@johpaz/hive-agents 0.0.35 → 0.0.37

package/packages/core/src/agent/compaction.ts

@@ -0,0 +1,240 @@
+/**
+ * Compaction — Fase 6.
+ *
+ * Compresses conversation history when token count exceeds threshold.
+ * Uses the active LLM to summarize old messages, preserving:
+ *   - User data and preferences
+ *   - Decisions made
+ *   - Tool results
+ *   - Context needed to continue
+ *
+ * Saves summary to `summaries` table. Original messages are kept (audit trail)
+ * but the Context Compiler uses the summary instead of old messages.
+ *
+ * Also implements "tool result clearing": replaces old tool results with
+ * short summaries in the in-memory message array before model calls.
+ */
+
+import { logger } from "../utils/logger"
+import {
+  getTotalTokens,
+  getHistory,
+  getSummary,
+  saveSummary,
+  toAPIMessages,
+  getMessageCount,
+} from "./conversation-store"
+import { estimateTokens } from "../utils/toon"
+import { callLLM, resolveProviderConfig, type ContentPart } from "./llm-client"
+import { getDb } from "../storage/sqlite"
+
+const log = logger.child("compaction")
+
+// Token budget: compress when stored tokens exceed this threshold
+// Will be overridden by model's context_window at runtime if available
+const COMPACT_TOKEN_THRESHOLD = 32000  // ~25% of 128K default context window
+const KEEP_LAST_N_MESSAGES = 5         // always keep most recent N messages
+const TOOL_RESULT_MAX_CHARS = 200      // max chars for old tool results after clearing
+const MAX_TRANSCRIPT_MSGS = 30         // cap messages sent to summarizer (avoids OOM on small models)
+const MAX_MSG_CHARS = 300              // chars per message in transcript
+
+/**
+ * Check if compaction is needed and run it if so.
+ * Called at the start of each agent loop iteration.
+ */
+export async function maybeCompact(
+  threadId: string,
+  notify?: { channel: string; userId: string }
+): Promise<void> {
+  try {
+    const totalTokens = getTotalTokens(threadId)
+
+    // Use model's context window if available, otherwise use default
+    const db = getDb()
+    let effectiveThreshold = COMPACT_TOKEN_THRESHOLD
+    try {
+      const coordinator = db.query<any, []>(
+        "SELECT a.model_id FROM agents a WHERE a.role = 'coordinator' LIMIT 1"
+      ).get()
+      if (coordinator?.model_id) {
+        const modelRow = db.query<any, [string]>(
+          "SELECT context_window FROM models WHERE id = ?"
+        ).get(coordinator.model_id.replace(/^[^/]+\//, ''))
+        if (modelRow?.context_window) {
+          effectiveThreshold = Math.floor(modelRow.context_window * 0.25)
+        }
+      }
+    } catch { /* use default threshold */ }
+
+    if (totalTokens < effectiveThreshold) return
+
+    const summary = getSummary(threadId)
+    const totalMessages = getMessageCount(threadId)
+
+    // Already summarized up to near the current state
+    if (summary && summary.last_message_id > totalMessages - KEEP_LAST_N_MESSAGES) return
+
+    log.info(`[compaction] Compacting thread=${threadId} tokens=${totalTokens}`)
+    await compactThread(threadId, notify)
+  } catch (err) {
+    log.warn("[compaction] Error during compaction check:", err)
+  }
+}
+
+/**
+ * Compress a thread's history into a summary.
+ */
+export async function compactThread(
+  threadId: string,
+  notify?: { channel: string; userId: string }
+): Promise<void> {
+  const allMessages = getHistory(threadId)
+  if (allMessages.length <= KEEP_LAST_N_MESSAGES) return
+
+  // Find a clean cut point: the "keep" side must begin with a user turn so
+  // we never leave orphaned tool messages at the start of the visible window.
+  let cutIndex = allMessages.length - KEEP_LAST_N_MESSAGES
+  while (cutIndex > 0 && allMessages[cutIndex]?.role !== "user") {
+    cutIndex--
+  }
+  if (cutIndex <= 0) {
+    log.info(`[compaction] No clean user-turn boundary found — skipping`)
+    return
+  }
+
+  const toSummarize = allMessages.slice(0, cutIndex)
+  if (toSummarize.length === 0) return
+
+  const lastSummarizedId = toSummarize[toSummarize.length - 1].id
+
+  const existingSummary = getSummary(threadId)
+  if (existingSummary && existingSummary.last_message_id >= lastSummarizedId) return
+
+  // Cap transcript to avoid overflowing small model contexts
+  const capped = toSummarize.slice(-MAX_TRANSCRIPT_MSGS)
+  const apiMessages = toAPIMessages(capped)
+  const transcript = apiMessages
+    .map((m) => {
+      const text = typeof m.content === "string"
+        ? m.content
+        : Array.isArray(m.content)
+          ? m.content.filter(p => p.type === "text").map(p => (p as any).text).join("\n")
+          : ""
+      return `[${m.role.toUpperCase()}]: ${text.substring(0, MAX_MSG_CHARS)}`
+    })
+    .join("\n\n")
+
+  const db = getDb()
+  const coordinator = db.query<any, []>(
+    "SELECT provider_id, model_id FROM agents WHERE role = 'coordinator' LIMIT 1"
+  ).get()
+
+  const providerCfg = await resolveProviderConfig(
+    coordinator?.provider_id || "openai",
+    coordinator?.model_id || "gpt-4o-mini"
+  )
+
+  const summaryResponse = await callLLM({
+    ...providerCfg,
+    messages: [
+      {
+        role: "system",
+        content:
+          "You are a conversation summarizer. Create a concise summary preserving: " +
+          "user preferences, decisions made, important facts, tool results, and context needed to continue.",
+      },
+      {
+        role: "user",
+        content: `Summarize this conversation (${toSummarize.length} messages) in 3-5 sentences:\n\n${transcript}`,
+      },
+    ],
+  })
+
+  const summary = summaryResponse.content.trim()
+  if (!summary) return
+
+  saveSummary(threadId, summary, toSummarize.length, lastSummarizedId)
+  log.info(
+    `[compaction] Thread ${threadId} compacted: ${toSummarize.length} msgs → ${estimateTokens(summary)} tokens`
+  )
+
+  // Notify user in their active channel (non-critical)
+  if (notify?.channel && notify?.userId) {
+    try {
+      const { sendToUserChannel } = await import("../gateway/channel-notify")
+      await sendToUserChannel(
+        notify.channel,
+        notify.userId,
+        `🗜️ Resumí ${toSummarize.length} mensajes anteriores para mantener el contexto limpio.`
+      )
+    } catch {
+      // Non-critical — don't break the flow if notification fails
+    }
+  }
+}
+
+/**
+ * Clear old tool results in-memory to reduce tokens before a model call.
+ * Does NOT modify the database — only the in-memory messages array.
+ * 
+ * Strategy: COMPRESS (Context Engineering)
+ * - Replaces old tool results with short summaries
+ * - Keeps recent tool results intact (keepLastN)
+ * - Uses TOON format for compact representation
+ */
+export function clearOldToolResults<T extends { role: string; content: string | ContentPart[] }>(
+  messages: T[],
+  keepLastN = 6
+): T[] {
+  if (messages.length <= keepLastN) return messages
+  const cutoffIndex = messages.length - keepLastN
+
+  return messages.map((msg, i) => {
+    if (i >= cutoffIndex) return msg
+    
+    if (msg.role === "tool" && typeof msg.content === "string") {
+      // For tool results older than keepLastN, summarize
+      if (msg.content.length > TOOL_RESULT_MAX_CHARS) {
+        // Try to extract key info from TOON/JSON format
+        let summary = msg.content.substring(0, TOOL_RESULT_MAX_CHARS)
+        
+        // If it looks like JSON/TOON, add a marker
+        if (msg.content.trim().startsWith('{') || msg.content.trim().includes(':')) {
+          summary = `[Tool result summarized: ${summary}...]`
+        } else {
+          summary = `[Result truncated: ${summary}...]`
+        }
+        
+        return {
+          ...msg,
+          content: summary,
+        }
+      }
+    }
+    
+    return msg
+  })
+}
+
+/**
+ * Summarize a tool result to a single line
+ * Used for very old tool results (> 10 turns)
+ */
+export function summarizeToolResult(content: string, toolName?: string): string {
+  // Try to extract success/failure status
+  const isError = content.includes('error') || content.includes('failed') || content.startsWith('[Tool Error]')
+  const isSuccess = content.includes('ok') || content.includes('success') || content.includes('true')
+  
+  // Try to extract key result field from JSON/TOON
+  let keyInfo = ""
+  try {
+    // Simple extraction of first key value
+    const firstLine = content.split('\n')[0].substring(0, 80)
+    keyInfo = firstLine
+  } catch {
+    keyInfo = content.substring(0, 80)
+  }
+  
+  const status = isError ? "failed" : isSuccess ? "success" : "completed"
+  return `[${toolName || 'Tool'} ${status}: ${keyInfo}...]`
+}