDeepSeekRequest(this.config)) {
+        payload.thinking = {
+          type: this.config.enableThinking ? "enabled" : "disabled"
+        };
+      } else {
+        payload.enable_thinking = this.config.enableThinking;
+      }
     }
     const headers = {
       "Content-Type": "application/json",
@@ -465,6 +471,9 @@ var LLMClient = class {
     };
   }
 };
+function isDeepSeekRequest(config) {
+  return /deepseek/i.test(config.model) || /deepseek\.com/i.test(config.baseUrl);
+}
 var ApiError = class extends Error {
   constructor(status, body) {
     super(`API error ${status}: ${body}`);
@@ -4154,14 +4163,6 @@ var BaseMemoryLayer = class {
     const timestamp = Date.now();
     const msg = this.createMessage(id, role, content, this.currentTurn, timestamp);
     this.rawHistory.push(msg);
-    if (this.rawHistory.length > this.config.maxRawHistory) {
-      const excess = this.rawHistory.length - this.config.maxRawHistory;
-      const lastPrunedTurn = this.rawHistory[excess - 1].turn;
-      this.rawHistory.splice(0, excess);
-      if (this.compactionCursor.lastCompactedTurn < lastPrunedTurn) {
-        this.compactionCursor.lastCompactedTurn = lastPrunedTurn;
-      }
-    }
     this.onMessageAppended(msg);
     return id;
   }
@@ -5227,7 +5228,6 @@ var AgentMemoryLayer = class extends BaseMemoryLayer {
         compactionTokenThreshold: config.compactionTokenThreshold,
         protectedContextTokens: config.protectedContextTokens,
         stateBudgetTokens: config.stateBudgetTokens,
-        maxRawHistory: config.maxRawHistory,
         compactionDebounceTurns: config.compactionDebounceTurns,
         systemPromptOverhead: config.systemPromptOverhead,
         tokenEstimator: config.tokenEstimator,
@@ -5734,7 +5734,6 @@ var SHARED_WORKSPACE_AGENT_MEMORY_PRESET = {
   //  60_000
   stateBudgetTokens: Math.floor(DEFAULT_MAX_CONTEXT_TOKENS * 0.05),
   //  10_000
-  maxRawHistory: 2e3,
   compactionDebounceTurns: 3,
   systemPromptOverhead: 4e3,
   toolMemoryExtractor: DEFAULT_TOOL_MEMORY_EXTRACTOR
@@ -5744,7 +5743,6 @@ function createAgentMemoryLayer(_preset, overrides) {
     compactionTokenThreshold: SHARED_WORKSPACE_AGENT_MEMORY_PRESET.compactionTokenThreshold,
     protectedContextTokens: SHARED_WORKSPACE_AGENT_MEMORY_PRESET.protectedContextTokens,
     stateBudgetTokens: SHARED_WORKSPACE_AGENT_MEMORY_PRESET.stateBudgetTokens,
-    maxRawHistory: SHARED_WORKSPACE_AGENT_MEMORY_PRESET.maxRawHistory,
     compactionDebounceTurns: SHARED_WORKSPACE_AGENT_MEMORY_PRESET.compactionDebounceTurns,
     systemPromptOverhead: SHARED_WORKSPACE_AGENT_MEMORY_PRESET.systemPromptOverhead,
     tokenEstimator: overrides?.tokenEstimator ?? new CharTokenEstimator(),