@cuylabs/agent-core 0.9.0 → 0.11.0

package/dist/chunk-WWYYNWEW.js

@@ -1,259 +0,0 @@
-// src/context/estimation.ts
-function estimateTokens(text) {
-  return Math.ceil(text.length / 4);
-}
-function estimateMessageTokens(message) {
-  if (typeof message.content === "string") {
-    return estimateTokens(message.content);
-  }
-  if (Array.isArray(message.content)) {
-    let total = 0;
-    for (const part of message.content) {
-      if (typeof part === "string") {
-        total += estimateTokens(part);
-      } else if ("text" in part && typeof part.text === "string") {
-        total += estimateTokens(part.text);
-      } else if ("type" in part && part.type === "image") {
-        total += 765;
-      }
-    }
-    return total;
-  }
-  return 0;
-}
-function estimateConversationTokens(messages) {
-  let total = 0;
-  for (const message of messages) {
-    total += estimateMessageTokens(message);
-    total += 4;
-  }
-  return total;
-}
-
-// src/agent/types/compaction.ts
-var PRUNE_PROTECTED_TOOLS = ["skill"];
-
-// src/context/pruning.ts
-var DEFAULT_CONTEXT_LIMITS = {
-  contextWindow: 128e3,
-  reserveTokens: 16e3,
-  // Reserve for output
-  protectedTokens: 4e4,
-  // Keep recent 40 k tokens
-  pruneMinimum: 2e4
-  // Don't prune until 20 k tokens
-};
-function isContextOverflowing(tokens, limits = DEFAULT_CONTEXT_LIMITS) {
-  const threshold = limits.contextWindow - limits.reserveTokens;
-  return tokens > threshold;
-}
-function shouldPruneContext(tokens, limits = DEFAULT_CONTEXT_LIMITS) {
-  if (tokens < limits.pruneMinimum) return false;
-  return isContextOverflowing(tokens, limits);
-}
-function findCutPoint(messages, protectedTokens = DEFAULT_CONTEXT_LIMITS.protectedTokens) {
-  if (messages.length === 0) return 0;
-  let tokensFromEnd = 0;
-  let cutIndex = messages.length;
-  for (let i = messages.length - 1; i >= 0; i--) {
-    tokensFromEnd += estimateMessageTokens(messages[i]);
-    if (tokensFromEnd >= protectedTokens) {
-      cutIndex = i;
-      break;
-    }
-  }
-  if (cutIndex <= 1) return 0;
-  const startIndex = cutIndex >= messages.length ? messages.length - 1 : cutIndex;
-  for (let i = startIndex; i >= 1; i--) {
-    const msg = messages[i];
-    const prevMsg = messages[i - 1];
-    if (!msg || !prevMsg) continue;
-    if (msg.role === "tool") continue;
-    if (prevMsg.role === "assistant" || prevMsg.role === "user") {
-      return i;
-    }
-  }
-  return 0;
-}
-function pruneToolResults(messages, protectedTokens = DEFAULT_CONTEXT_LIMITS.protectedTokens, options) {
-  const protectedToolSet = /* @__PURE__ */ new Set([
-    ...PRUNE_PROTECTED_TOOLS,
-    ...options?.protectedTools ?? []
-  ]);
-  let tokensFromEnd = 0;
-  const tokenPositions = [];
-  for (let i = messages.length - 1; i >= 0; i--) {
-    tokensFromEnd += estimateMessageTokens(messages[i]);
-    tokenPositions[i] = tokensFromEnd;
-  }
-  return messages.map((msg, i) => {
-    if (tokenPositions[i] < protectedTokens) return msg;
-    if (!("role" in msg) || msg.role !== "tool") return msg;
-    const toolMsg = msg;
-    if ("compactedAt" in toolMsg && toolMsg.compactedAt) return msg;
-    if (toolMsg.toolName && protectedToolSet.has(toolMsg.toolName)) return msg;
-    const currentTokens = estimateTokens(toolMsg.content);
-    if (currentTokens < 500) return msg;
-    return {
-      ...toolMsg,
-      content: `[Output pruned - was ${currentTokens} tokens]`,
-      compactedAt: Date.now()
-    };
-  });
-}
-
-// src/context/summarization.ts
-import { generateText } from "ai";
-var DEFAULT_SUMMARY_PROMPT = `You are summarizing a conversation to continue it with context.
-
-Create a structured summary that captures:
-1. **Goal**: What the user is trying to accomplish
-2. **Progress**: What has been done so far
-3. **Decisions**: Key decisions made during the conversation
-4. **Current State**: Where we left off
-5. **Next Steps**: What should happen next
-
-Be concise but comprehensive. Include specific file paths, function names, and technical details that would be lost otherwise.
-
-Format as a clear summary that could be given to another assistant to continue the work.`;
-async function generateSummary(messages, options) {
-  const conversationText = messages.map((m) => {
-    const role = m.role.toUpperCase();
-    const content = typeof m.content === "string" ? m.content : JSON.stringify(m.content);
-    return `[${role}]: ${content}`;
-  }).join("\n\n");
-  const prompt = options.customPrompt || DEFAULT_SUMMARY_PROMPT;
-  const { text } = await generateText({
-    model: options.model,
-    maxOutputTokens: options.maxTokens ?? 2e3,
-    system: prompt,
-    prompt: `Summarize this conversation:
-
-${conversationText}`
-  });
-  return text;
-}
-async function pruneContext(messages, options = {}) {
-  const limits = options.limits ?? DEFAULT_CONTEXT_LIMITS;
-  let currentMessages = [...messages];
-  let tokensRemoved = 0;
-  let removedCount = 0;
-  let summarized = false;
-  let summary;
-  const initialTokens = estimateConversationTokens(currentMessages);
-  if (!shouldPruneContext(initialTokens, limits)) {
-    return { messages: currentMessages, removedCount: 0, tokensRemoved: 0, summarized: false };
-  }
-  const prunedMessages = pruneToolResults(currentMessages, limits.protectedTokens);
-  const afterPruneTokens = estimateConversationTokens(prunedMessages);
-  tokensRemoved = initialTokens - afterPruneTokens;
-  currentMessages = prunedMessages;
-  if (!isContextOverflowing(afterPruneTokens, limits)) {
-    return { messages: currentMessages, removedCount: 0, tokensRemoved, summarized: false };
-  }
-  const cutIndex = findCutPoint(currentMessages, limits.protectedTokens);
-  if (cutIndex === 0) {
-    return { messages: currentMessages, removedCount: 0, tokensRemoved, summarized: false };
-  }
-  const toSummarize = currentMessages.slice(0, cutIndex);
-  const toKeep = currentMessages.slice(cutIndex);
-  removedCount = toSummarize.length;
-  tokensRemoved += estimateConversationTokens(toSummarize);
-  if (options.model) {
-    summary = await generateSummary(toSummarize, {
-      model: options.model,
-      customPrompt: options.summaryPrompt
-    });
-    summarized = true;
-    const summaryMessage = {
-      id: crypto.randomUUID(),
-      role: "system",
-      content: `## Previous Conversation Summary
-
-${summary}`,
-      createdAt: /* @__PURE__ */ new Date()
-    };
-    currentMessages = [summaryMessage, ...toKeep];
-  } else {
-    currentMessages = toKeep;
-  }
-  return { messages: currentMessages, removedCount, tokensRemoved, summarized, summary };
-}
-
-// src/context/manager.ts
-var ContextManager = class {
-  limits;
-  model;
-  summaryPrompt;
-  constructor(options) {
-    this.limits = { ...DEFAULT_CONTEXT_LIMITS, ...options?.limits };
-    this.model = options?.model;
-    this.summaryPrompt = options?.summaryPrompt;
-  }
-  /** Get a copy of the current context limits. */
-  getLimits() {
-    return { ...this.limits };
-  }
-  /** Update context limits (e.g. when switching models). */
-  setLimits(limits) {
-    this.limits = { ...this.limits, ...limits };
-  }
-  /** Set the model used for summarisation. */
-  setModel(model) {
-    this.model = model;
-  }
-  /** Estimate total tokens for a message array. */
-  estimateTokens(messages) {
-    return estimateConversationTokens(messages);
-  }
-  /** Check whether the context is overflowing. */
-  isOverflowing(messages) {
-    const tokens = this.estimateTokens(messages);
-    return isContextOverflowing(tokens, this.limits);
-  }
-  /** Check whether pruning should be triggered. */
-  shouldPrune(messages) {
-    const tokens = this.estimateTokens(messages);
-    return shouldPruneContext(tokens, this.limits);
-  }
-  /** Prune context to fit within limits. */
-  async prune(messages) {
-    return pruneContext(messages, {
-      model: this.model,
-      limits: this.limits,
-      summaryPrompt: this.summaryPrompt
-    });
-  }
-  /**
-   * Get a snapshot of token statistics.
-   *
-   * Useful for dashboards, logging, or deciding whether to prune.
-   */
-  getStats(messages) {
-    const tokens = this.estimateTokens(messages);
-    const limit = this.limits.contextWindow - this.limits.reserveTokens;
-    return {
-      tokens,
-      limit,
-      available: Math.max(0, limit - tokens),
-      utilizationPercent: Math.round(tokens / limit * 100),
-      isOverflowing: isContextOverflowing(tokens, this.limits),
-      shouldPrune: shouldPruneContext(tokens, this.limits)
-    };
-  }
-};
-
-export {
-  estimateTokens,
-  estimateMessageTokens,
-  estimateConversationTokens,
-  PRUNE_PROTECTED_TOOLS,
-  DEFAULT_CONTEXT_LIMITS,
-  isContextOverflowing,
-  shouldPruneContext,
-  findCutPoint,
-  pruneToolResults,
-  generateSummary,
-  pruneContext,
-  ContextManager
-};

package/dist/context/index.d.ts

@@ -1,259 +0,0 @@
-import { M as Message } from '../messages-BYWGn8TY.js';
-import { ModelMessage, LanguageModel } from 'ai';
-
-/**
- * Token Estimation Utilities
- *
- * Provides lightweight heuristic-based token counting for messages
- * and conversations. Uses the chars/4 approximation — simple but
- * effective for context-window planning decisions.
- *
- * These are *estimates*, not exact counts. For billing or hard limits,
- * use the provider's native tokeniser instead.
- */
-
-/**
- * Estimate token count for a plain string.
- *
- * Uses the widely-accepted `chars / 4` heuristic.
- *
- * @param text - Text to estimate
- * @returns Estimated token count (always ≥ 1 for non-empty input)
- */
-declare function estimateTokens(text: string): number;
-/**
- * Estimate token count for a single message.
- *
- * Handles:
- * - Plain string content
- * - Multi-part / multimodal arrays (text parts + images)
- *
- * @param message - A `Message` (internal) or `ModelMessage` (AI SDK)
- * @returns Estimated token count
- */
-declare function estimateMessageTokens(message: Message | ModelMessage): number;
-/**
- * Estimate total tokens for an entire conversation.
- *
- * Adds a small per-message overhead (≈ 4 tokens) for message
- * framing (`role`, delimiters, etc.).
- *
- * @param messages - Array of messages to estimate
- * @returns Estimated total token count
- */
-declare function estimateConversationTokens(messages: (Message | ModelMessage)[]): number;
-
-/**
- * Context Overflow Detection & Pruning
- *
- * Detects when a conversation is approaching the model's context-window
- * limit and prunes old / redundant content to stay within bounds.
- *
- * Two pruning strategies are available:
- * 1. **Tool-result pruning** — replaces large, stale tool outputs with
- *    compact placeholders (lightweight, no model call).
- * 2. **Conversation cutting** — identifies a safe cut-point and removes
- *    older messages entirely (optionally summarised by the manager).
- */
-
-/**
- * Context limits configuration.
- *
- * All values are in *estimated* tokens (see {@link estimateTokens}).
- */
-interface ContextLimits {
-    /** Maximum context window size in tokens */
-    contextWindow: number;
-    /** Reserve tokens for output generation */
-    reserveTokens: number;
-    /** Protect this many recent tokens from pruning */
-    protectedTokens: number;
-    /** Minimum tokens to trigger pruning (avoid pruning tiny contexts) */
-    pruneMinimum: number;
-}
-/**
- * Default context limits.
- * Based on typical 128 k context-window models.
- */
-declare const DEFAULT_CONTEXT_LIMITS: ContextLimits;
-/**
- * Check whether the context is overflowing.
- *
- * @param tokens - Current estimated token count
- * @param limits - Context limits (defaults to 128 k window)
- * @returns `true` if context exceeds the safe threshold
- */
-declare function isContextOverflowing(tokens: number, limits?: ContextLimits): boolean;
-/**
- * Check whether pruning should be triggered.
- *
- * Returns `false` when the conversation is still small enough that
- * pruning would be wasteful (below {@link ContextLimits.pruneMinimum}).
- */
-declare function shouldPruneContext(tokens: number, limits?: ContextLimits): boolean;
-/**
- * Result of a pruning operation.
- */
-interface PruneResult {
-    /** Messages after pruning (may include a summary message) */
-    messages: Message[];
-    /** Number of messages removed */
-    removedCount: number;
-    /** Estimated tokens removed */
-    tokensRemoved: number;
-    /** Whether summarisation was used */
-    summarized: boolean;
-    /** The summary content, if generated */
-    summary?: string;
-}
-/**
- * Find a safe index at which the conversation can be "cut".
- *
- * Rules:
- * - Never cut in the middle of a tool-call sequence
- * - Prefer cutting after assistant / user messages
- * - Keep at least {@link protectedTokens} tokens at the end
- *
- * @param messages       - Full message array
- * @param protectedTokens - Tokens to preserve at the end
- * @returns Cut index (exclusive — remove messages *before* this index).
- *          Returns `0` when no safe cut-point exists.
- */
-declare function findCutPoint(messages: Message[], protectedTokens?: number): number;
-/**
- * Prune old, large tool results from the conversation.
- *
- * Prune strategy:
- * - Walks backwards through messages
- * - Protects recent outputs (within {@link protectedTokens})
- * - Skips protected tools (e.g. "skill")
- * - Stamps pruned messages with a `compactedAt` timestamp
- *
- * @param messages       - Messages to prune
- * @param protectedTokens - Don't prune tool results in the last N tokens
- * @param options        - Extra tool names to protect
- * @returns New message array with large tool outputs replaced
- */
-declare function pruneToolResults(messages: Message[], protectedTokens?: number, options?: {
-    /** Additional tools to protect from pruning */
-    protectedTools?: string[];
-}): Message[];
-
-/**
- * Context Summarisation
- *
- * When tool-result pruning alone isn't enough to stay within the
- * context window, the manager can *cut* the conversation and replace
- * the removed portion with an LLM-generated summary.
- *
- * This module provides the summary-generation logic and the
- * top-level `pruneContext()` orchestration function.
- */
-
-/**
- * Options for summary generation.
- */
-interface SummarizationOptions {
-    /** Model used to generate the summary */
-    model: LanguageModel;
-    /** Max tokens for the summary output */
-    maxTokens?: number;
-    /** Custom summarisation system prompt */
-    customPrompt?: string;
-}
-/**
- * Options for {@link pruneContext}.
- */
-interface PruneContextOptions {
-    /** Model for summarisation (omit to skip summary generation) */
-    model?: LanguageModel;
-    /** Context limits to enforce */
-    limits?: ContextLimits;
-    /** Custom summarisation prompt */
-    summaryPrompt?: string;
-}
-/**
- * Generate a natural-language summary of a message sequence.
- *
- * @param messages - Messages to summarise
- * @param options  - Model & prompt configuration
- * @returns Summary text
- */
-declare function generateSummary(messages: Message[], options: SummarizationOptions): Promise<string>;
-/**
- * Prune a conversation to fit within context-window limits.
- *
- * Strategy (in order):
- * 1. Prune old tool outputs (lightweight, no model call)
- * 2. If still overflowing, find a safe cut-point and remove
- *    older messages — optionally generating an LLM summary.
- *
- * @param messages - Current message array
- * @param options  - Limits, model, and prompt overrides
- * @returns A {@link PruneResult} with the trimmed messages
- */
-declare function pruneContext(messages: Message[], options?: PruneContextOptions): Promise<PruneResult>;
-
-/**
- * Context Manager
- *
- * Stateful wrapper around the context-management primitives.
- * Tracks per-session context limits and provides a single entry
- * point for token estimation, overflow detection, and pruning.
- */
-
-/**
- * Per-session context manager.
- *
- * Holds the active context-window limits and provides convenience
- * methods for checking, reporting, and pruning context.
- *
- * @example
- * ```typescript
- * const ctx = new ContextManager({ limits: { contextWindow: 200_000 } });
- *
- * if (ctx.shouldPrune(messages)) {
- *   const result = await ctx.prune(messages);
- *   console.log(`Removed ${result.removedCount} messages`);
- * }
- * ```
- */
-declare class ContextManager {
-    private limits;
-    private model?;
-    private summaryPrompt?;
-    constructor(options?: {
-        limits?: Partial<ContextLimits>;
-        model?: LanguageModel;
-        summaryPrompt?: string;
-    });
-    /** Get a copy of the current context limits. */
-    getLimits(): ContextLimits;
-    /** Update context limits (e.g. when switching models). */
-    setLimits(limits: Partial<ContextLimits>): void;
-    /** Set the model used for summarisation. */
-    setModel(model: LanguageModel): void;
-    /** Estimate total tokens for a message array. */
-    estimateTokens(messages: (Message | ModelMessage)[]): number;
-    /** Check whether the context is overflowing. */
-    isOverflowing(messages: (Message | ModelMessage)[]): boolean;
-    /** Check whether pruning should be triggered. */
-    shouldPrune(messages: (Message | ModelMessage)[]): boolean;
-    /** Prune context to fit within limits. */
-    prune(messages: Message[]): Promise<PruneResult>;
-    /**
-     * Get a snapshot of token statistics.
-     *
-     * Useful for dashboards, logging, or deciding whether to prune.
-     */
-    getStats(messages: (Message | ModelMessage)[]): {
-        tokens: number;
-        limit: number;
-        available: number;
-        utilizationPercent: number;
-        isOverflowing: boolean;
-        shouldPrune: boolean;
-    };
-}
-
-export { type ContextLimits, ContextManager, DEFAULT_CONTEXT_LIMITS, type PruneContextOptions, type PruneResult, type SummarizationOptions, estimateConversationTokens, estimateMessageTokens, estimateTokens, findCutPoint, generateSummary, isContextOverflowing, pruneContext, pruneToolResults, shouldPruneContext };

package/dist/context/index.js

@@ -1,26 +0,0 @@
-import {
-  ContextManager,
-  DEFAULT_CONTEXT_LIMITS,
-  estimateConversationTokens,
-  estimateMessageTokens,
-  estimateTokens,
-  findCutPoint,
-  generateSummary,
-  isContextOverflowing,
-  pruneContext,
-  pruneToolResults,
-  shouldPruneContext
-} from "../chunk-WWYYNWEW.js";
-export {
-  ContextManager,
-  DEFAULT_CONTEXT_LIMITS,
-  estimateConversationTokens,
-  estimateMessageTokens,
-  estimateTokens,
-  findCutPoint,
-  generateSummary,
-  isContextOverflowing,
-  pruneContext,
-  pruneToolResults,
-  shouldPruneContext
-};

package/dist/events-CE72w8W4.d.ts

@@ -1,149 +0,0 @@
-import * as ai from 'ai';
-import { T as TokenUsage, M as Message } from './messages-BYWGn8TY.js';
-
-/** Agent status for UI display */
-type AgentStatus = "idle" | "processing" | "thinking" | "reasoning" | "calling-tool" | "waiting-approval" | "error";
-/** Approval request for UI */
-interface ApprovalEvent {
-    id: string;
-    tool: string;
-    args: unknown;
-    description: string;
-    risk: "safe" | "moderate" | "dangerous";
-}
-/** Neutral turn-commit boundaries for runtime/durability integrations */
-type AgentTurnBoundaryKind = "input-commit-start" | "input-commit-finish" | "intervention-commit-start" | "intervention-commit-finish" | "step-commit-start" | "step-commit-finish" | "output-commit-start" | "output-commit-finish";
-/**
- * Events emitted during agent execution
- *
- * These events are designed for UI consumption:
- * - status: Overall agent state for status indicators
- * - approval-request: User confirmation needed
- * - progress: Step counts for progress bars
- */
-type AgentEvent = {
-    type: "status";
-    status: AgentStatus;
-} | {
-    type: "approval-request";
-    request: ApprovalEvent;
-} | {
-    type: "approval-resolved";
-    id: string;
-    action: "allow" | "deny" | "remember";
-} | {
-    type: "step-start";
-    step: number;
-    maxSteps: number;
-} | {
-    type: "step-finish";
-    step: number;
-    usage?: TokenUsage;
-    finishReason?: string;
-} | {
-    type: "turn-boundary";
-    boundary: AgentTurnBoundaryKind;
-    step?: number;
-    messageRole?: Message["role"];
-    pendingToolCallCount?: number;
-} | {
-    type: "message";
-    message: Message;
-} | {
-    type: "text-start";
-} | {
-    type: "text-delta";
-    text: string;
-} | {
-    type: "text-end";
-} | {
-    type: "reasoning-start";
-    id: string;
-} | {
-    type: "reasoning-delta";
-    id: string;
-    text: string;
-} | {
-    type: "reasoning-end";
-    id: string;
-} | {
-    type: "tool-start";
-    toolName: string;
-    toolCallId: string;
-    input: unknown;
-} | {
-    type: "tool-result";
-    toolName: string;
-    toolCallId: string;
-    result: unknown;
-} | {
-    type: "tool-error";
-    toolName: string;
-    toolCallId: string;
-    error: string;
-} | {
-    type: "computer-call";
-    callId: string;
-    action: unknown;
-    pendingSafetyChecks?: unknown[];
-} | {
-    type: "computer-result";
-    callId: string;
-    result: unknown;
-} | {
-    type: "intervention-applied";
-    id: string;
-    message: string;
-} | {
-    type: "doom-loop";
-    toolName: string;
-    repeatCount: number;
-} | {
-    type: "context-overflow";
-    inputTokens: number;
-    limit: number;
-} | {
-    type: "turn-summary";
-    turnId: string;
-    files: Array<{
-        path: string;
-        type: "created" | "modified" | "deleted" | "unchanged";
-        additions: number;
-        deletions: number;
-    }>;
-    additions: number;
-    deletions: number;
-} | {
-    type: "retry";
-    attempt: number;
-    delayMs: number;
-    error: Error;
-} | {
-    type: "error";
-    error: Error;
-} | {
-    type: "complete";
-    usage?: TokenUsage;
-    output?: string;
-};
-/**
- * Step-processing result - what happens after reducing one streamed step
- */
-type StepProcessingResult = "continue" | "stop" | "compact";
-/**
- * @deprecated Use `StepProcessingResult`.
- */
-type ProcessorResult = StepProcessingResult;
-/**
- * Stream input for model inference
- */
-interface StreamInput {
-    sessionID: string;
-    model: ai.LanguageModel;
-    system: string[];
-    messages: ai.ModelMessage[];
-    abort: AbortSignal;
-    tools: Record<string, unknown>;
-}
-
-export type { AgentEvent as A, ProcessorResult as P, StepProcessingResult as S, AgentStatus as a, AgentTurnBoundaryKind as b, ApprovalEvent as c, StreamInput as d };