@earendil-works/pi-agent-core

Stateful agent with tool execution and event streaming. Built on @earendil-works/pi-ai.

Installation

npm install @earendil-works/pi-agent-core

Quick Start

import { Agent } from "@earendil-works/pi-agent-core";\nimport { getModel } from "@earendil-works/pi-ai";\n\nconst agent = new Agent({\n  initialState: {\n    systemPrompt: "You are a helpful assistant.",\n    model: getModel("anthropic", "claude-sonnet-4-20250514"),\n  },\n});\n\nagent.subscribe((event) => {\n  if (event.type === "message_update" && event.assistantMessageEvent.type === "text_delta") {\n    // Stream just the new text chunk\n    process.stdout.write(event.assistantMessageEvent.delta);\n  }\n});\n\nawait agent.prompt("Hello!");

Core Concepts

AgentMessage vs LLM Message

The agent works with AgentMessage, a flexible type that can include:

Standard LLM messages (user, assistant, toolResult)
Custom app-specific message types via declaration merging

LLMs only understand user, assistant, and toolResult. The convertToLlm function bridges this gap by filtering and transforming messages before each LLM call.

Message Flow

AgentMessage[] → transformContext() → AgentMessage[] → convertToLlm() → Message[] → LLM\n                    (optional)                           (required)\n

transformContext: Prune old messages, inject external context
convertToLlm: Filter out UI-only messages, convert custom types to LLM format

Event Flow

The agent emits events for UI updates. Understanding the event sequence helps build responsive interfaces.

prompt() Event Sequence

When you call prompt("Hello"):

prompt("Hello")\n├─ agent_start\n├─ turn_start\n├─ message_start   { message: userMessage }      // Your prompt\n├─ message_end     { message: userMessage }\n├─ message_start   { message: assistantMessage } // LLM starts responding\n├─ message_update  { message: partial... }       // Streaming chunks\n├─ message_update  { message: partial... }\n├─ message_end     { message: assistantMessage } // Complete response\n├─ turn_end        { message, toolResults: [] }\n└─ agent_end       { messages: [...] }\n

With Tool Calls

If the assistant calls tools, the loop continues:

prompt("Read config.json")\n├─ agent_start\n├─ turn_start\n├─ message_start/end  { userMessage }\n├─ message_start      { assistantMessage with toolCall }\n├─ message_update...\n├─ message_end        { assistantMessage }\n├─ tool_execution_start  { toolCallId, toolName, args }\n├─ tool_execution_update { partialResult }           // If tool streams\n├─ tool_execution_end    { toolCallId, result }\n├─ message_start/end  { toolResultMessage }\n├─ turn_end           { message, toolResults: [toolResult] }\n│\n├─ turn_start                                        // Next turn\n├─ message_start      { assistantMessage }           // LLM responds to tool result\n├─ message_update...\n├─ message_end\n├─ turn_end\n└─ agent_end\n

Tool execution mode is configurable:

parallel (default): preflight tool calls sequentially, execute allowed tools concurrently, emit tool_execution_end as soon as each tool is finalized, then emit toolResult messages and turn_end.toolResults in assistant source order
sequential: execute tool calls one by one, matching the historical behavior

In parallel mode, tool completion events follow tool completion order, but persisted toolResult messages still follow assistant source order.

The mode can be set globally via toolExecution in the agent config, or per-tool via executionMode on AgentTool. If any tool call in a batch targets a tool with executionMode: "sequential", the entire batch executes sequentially regardless of the global setting.

The beforeToolCall hook runs after tool_execution_start and validated argument parsing. It can block execution. The afterToolCall hook runs after tool execution finishes and before tool_execution_end and final tool result message events are emitted.

Tools can also return terminate: true to hint that the automatic follow-up LLM call should be skipped. The loop only stops early when every finalized tool result in that batch sets terminate: true. Mixed batches continue normally.

Low-level loop callers can set shouldStopAfterTurn to stop gracefully after the current turn completes:

const stream = agentLoop(prompts, context, {\n  model,\n  convertToLlm,\n  shouldStopAfterTurn: async ({ message, toolResults, context, newMessages }) => {\n    return shouldCompactBeforeNextTurn(context.messages);\n  },\n});

shouldStopAfterTurn runs after turn_end is emitted and after the assistant response and any tool executions have completed normally. If it returns true, the loop emits agent_end and exits before polling steering or follow-up queues, and before starting another LLM call. It does not abort the provider stream, does not cancel running tools, and does not alter the assistant message stop reason.

When you use the Agent class, assistant message_end processing is treated as a barrier before tool preflight begins. That means beforeToolCall sees agent state that already includes the assistant message that requested the tool call.

continue() Event Sequence

continue() resumes from existing context without adding a new message. Use it for retries after errors.

// After an error, retry from current state\nawait agent.continue();

The last message in context must be user or toolResult (not assistant).

Event Types

EventDescriptionagent_startAgent begins processingagent_endFinal event for the run. Awaited subscribers for this event still count toward settlementturn_startNew turn begins (one LLM call + tool executions)turn_endTurn completes with assistant message and tool resultsmessage_startAny message begins (user, assistant, toolResult)message_updateAssistant only. Includes assistantMessageEvent with deltamessage_endMessage completestool_execution_startTool beginstool_execution_updateTool streams progresstool_execution_endTool completes.

Agent.subscribe() listeners are awaited in registration order. agent_end means no more loop events will be emitted, but await agent.waitForIdle() and await agent.prompt(...) only settle after awaited agent_end listeners finish.

Agent Options

const agent = new Agent({\n  // Initial state\n  initialState: {\n    systemPrompt: string,\n    model: Model<any>,\n    thinkingLevel: "off" | "minimal" | "low" | "medium" | "high" | "xhigh",\n    tools: AgentTool<any>[],\n    messages: AgentMessage[],\n  },\n\n  // Convert AgentMessage[] to LLM Message[] (required for custom message types)\n  convertToLlm: (messages) => messages.filter(...),\n\n  // Transform context before convertToLlm (for pruning, compaction)\n  transformContext: async (messages, signal) => pruneOldMessages(messages),\n\n  // Steering mode: "one-at-a-time" (default) or "all"\n  steeringMode: "one-at-a-time",\n\n  // Follow-up mode: "one-at-a-time" (default) or "all"\n  followUpMode: "one-at-a-time",\n\n  // Custom stream function (for proxy backends)\n  streamFn: streamProxy,\n\n  // Session ID for provider caching\n  sessionId: "session-123",\n\n  // Dynamic API key resolution (for expiring OAuth tokens)\n  getApiKey: async (provider) => refreshToken(),\n\n  // Tool execution mode: "parallel" (default) or "sequential"\n  toolExecution: "parallel",\n\n  // Preflight each tool call after args are validated. Can block execution.\n  beforeToolCall: async ({ toolCall, args, context }) => {\n    if (toolCall.name === "bash") {\n      return { block: true, reason: "bash is disabled" };\n    }\n  },\n\n  // Postprocess each tool result before final tool events are emitted.\n  afterToolCall: async ({ toolCall, result, isError, context }) => {\n    if (toolCall.name === "notify_done" && !isError) {\n      return { terminate: true };\n    }\n    if (!isError) {\n      return { details: { ...result.details, audited: true } };\n    }\n  },\n\n  // Custom thinking budgets for token-based providers\n  thinkingBudgets: {\n    minimal: 128,\n    low: 512,\n    medium: 1024,\n    high: 2048,\n  },\n});

Agent State

interface AgentState {\n  systemPrompt: string;\n  model: Model<any>;\n  thinkingLevel: ThinkingLevel;\n  tools: AgentTool<any>[];\n  messages: AgentMessage[];\n  readonly isStreaming: boolean;\n  readonly streamingMessage?: AgentMessage;\n  readonly pendingToolCalls: ReadonlySet<string>;\n  readonly errorMessage?: string;\n}

Access state via agent.state.

Assigning agent.state.tools = [...] or agent.state.messages = [...] copies the top-level array before storing it. Mutating the returned array mutates the current agent state.

During streaming, agent.state.streamingMessage contains the current partial assistant message.

agent.state.isStreaming remains true until the run fully settles, including awaited agent_end subscribers.

Methods

Prompting

// Text prompt\nawait agent.prompt("Hello");\n\n// With images\nawait agent.prompt("What's in this image?", [\n  { type: "image", data: base64Data, mimeType: "image/jpeg" }\n]);\n\n// AgentMessage directly\nawait agent.prompt({ role: "user", content: "Hello", timestamp: Date.now() });\n\n// Continue from current context (last message must be user or toolResult)\nawait agent.continue();

State Management

agent.state.systemPrompt = "New prompt";\nagent.state.model = getModel("openai", "gpt-4o");\nagent.state.thinkingLevel = "medium";\nagent.state.tools = [myTool];\nagent.toolExecution = "sequential";\nagent.beforeToolCall = async ({ toolCall }) => undefined;\nagent.afterToolCall = async ({ toolCall, result }) => undefined;\nagent.state.messages = newMessages; // top-level array is copied\nagent.state.messages.push(message);\nagent.reset();

Session and Thinking Budgets

agent.sessionId = "session-123";\n\nagent.thinkingBudgets = {\n  minimal: 128,\n  low: 512,\n  medium: 1024,\n  high: 2048,\n};

Control

agent.abort();           // Cancel current operation\nawait agent.waitForIdle(); // Wait for completion

Events

const unsubscribe = agent.subscribe(async (event, signal) => {\n  if (event.type === "agent_end") {\n    // Final barrier work for the run\n    await flushSessionState(signal);\n  }\n});\nunsubscribe();

Steering and Follow-up

Steering messages let you interrupt the agent while tools are running. Follow-up messages let you queue work after the agent would otherwise stop.

agent.steeringMode = "one-at-a-time";\nagent.followUpMode = "one-at-a-time";\n\n// While agent is running tools\nagent.steer({\n  role: "user",\n  content: "Stop! Do this instead.",\n  timestamp: Date.now(),\n});\n\n// After the agent finishes its current work\nagent.followUp({\n  role: "user",\n  content: "Also summarize the result.",\n  timestamp: Date.now(),\n});\n\nconst steeringMode = agent.steeringMode;\nconst followUpMode = agent.followUpMode;\n\nagent.clearSteeringQueue();\nagent.clearFollowUpQueue();\nagent.clearAllQueues();

Use clearSteeringQueue, clearFollowUpQueue, or clearAllQueues to drop queued messages.

When steering messages are detected after a turn completes:

All tool calls from the current assistant message have already finished
Steering messages are injected
The LLM responds on the next turn

Follow-up messages are checked only when there are no more tool calls and no steering messages. If any are queued, they are injected and another turn runs.

Custom Message Types

Extend AgentMessage via declaration merging:

declare module "@earendil-works/pi-agent-core" {\n  interface CustomAgentMessages {\n    notification: { role: "notification"; text: string; timestamp: number };\n  }\n}\n\n// Now valid\nconst msg: AgentMessage = { role: "notification", text: "Info", timestamp: Date.now() };

Handle custom types in convertToLlm:

const agent = new Agent({\n  convertToLlm: (messages) => messages.flatMap(m => {\n    if (m.role === "notification") return []; // Filter out\n    return [m];\n  }),\n});

Tools

Define tools using AgentTool:

import { Type } from "typebox";\n\nconst readFileTool: AgentTool = {\n  name: "read_file",\n  label: "Read File",  // For UI display\n  description: "Read a file's contents",\n  parameters: Type.Object({\n    path: Type.String({ description: "File path" }),\n  }),\n  // Override execution mode for this tool (optional).\n  // "sequential" forces the entire batch to run one at a time.\n  // "parallel" allows concurrent execution with other tool calls.\n  // If omitted, the global toolExecution config applies.\n  executionMode: "sequential",\n  execute: async (toolCallId, params, signal, onUpdate) => {\n    const content = await fs.readFile(params.path, "utf-8");\n\n    // Optional: stream progress\n    onUpdate?.({ content: [{ type: "text", text: "Reading..." }], details: {} });\n\n    // Optional: add `terminate: true` here to skip the automatic follow-up LLM call\n    // when every finalized tool result in the batch does the same.\n    return {\n      content: [{ type: "text", text: content }],\n      details: { path: params.path, size: content.length },\n    };\n  },\n};\n\nagent.state.tools = [readFileTool];

Error Handling

Throw an error when a tool fails. Do not return error messages as content.

execute: async (toolCallId, params, signal, onUpdate) => {\n  if (!fs.existsSync(params.path)) {\n    throw new Error(`File not found: ${params.path}`);\n  }\n  // Return content only on success\n  return { content: [{ type: "text", text: "..." }] };\n}

Thrown errors are caught by the agent and reported to the LLM as tool errors with isError: true.

Return terminate: true from execute() or afterToolCall to hint that the agent should stop after the current tool batch. This only takes effect when every finalized tool result in the batch is terminating. The hint is runtime-only; emitted toolResult transcript messages remain standard LLM tool results.

Proxy Usage

For browser apps that proxy through a backend:

import { Agent, streamProxy } from "@earendil-works/pi-agent-core";\n\nconst agent = new Agent({\n  streamFn: (model, context, options) =>\n    streamProxy(model, context, {\n      ...options,\n      authToken: "...",\n      proxyUrl: "https://your-server.com",\n    }),\n});

Low-Level API

For direct control without the Agent class:

import { agentLoop, agentLoopContinue } from "@earendil-works/pi-agent-core";\n\nconst context: AgentContext = {\n  systemPrompt: "You are helpful.",\n  messages: [],\n  tools: [],\n};\n\nconst config: AgentLoopConfig = {\n  model: getModel("openai", "gpt-4o"),\n  convertToLlm: (msgs) => msgs.filter(m => ["user", "assistant", "toolResult"].includes(m.role)),\n  toolExecution: "parallel",  // overridden by per-tool executionMode if set\n  beforeToolCall: async ({ toolCall, args, context }) => undefined,\n  afterToolCall: async ({ toolCall, result, isError, context }) => undefined,\n};\n\nconst userMessage = { role: "user", content: "Hello", timestamp: Date.now() };\n\nfor await (const event of agentLoop([userMessage], context, config)) {\n  console.log(event.type);\n}\n\n// Continue from existing context\nfor await (const event of agentLoopContinue(context, config)) {\n  console.log(event.type);\n}

These low-level streams are observational. They preserve event order, but they do not wait for your async event handling to settle before later producer phases continue. If you need message processing to act as a barrier before tool preflight, use the Agent class instead of raw agentLoop() or agentLoopContinue().

","enableWebSigning","origin","homeMetadata",{"_1038":1816,"_1817":1062,"_1818":1819},"originHomeId",{"_1055":1026,"_37":1853,"_1030":2184,"_1023":-5,"_1057":-5,"_1058":-5,"_1059":-5,"_1060":-5},[],"meta",{"_2187":2188,"_2318":2319},"values",{"_2189":2190,"_2192":2193,"_2194":2195,"_2196":2197,"_2199":2200,"_2201":2202,"_2203":2204,"_2206":2207,"_2208":2209,"_2210":2211,"_2212":2213,"_2214":2215,"_2216":2217,"_2218":2219,"_2220":2221,"_2222":2223,"_2224":2225,"_2226":2227,"_2228":2229,"_2230":2231,"_2232":2233,"_2234":2235,"_2236":2237,"_2238":2239,"_2240":2241,"_2242":2243,"_2244":2245,"_2246":2247,"_2248":2249,"_2250":2251,"_2252":2253,"_2254":2255,"_2256":2257,"_2258":2259,"_2260":2261,"_2262":2263,"_2264":2265,"_2266":2267,"_2268":2269,"_2270":2271,"_2272":2273,"_2274":2275,"_2276":2277,"_2278":2279,"_2280":2281,"_2282":2283,"_2284":2285,"_2286":2287,"_2288":2289,"_2290":2291,"_2292":2293,"_2294":2295,"_2296":2297,"_2298":2299,"_2300":2301,"_2302":2303,"_2304":2305,"_2306":2307,"_2308":2309,"_2310":2311,"_2312":2313,"_2314":2315,"_2316":2317},"document.generationInfo.generation",[2191],"bigint","dehydratedState.queries.0.state.data.document.generationInfo.generation",[2191],"dehydratedState.queries.1.state.data.document.generationInfo.generation",[2191],"dehydratedState.queries.1.queryKey.2",[2198],"undefined","dehydratedState.queries.2.state.data.0.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.0.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.0.sortTime",[2205],"Date","dehydratedState.queries.2.state.data.0.activitySummary.latestCommentTime",[2198],"dehydratedState.queries.2.state.data.0.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.0.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.0.redirectInfo",[2198],"dehydratedState.queries.2.state.data.1.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.1.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.1.sortTime",[2205],"dehydratedState.queries.2.state.data.1.activitySummary.latestCommentTime",[2198],"dehydratedState.queries.2.state.data.1.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.1.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.1.redirectInfo",[2198],"dehydratedState.queries.2.state.data.2.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.2.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.2.sortTime",[2205],"dehydratedState.queries.2.state.data.2.activitySummary.latestCommentTime.seconds",[2191],"dehydratedState.queries.2.state.data.2.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.2.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.2.redirectInfo",[2198],"dehydratedState.queries.2.state.data.3.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.3.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.3.sortTime",[2205],"dehydratedState.queries.2.state.data.3.activitySummary.latestCommentTime",[2198],"dehydratedState.queries.2.state.data.3.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.3.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.3.redirectInfo",[2198],"dehydratedState.queries.2.state.data.4.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.4.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.4.sortTime",[2205],"dehydratedState.queries.2.state.data.4.activitySummary.latestCommentTime.seconds",[2191],"dehydratedState.queries.2.state.data.4.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.4.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.4.redirectInfo",[2198],"dehydratedState.queries.2.state.data.5.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.5.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.5.sortTime",[2205],"dehydratedState.queries.2.state.data.5.activitySummary.latestCommentTime",[2198],"dehydratedState.queries.2.state.data.5.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.5.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.5.redirectInfo",[2198],"dehydratedState.queries.2.state.data.6.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.6.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.6.sortTime",[2205],"dehydratedState.queries.2.state.data.6.activitySummary.latestCommentTime.seconds",[2191],"dehydratedState.queries.2.state.data.6.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.6.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.6.redirectInfo",[2198],"dehydratedState.queries.2.state.data.7.createTime.seconds",[2191],"dehydratedState.queries.2.state.data.7.updateTime.seconds",[2191],"dehydratedState.queries.2.state.data.7.sortTime",[2205],"dehydratedState.queries.2.state.data.7.activitySummary.latestCommentTime",[2198],"dehydratedState.queries.2.state.data.7.activitySummary.latestChangeTime.seconds",[2191],"dehydratedState.queries.2.state.data.7.generationInfo.generation",[2191],"dehydratedState.queries.2.state.data.7.redirectInfo",[2198],"dehydratedState.queries.5.state.data.document.generationInfo.generation",[2191],"dehydratedState.queries.5.queryKey.2",[2198],"dehydratedState.queries.6.state.data.metadata.summary",[2198],"referentialEqualities",{"_2189":2320,"_2201":2321,"_2210":2322,"_2214":2323,"_2230":2324,"_2242":2325,"_2258":2326,"_2272":2327,"_2286":2328,"_2300":2329},[2192],[2208],[2312],[2216,2222],[2236],[2244,2250],[2264],[2278],[2292],[2306],"actionData","errors"]