agents 0.13.1 → 0.13.3

package/dist/compaction-helpers-fJyf8j4m.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"compaction-helpers-fJyf8j4m.js","names":[],"sources":["../src/experimental/memory/utils/tokens.ts","../src/experimental/memory/utils/compaction-helpers.ts"],"sourcesContent":["/**\n * Token Estimation Utilities\n *\n * IMPORTANT: These are heuristic estimates, not actual tokenizer counts.\n *\n * We intentionally avoid real tokenizers (e.g. tiktoken, sentencepiece) because:\n * - A single tiktoken instance costs ~80-120MB of heap\n * - Cloudflare Workers have tight memory limits (128MB)\n * - For compaction thresholds, a conservative estimate is sufficient\n *\n * The hybrid approach (max of character-based and word-based estimates) handles\n * both dense token content (JSON, code) and natural language reasonably well.\n *\n * Calibration notes:\n * - Character-based: ~4 chars per token (conservative, from OpenAI guidance)\n * - Word-based: ~1.3 tokens per word (empirical, from Mastra's memory system)\n * - Per-message overhead: ~4 tokens for role/framing (empirical)\n *\n * These ratios are tuned for English. CJK, emoji-heavy, or highly technical\n * content may have different ratios. The conservative estimates help ensure\n * compaction triggers before context windows are actually exceeded.\n */\n\nimport type { SessionMessage } from \"../session/types\";\n\n/** Approximate characters per token for English text */\nexport const CHARS_PER_TOKEN = 4;\n\n/** Approximate token multiplier per whitespace-separated word */\nexport const WORDS_TOKEN_MULTIPLIER = 1.3;\n\n/** Approximate overhead tokens per message (role, framing) */\nexport const TOKENS_PER_MESSAGE = 4;\n\n/**\n * Estimate token count for a string using a hybrid heuristic.\n *\n * Takes the max of two estimates:\n * - Character-based: `length / 4` — better for dense content (JSON, code, URLs)\n * - Word-based: `words * 1.3` — better for natural language prose\n *\n * This is a heuristic. Do not use where exact counts are required.\n */\nexport function estimateStringTokens(text: string): number {\n  if (!text) return 0;\n  const charEstimate = text.length / CHARS_PER_TOKEN;\n  const wordEstimate =\n    text.split(/\\s+/).filter(Boolean).length * WORDS_TOKEN_MULTIPLIER;\n  return Math.ceil(Math.max(charEstimate, wordEstimate));\n}\n\nfunction estimateUnknownTokens(value: unknown): number {\n  if (value === null || value === undefined) return 0;\n  if (typeof value === \"string\") return estimateStringTokens(value);\n\n  try {\n    return estimateStringTokens(JSON.stringify(value));\n  } catch {\n    return estimateStringTokens(String(value));\n  }\n}\n\n/**\n * Estimate total token count for an array of UIMessages.\n *\n * Walks each message's parts (text, reasoning, tool invocations, tool results)\n * and applies per-message overhead.\n *\n * This is a heuristic. Do not use where exact counts are required.\n */\nexport function estimateMessageTokens(messages: SessionMessage[]): number {\n  let tokens = 0;\n  for (const msg of messages) {\n    tokens += TOKENS_PER_MESSAGE;\n    for (const part of msg.parts) {\n      if (part.type === \"text\" || part.type === \"reasoning\") {\n        tokens += estimateUnknownTokens(part.text ?? part.reasoning);\n      } else if (\n        part.type.startsWith(\"tool-\") ||\n        part.type === \"dynamic-tool\"\n      ) {\n        tokens += estimateUnknownTokens(part.input);\n        tokens += estimateUnknownTokens(part.output ?? part.result);\n      } else if (part.text !== undefined) {\n        tokens += estimateUnknownTokens(part.text);\n      } else if (part.result !== undefined) {\n        tokens += estimateUnknownTokens(part.result);\n      }\n    }\n  }\n  return tokens;\n}\n","/**\n * Compaction Helpers\n *\n * Utilities for full compaction (LLM-based summarization).\n * Used by the reference compaction implementation and available\n * for custom CompactFunction implementations.\n */\n\nimport type { SessionMessage } from \"../session/types\";\nimport { estimateMessageTokens } from \"./tokens\";\n\n// ── Compaction ID constants ─────────────────────────────────────────\n\n/** Prefix for all compaction messages (overlays and summaries) */\nexport const COMPACTION_PREFIX = \"compaction_\";\n\n/** Check if a message is a compaction message */\nexport function isCompactionMessage(msg: SessionMessage): boolean {\n  return msg.id.startsWith(COMPACTION_PREFIX);\n}\n\n// ── Tool Pair Alignment ──────────────────────────────────────────────\n\n/**\n * Check if a message contains tool invocations.\n */\nfunction hasToolCalls(msg: SessionMessage): boolean {\n  return msg.parts.some(\n    (p) => p.type.startsWith(\"tool-\") || p.type === \"dynamic-tool\"\n  );\n}\n\n/**\n * Get tool call IDs from a message's parts.\n */\nfunction getToolCallIds(msg: SessionMessage): Set<string> {\n  const ids = new Set<string>();\n  for (const part of msg.parts) {\n    if (\n      (part.type.startsWith(\"tool-\") || part.type === \"dynamic-tool\") &&\n      \"toolCallId\" in part\n    ) {\n      ids.add((part as { toolCallId: string }).toolCallId);\n    }\n  }\n  return ids;\n}\n\n/**\n * Check if a message is a tool result referencing a specific call ID.\n */\nfunction isToolResultFor(msg: SessionMessage, callIds: Set<string>): boolean {\n  return msg.parts.some(\n    (p) =>\n      (p.type.startsWith(\"tool-\") || p.type === \"dynamic-tool\") &&\n      \"toolCallId\" in p &&\n      callIds.has((p as { toolCallId: string }).toolCallId)\n  );\n}\n\n/**\n * Align a boundary index forward to avoid splitting tool call/result groups.\n * If the boundary falls between an assistant message with tool calls and its\n * tool results, move it forward past the results.\n */\nexport function alignBoundaryForward(\n  messages: SessionMessage[],\n  idx: number\n): number {\n  if (idx <= 0 || idx >= messages.length) return idx;\n\n  // Check if the message before the boundary has tool calls\n  const prev = messages[idx - 1];\n  if (prev.role === \"assistant\" && hasToolCalls(prev)) {\n    const callIds = getToolCallIds(prev);\n    // Skip forward past any tool results for these calls\n    while (idx < messages.length && isToolResultFor(messages[idx], callIds)) {\n      idx++;\n    }\n  }\n\n  return idx;\n}\n\n/**\n * Align a boundary index backward to avoid splitting tool call/result groups.\n * If the boundary falls in the middle of tool results, move it backward to\n * include the assistant message that made the calls.\n */\nexport function alignBoundaryBackward(\n  messages: SessionMessage[],\n  idx: number\n): number {\n  if (idx <= 0 || idx >= messages.length) return idx;\n\n  // If the message at idx is a tool result, walk backward to find the call\n  while (idx > 0) {\n    const msg = messages[idx];\n    if (msg.role === \"assistant\" && hasToolCalls(msg)) {\n      break; // This is a tool call message — include it\n    }\n    // Check if this looks like a tool result (assistant message following another)\n    const prev = messages[idx - 1];\n    if (prev.role === \"assistant\" && hasToolCalls(prev)) {\n      const callIds = getToolCallIds(prev);\n      if (isToolResultFor(msg, callIds)) {\n        idx--; // Move back to include the call\n        continue;\n      }\n    }\n    break;\n  }\n\n  return idx;\n}\n\n// ── Token-Budget Tail Protection ─────────────────────────────────────\n\n/**\n * Find the compression end boundary using a token budget for the tail.\n * Walks backward from the end, accumulating tokens until budget is reached.\n * Returns the index where compression should stop (everything from this\n * index onward is protected).\n *\n * @param messages All messages\n * @param headEnd Index where the protected head ends (compression starts here)\n * @param tailTokenBudget Maximum tokens to keep in the tail\n * @param minTailMessages Minimum messages to protect in the tail (fallback)\n */\nexport function findTailCutByTokens(\n  messages: SessionMessage[],\n  headEnd: number,\n  tailTokenBudget = 20000,\n  minTailMessages = 2\n): number {\n  const n = messages.length;\n  let accumulated = 0;\n  let tokenCut = n;\n\n  for (let i = n - 1; i >= headEnd; i--) {\n    const msgTokens = estimateMessageTokens([messages[i]]);\n\n    if (accumulated + msgTokens > tailTokenBudget && tokenCut < n) {\n      // Budget exceeded and we already have at least one tail message\n      break;\n    }\n    accumulated += msgTokens;\n    tokenCut = i;\n  }\n\n  // Protect whichever is larger: token-based tail or minTailMessages\n  const minCut = n - minTailMessages;\n  const cutIdx = minCut >= headEnd ? Math.min(tokenCut, minCut) : tokenCut;\n\n  // Align to avoid splitting tool groups\n  return alignBoundaryBackward(messages, cutIdx);\n}\n\n// ── Tool Pair Sanitization ───────────────────────────────────────────\n\n/**\n * Fix orphaned tool call/result pairs after compaction.\n *\n * Two failure modes:\n * 1. Tool result references a call_id whose assistant tool_call was removed\n *    → Remove the orphaned result\n * 2. Assistant has tool_calls whose results were dropped\n *    → Add stub results so the API doesn't error\n *\n * @param messages Messages after compaction\n * @returns Sanitized messages with no orphaned pairs\n */\nexport function sanitizeToolPairs(\n  messages: SessionMessage[]\n): SessionMessage[] {\n  // Build set of surviving tool call IDs (from assistant messages)\n  const survivingCallIds = new Set<string>();\n  for (const msg of messages) {\n    if (msg.role === \"assistant\") {\n      for (const id of getToolCallIds(msg)) {\n        survivingCallIds.add(id);\n      }\n    }\n  }\n\n  // Build set of tool result IDs\n  const resultCallIds = new Set<string>();\n  for (const msg of messages) {\n    for (const part of msg.parts) {\n      if (\n        (part.type.startsWith(\"tool-\") || part.type === \"dynamic-tool\") &&\n        \"toolCallId\" in part &&\n        \"output\" in part\n      ) {\n        resultCallIds.add((part as { toolCallId: string }).toolCallId);\n      }\n    }\n  }\n\n  // Remove orphaned results (results whose calls were dropped)\n  const orphanedResults = new Set<string>();\n  for (const id of resultCallIds) {\n    if (!survivingCallIds.has(id)) {\n      orphanedResults.add(id);\n    }\n  }\n\n  let result = messages;\n  if (orphanedResults.size > 0) {\n    result = result.map((msg) => {\n      const filteredParts = msg.parts.filter((part) => {\n        if (\n          (part.type.startsWith(\"tool-\") || part.type === \"dynamic-tool\") &&\n          \"toolCallId\" in part &&\n          \"output\" in part\n        ) {\n          return !orphanedResults.has(\n            (part as { toolCallId: string }).toolCallId\n          );\n        }\n        return true;\n      });\n      if (filteredParts.length !== msg.parts.length) {\n        return { ...msg, parts: filteredParts } as SessionMessage;\n      }\n      return msg;\n    });\n  }\n\n  // Add stub results for calls whose results were dropped\n  const missingResults = new Set<string>();\n  for (const id of survivingCallIds) {\n    if (!resultCallIds.has(id) && !orphanedResults.has(id)) {\n      missingResults.add(id);\n    }\n  }\n\n  if (missingResults.size > 0) {\n    const patched: SessionMessage[] = [];\n    for (const msg of result) {\n      patched.push(msg);\n      if (msg.role === \"assistant\") {\n        for (const id of getToolCallIds(msg)) {\n          if (missingResults.has(id)) {\n            // Find the tool name from the call\n            const callPart = msg.parts.find(\n              (p) =>\n                \"toolCallId\" in p &&\n                (p as { toolCallId: string }).toolCallId === id\n            ) as { toolName?: string } | undefined;\n\n            patched.push({\n              id: `stub-${id}`,\n              role: \"assistant\",\n              parts: [\n                {\n                  type: \"tool-result\" as const,\n                  toolCallId: id,\n                  toolName: callPart?.toolName ?? \"unknown\",\n                  result:\n                    \"[Result from earlier conversation — see context summary above]\"\n                } as unknown as SessionMessage[\"parts\"][number]\n              ],\n              createdAt: new Date()\n            } as SessionMessage);\n          }\n        }\n      }\n    }\n    result = patched;\n  }\n\n  // Remove empty messages (all parts filtered out)\n  return result.filter((msg) => msg.parts.length > 0);\n}\n\n// ── Summary Budget ───────────────────────────────────────────────────\n\n/**\n * Compute a summary token budget based on the content being compressed.\n * 20% of the compressed content, clamped to 2K-8K tokens.\n */\nexport function computeSummaryBudget(messages: SessionMessage[]): number {\n  const contentTokens = estimateMessageTokens(messages);\n  // Summary is ~20% of the content being compressed.\n  // The summary replaces the compressed middle, so it's sized relative\n  // to what it's replacing — not the tail budget (they occupy different\n  // slots in the context window).\n  const budget = Math.floor(contentTokens * 0.2);\n  return Math.max(100, budget);\n}\n\n// ── Structured Summary Prompt ────────────────────────────────────────\n\n/**\n * Build a prompt for LLM summarization of compressed messages.\n *\n * @param messages Messages to summarize\n * @param previousSummary Previous summary for iterative updates (or null for first compaction)\n * @param budget Target token count for the summary\n */\nexport function buildSummaryPrompt(\n  messages: SessionMessage[],\n  previousSummary: string | null,\n  budget: number\n): string {\n  const content = messages\n    .map((msg) => {\n      const textParts = msg.parts\n        .filter((p) => p.type === \"text\")\n        .map((p) => (p as { text: string }).text)\n        .join(\"\\n\");\n\n      const toolParts = msg.parts\n        .filter((p) => p.type.startsWith(\"tool-\") || p.type === \"dynamic-tool\")\n        .map((p) => {\n          const tp = p as {\n            toolName?: string;\n            input?: unknown;\n            output?: unknown;\n          };\n          const parts = [`[Tool: ${tp.toolName ?? \"unknown\"}]`];\n          if (tp.input)\n            parts.push(`Input: ${JSON.stringify(tp.input).slice(0, 500)}`);\n          if (tp.output)\n            parts.push(`Output: ${String(tp.output).slice(0, 500)}`);\n          return parts.join(\"\\n\");\n        })\n        .join(\"\\n\");\n\n      return `[${msg.role}]\\n${textParts}${toolParts ? \"\\n\" + toolParts : \"\"}`;\n    })\n    .join(\"\\n\\n---\\n\\n\");\n\n  if (previousSummary) {\n    return `You are updating a conversation summary. A previous summary exists below. New conversation turns have occurred since then and need to be incorporated.\n\nPREVIOUS SUMMARY:\n${previousSummary}\n\nNEW TURNS TO INCORPORATE:\n${content}\n\nUpdate the summary. PRESERVE existing information that is still relevant. ADD new information. Remove information only if it is clearly obsolete.\n\n## Topic\n[What the conversation is about]\n\n## Key Points\n[Important information, decisions, and conclusions from the conversation]\n\n## Current State\n[Where things stand now — what has been done, what is in progress]\n\n## Open Items\n[Unresolved questions, pending tasks, or next steps discussed]\n\nTarget ~${budget} tokens. Be factual — only include information that was explicitly discussed in the conversation. Do NOT invent file paths, commands, or details that were not mentioned. Write only the summary body.`;\n  }\n\n  return `Create a concise summary of this conversation that preserves the important information for future context.\n\nCONVERSATION TO SUMMARIZE:\n${content}\n\nUse this structure:\n\n## Topic\n[What the conversation is about]\n\n## Key Points\n[Important information, decisions, and conclusions from the conversation]\n\n## Current State\n[Where things stand now — what has been done, what is in progress]\n\n## Open Items\n[Unresolved questions, pending tasks, or next steps discussed]\n\nTarget ~${budget} tokens. Be factual — only include information that was explicitly discussed in the conversation. Do NOT invent file paths, commands, or details that were not mentioned. Write only the summary body.`;\n}\n\n// ── Reference Compaction Implementation ──────────────────────────────\n\n/**\n * Result of a compaction function — describes the overlay to store.\n */\nexport interface CompactResult {\n  /** First message ID in the compacted range */\n  fromMessageId: string;\n  /** Last message ID in the compacted range */\n  toMessageId: string;\n  /** Summary text to store as the overlay */\n  summary: string;\n}\n\nexport interface CompactOptions {\n  /**\n   * Function to call the LLM for summarization.\n   * Takes a user prompt string, returns the LLM's text response.\n   */\n  summarize: (prompt: string) => Promise<string>;\n\n  /** Number of head messages to protect (default: 2) */\n  protectHead?: number;\n\n  /** Token budget for tail protection (default: 20000) */\n  tailTokenBudget?: number;\n\n  /** Minimum tail messages to protect (default: 2) */\n  minTailMessages?: number;\n}\n\n/**\n * Reference compaction implementation.\n *\n * Implements the full hermes-style compaction algorithm:\n * 1. Protect head messages (first N)\n * 2. Protect tail by token budget (walk backward)\n * 3. Align boundaries to tool call groups\n * 4. Summarize middle section with LLM (structured format)\n * 5. Sanitize orphaned tool pairs\n * 6. Iterative summary updates on subsequent compactions\n *\n * @example\n * ```typescript\n * import { createCompactFunction } from \"agents/experimental/memory/utils\";\n *\n * const session = new Session(provider, {\n *   compaction: {\n *     tokenThreshold: 100000,\n *     fn: createCompactFunction({\n *       summarize: (prompt) => generateText({ model, prompt }).then(r => r.text)\n *     })\n *   }\n * });\n * ```\n */\nexport function createCompactFunction(opts: CompactOptions) {\n  const protectHead = opts.protectHead ?? 3;\n  const tailTokenBudget = opts.tailTokenBudget ?? 20000;\n  const minTailMessages = opts.minTailMessages ?? 2;\n\n  return async (messages: SessionMessage[]): Promise<CompactResult | null> => {\n    if (messages.length <= protectHead + minTailMessages) {\n      return null;\n    }\n\n    // 1. Find compression boundaries\n    let compressStart = protectHead;\n    compressStart = alignBoundaryForward(messages, compressStart);\n\n    let compressEnd = findTailCutByTokens(\n      messages,\n      compressStart,\n      tailTokenBudget,\n      minTailMessages\n    );\n\n    if (compressEnd <= compressStart) {\n      return null;\n    }\n\n    // Filter out compaction overlay messages — they have virtual IDs\n    // and should not be included in the summary prompt or used as range IDs\n    const middleMessages = messages\n      .slice(compressStart, compressEnd)\n      .filter((m) => !isCompactionMessage(m));\n\n    if (middleMessages.length === 0) return null;\n\n    // 2. Generate summary — extract previous summary from compaction overlays\n    const existingCompaction = messages.find(isCompactionMessage);\n    const previousSummary = existingCompaction\n      ? existingCompaction.parts\n          .filter((p) => p.type === \"text\")\n          .map((p) => (p as { text: string }).text)\n          .join(\"\\n\")\n      : null;\n\n    const budget = computeSummaryBudget(middleMessages);\n    const prompt = buildSummaryPrompt(middleMessages, previousSummary, budget);\n    const summary = await opts.summarize(prompt);\n\n    if (!summary.trim()) return null;\n\n    return {\n      fromMessageId: middleMessages[0].id,\n      toMessageId: middleMessages[middleMessages.length - 1].id,\n      summary\n    };\n  };\n}\n"],"mappings":";;AA0BA,MAAa,kBAAkB;;AAG/B,MAAa,yBAAyB;;AAGtC,MAAa,qBAAqB;;;;;;;;;;AAWlC,SAAgB,qBAAqB,MAAsB;CACzD,IAAI,CAAC,MAAM,OAAO;CAClB,MAAM,eAAe,KAAK,SAAA;CAC1B,MAAM,eACJ,KAAK,MAAM,KAAK,EAAE,OAAO,OAAO,EAAE,SAAS;CAC7C,OAAO,KAAK,KAAK,KAAK,IAAI,cAAc,YAAY,CAAC;AACvD;AAEA,SAAS,sBAAsB,OAAwB;CACrD,IAAI,UAAU,QAAQ,UAAU,KAAA,GAAW,OAAO;CAClD,IAAI,OAAO,UAAU,UAAU,OAAO,qBAAqB,KAAK;CAEhE,IAAI;EACF,OAAO,qBAAqB,KAAK,UAAU,KAAK,CAAC;CACnD,QAAQ;EACN,OAAO,qBAAqB,OAAO,KAAK,CAAC;CAC3C;AACF;;;;;;;;;AAUA,SAAgB,sBAAsB,UAAoC;CACxE,IAAI,SAAS;CACb,KAAK,MAAM,OAAO,UAAU;EAC1B,UAAA;EACA,KAAK,MAAM,QAAQ,IAAI,OACrB,IAAI,KAAK,SAAS,UAAU,KAAK,SAAS,aACxC,UAAU,sBAAsB,KAAK,QAAQ,KAAK,SAAS;OACtD,IACL,KAAK,KAAK,WAAW,OAAO,KAC5B,KAAK,SAAS,gBACd;GACA,UAAU,sBAAsB,KAAK,KAAK;GAC1C,UAAU,sBAAsB,KAAK,UAAU,KAAK,MAAM;EAC5D,OAAO,IAAI,KAAK,SAAS,KAAA,GACvB,UAAU,sBAAsB,KAAK,IAAI;OACpC,IAAI,KAAK,WAAW,KAAA,GACzB,UAAU,sBAAsB,KAAK,MAAM;CAGjD;CACA,OAAO;AACT;;;;AC7EA,MAAa,oBAAoB;;AAGjC,SAAgB,oBAAoB,KAA8B;CAChE,OAAO,IAAI,GAAG,WAAW,iBAAiB;AAC5C;;;;AAOA,SAAS,aAAa,KAA8B;CAClD,OAAO,IAAI,MAAM,MACd,MAAM,EAAE,KAAK,WAAW,OAAO,KAAK,EAAE,SAAS,cAClD;AACF;;;;AAKA,SAAS,eAAe,KAAkC;CACxD,MAAM,sBAAM,IAAI,IAAY;CAC5B,KAAK,MAAM,QAAQ,IAAI,OACrB,KACG,KAAK,KAAK,WAAW,OAAO,KAAK,KAAK,SAAS,mBAChD,gBAAgB,MAEhB,IAAI,IAAK,KAAgC,UAAU;CAGvD,OAAO;AACT;;;;AAKA,SAAS,gBAAgB,KAAqB,SAA+B;CAC3E,OAAO,IAAI,MAAM,MACd,OACE,EAAE,KAAK,WAAW,OAAO,KAAK,EAAE,SAAS,mBAC1C,gBAAgB,KAChB,QAAQ,IAAK,EAA6B,UAAU,CACxD;AACF;;;;;;AAOA,SAAgB,qBACd,UACA,KACQ;CACR,IAAI,OAAO,KAAK,OAAO,SAAS,QAAQ,OAAO;CAG/C,MAAM,OAAO,SAAS,MAAM;CAC5B,IAAI,KAAK,SAAS,eAAe,aAAa,IAAI,GAAG;EACnD,MAAM,UAAU,eAAe,IAAI;EAEnC,OAAO,MAAM,SAAS,UAAU,gBAAgB,SAAS,MAAM,OAAO,GACpE;CAEJ;CAEA,OAAO;AACT;;;;;;AAOA,SAAgB,sBACd,UACA,KACQ;CACR,IAAI,OAAO,KAAK,OAAO,SAAS,QAAQ,OAAO;CAG/C,OAAO,MAAM,GAAG;EACd,MAAM,MAAM,SAAS;EACrB,IAAI,IAAI,SAAS,eAAe,aAAa,GAAG,GAC9C;EAGF,MAAM,OAAO,SAAS,MAAM;EAC5B,IAAI,KAAK,SAAS,eAAe,aAAa,IAAI;OAE5C,gBAAgB,KADJ,eAAe,IACA,CAAC,GAAG;IACjC;IACA;GACF;;EAEF;CACF;CAEA,OAAO;AACT;;;;;;;;;;;;AAeA,SAAgB,oBACd,UACA,SACA,kBAAkB,KAClB,kBAAkB,GACV;CACR,MAAM,IAAI,SAAS;CACnB,IAAI,cAAc;CAClB,IAAI,WAAW;CAEf,KAAK,IAAI,IAAI,IAAI,GAAG,KAAK,SAAS,KAAK;EACrC,MAAM,YAAY,sBAAsB,CAAC,SAAS,EAAE,CAAC;EAErD,IAAI,cAAc,YAAY,mBAAmB,WAAW,GAE1D;EAEF,eAAe;EACf,WAAW;CACb;CAGA,MAAM,SAAS,IAAI;CAInB,OAAO,sBAAsB,UAHd,UAAU,UAAU,KAAK,IAAI,UAAU,MAAM,IAAI,QAGnB;AAC/C;;;;;;;;;;;;;AAgBA,SAAgB,kBACd,UACkB;CAElB,MAAM,mCAAmB,IAAI,IAAY;CACzC,KAAK,MAAM,OAAO,UAChB,IAAI,IAAI,SAAS,aACf,KAAK,MAAM,MAAM,eAAe,GAAG,GACjC,iBAAiB,IAAI,EAAE;CAM7B,MAAM,gCAAgB,IAAI,IAAY;CACtC,KAAK,MAAM,OAAO,UAChB,KAAK,MAAM,QAAQ,IAAI,OACrB,KACG,KAAK,KAAK,WAAW,OAAO,KAAK,KAAK,SAAS,mBAChD,gBAAgB,QAChB,YAAY,MAEZ,cAAc,IAAK,KAAgC,UAAU;CAMnE,MAAM,kCAAkB,IAAI,IAAY;CACxC,KAAK,MAAM,MAAM,eACf,IAAI,CAAC,iBAAiB,IAAI,EAAE,GAC1B,gBAAgB,IAAI,EAAE;CAI1B,IAAI,SAAS;CACb,IAAI,gBAAgB,OAAO,GACzB,SAAS,OAAO,KAAK,QAAQ;EAC3B,MAAM,gBAAgB,IAAI,MAAM,QAAQ,SAAS;GAC/C,KACG,KAAK,KAAK,WAAW,OAAO,KAAK,KAAK,SAAS,mBAChD,gBAAgB,QAChB,YAAY,MAEZ,OAAO,CAAC,gBAAgB,IACrB,KAAgC,UACnC;GAEF,OAAO;EACT,CAAC;EACD,IAAI,cAAc,WAAW,IAAI,MAAM,QACrC,OAAO;GAAE,GAAG;GAAK,OAAO;EAAc;EAExC,OAAO;CACT,CAAC;CAIH,MAAM,iCAAiB,IAAI,IAAY;CACvC,KAAK,MAAM,MAAM,kBACf,IAAI,CAAC,cAAc,IAAI,EAAE,KAAK,CAAC,gBAAgB,IAAI,EAAE,GACnD,eAAe,IAAI,EAAE;CAIzB,IAAI,eAAe,OAAO,GAAG;EAC3B,MAAM,UAA4B,CAAC;EACnC,KAAK,MAAM,OAAO,QAAQ;GACxB,QAAQ,KAAK,GAAG;GAChB,IAAI,IAAI,SAAS;SACV,MAAM,MAAM,eAAe,GAAG,GACjC,IAAI,eAAe,IAAI,EAAE,GAAG;KAE1B,MAAM,WAAW,IAAI,MAAM,MACxB,MACC,gBAAgB,KACf,EAA6B,eAAe,EACjD;KAEA,QAAQ,KAAK;MACX,IAAI,QAAQ;MACZ,MAAM;MACN,OAAO,CACL;OACE,MAAM;OACN,YAAY;OACZ,UAAU,UAAU,YAAY;OAChC,QACE;MACJ,CACF;MACA,2BAAW,IAAI,KAAK;KACtB,CAAmB;IACrB;;EAGN;EACA,SAAS;CACX;CAGA,OAAO,OAAO,QAAQ,QAAQ,IAAI,MAAM,SAAS,CAAC;AACpD;;;;;AAQA,SAAgB,qBAAqB,UAAoC;CACvE,MAAM,gBAAgB,sBAAsB,QAAQ;CAKpD,MAAM,SAAS,KAAK,MAAM,gBAAgB,EAAG;CAC7C,OAAO,KAAK,IAAI,KAAK,MAAM;AAC7B;;;;;;;;AAWA,SAAgB,mBACd,UACA,iBACA,QACQ;CACR,MAAM,UAAU,SACb,KAAK,QAAQ;EACZ,MAAM,YAAY,IAAI,MACnB,QAAQ,MAAM,EAAE,SAAS,MAAM,EAC/B,KAAK,MAAO,EAAuB,IAAI,EACvC,KAAK,IAAI;EAEZ,MAAM,YAAY,IAAI,MACnB,QAAQ,MAAM,EAAE,KAAK,WAAW,OAAO,KAAK,EAAE,SAAS,cAAc,EACrE,KAAK,MAAM;GACV,MAAM,KAAK;GAKX,MAAM,QAAQ,CAAC,UAAU,GAAG,YAAY,UAAU,EAAE;GACpD,IAAI,GAAG,OACL,MAAM,KAAK,UAAU,KAAK,UAAU,GAAG,KAAK,EAAE,MAAM,GAAG,GAAG,GAAG;GAC/D,IAAI,GAAG,QACL,MAAM,KAAK,WAAW,OAAO,GAAG,MAAM,EAAE,MAAM,GAAG,GAAG,GAAG;GACzD,OAAO,MAAM,KAAK,IAAI;EACxB,CAAC,EACA,KAAK,IAAI;EAEZ,OAAO,IAAI,IAAI,KAAK,KAAK,YAAY,YAAY,OAAO,YAAY;CACtE,CAAC,EACA,KAAK,aAAa;CAErB,IAAI,iBACF,OAAO;;;EAGT,gBAAgB;;;EAGhB,QAAQ;;;;;;;;;;;;;;;;UAgBA,OAAO;CAGf,OAAO;;;EAGP,QAAQ;;;;;;;;;;;;;;;;UAgBA,OAAO;AACjB;;;;;;;;;;;;;;;;;;;;;;;;;;AA0DA,SAAgB,sBAAsB,MAAsB;CAC1D,MAAM,cAAc,KAAK,eAAe;CACxC,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,kBAAkB,KAAK,mBAAmB;CAEhD,OAAO,OAAO,aAA8D;EAC1E,IAAI,SAAS,UAAU,cAAc,iBACnC,OAAO;EAIT,IAAI,gBAAgB;EACpB,gBAAgB,qBAAqB,UAAU,aAAa;EAE5D,IAAI,cAAc,oBAChB,UACA,eACA,iBACA,eACF;EAEA,IAAI,eAAe,eACjB,OAAO;EAKT,MAAM,iBAAiB,SACpB,MAAM,eAAe,WAAW,EAChC,QAAQ,MAAM,CAAC,oBAAoB,CAAC,CAAC;EAExC,IAAI,eAAe,WAAW,GAAG,OAAO;EAGxC,MAAM,qBAAqB,SAAS,KAAK,mBAAmB;EAS5D,MAAM,SAAS,mBAAmB,gBARV,qBACpB,mBAAmB,MAChB,QAAQ,MAAM,EAAE,SAAS,MAAM,EAC/B,KAAK,MAAO,EAAuB,IAAI,EACvC,KAAK,IAAI,IACZ,MAEW,qBAAqB,cACoC,CAAC;EACzE,MAAM,UAAU,MAAM,KAAK,UAAU,MAAM;EAE3C,IAAI,CAAC,QAAQ,KAAK,GAAG,OAAO;EAE5B,OAAO;GACL,eAAe,eAAe,GAAG;GACjC,aAAa,eAAe,eAAe,SAAS,GAAG;GACvD;EACF;CACF;AACF"}

package/dist/experimental/memory/session/index.d.ts

@@ -1,9 +1,12 @@
 import {
-  A as SessionOptions,
+  A as SessionMessage,
   C as AgentSessionProvider,
   D as StoredCompaction,
   E as SessionProvider,
-  O as SessionMessage,
+  M as SessionOptions,
+  N as SessionTokenCounter,
+  O as CompactAfterOptions,
+  P as SessionTokenCounterInput,
   S as isSearchProvider,
   T as SearchResult,
   _ as R2SkillProvider,
@@ -11,7 +14,8 @@ import {
   f as ContextBlock,
   g as isWritableProvider,
   h as WritableContextProvider,
-  k as SessionMessagePart,
+  j as SessionMessagePart,
+  k as CompactionErrorHandler,
   m as ContextProvider,
   p as ContextConfig,
   r as CompactResult,
@@ -19,7 +23,7 @@ import {
   w as SqlProvider,
   x as SearchProvider,
   y as isSkillProvider
-} from "../../../compaction-helpers-DAe-xiVY.js";
+} from "../../../compaction-helpers-B-pG5J22.js";
 import { ToolSet } from "ai";
 
 //#region src/experimental/memory/session/session.d.ts
@@ -56,6 +60,8 @@ declare class Session {
   private _cachedPrompt?;
   private _compactionFn?;
   private _tokenThreshold?;
+  private _tokenCounter?;
+  private _compactionErrorHandler?;
   private _ready;
   private _restorePromise?;
   private _messageChangeListener?;
@@ -104,8 +110,19 @@ declare class Session {
   /**
    * Auto-compact when estimated token count exceeds the threshold.
    * Checked after each `appendMessage`. Requires `onCompaction()`.
+   *
+   * By default this uses a Workers-safe heuristic over stored messages plus
+   * the Session-managed frozen system prompt. Provide `tokenCounter` when you
+   * have model-reported usage or a tokenizer and need a stricter budget.
+   */
+  compactAfter(tokenThreshold: number, options?: CompactAfterOptions): this;
+  /**
+   * Handle failures from the automatic `compactAfter()` trigger.
+   *
+   * Manual `compact()` still reports errors through the existing session error
+   * broadcast path.
    */
-  compactAfter(tokenThreshold: number): this;
+  onCompactionError(handler: CompactionErrorHandler): this;
   /**
    * @internal
    * Framework hook for cache-owning callers that need to mirror message
@@ -153,6 +170,9 @@ declare class Session {
   getBranches(messageId: string): Promise<SessionMessage[]>;
   getPathLength(leafId?: string | null): Promise<number>;
   private _broadcast;
+  private _shouldEstimateTokens;
+  private _estimateTokenCount;
+  private _handleAutoCompactionError;
   private _emitStatus;
   private _emitError;
   appendMessage(
@@ -263,6 +283,8 @@ declare class SessionManager {
   private _cachedPrompt?;
   private _compactionFn?;
   private _tokenThreshold?;
+  private _tokenCounter?;
+  private _compactionErrorHandler?;
   private _sessions;
   private _historyLabel?;
   private _tableReady;
@@ -299,7 +321,11 @@ declare class SessionManager {
    * Auto-compact when estimated token count exceeds the threshold.
    * Propagated to all sessions. Requires `onCompaction()`.
    */
-  compactAfter(tokenThreshold: number): this;
+  compactAfter(tokenThreshold: number, options?: CompactAfterOptions): this;
+  /**
+   * Handle failures from automatic compaction in managed sessions.
+   */
+  onCompactionError(handler: CompactionErrorHandler): this;
   /**
    * Add a searchable context block that searches conversation history
    * across all sessions managed by this manager.
@@ -540,6 +566,8 @@ export {
   AgentContextProvider,
   AgentSearchProvider,
   AgentSessionProvider,
+  type CompactAfterOptions,
+  type CompactionErrorHandler,
   type ContextBlock,
   type ContextConfig,
   type ContextProvider,
@@ -561,6 +589,8 @@ export {
   type SessionMessagePart,
   type SessionOptions,
   type SessionProvider,
+  type SessionTokenCounter,
+  type SessionTokenCounterInput,
   type SkillProvider,
   type SqlProvider,
   type StoredCompaction,

package/dist/experimental/memory/session/index.js

@@ -1,5 +1,5 @@
 import "../../../types.js";
-import { m as estimateStringTokens, p as estimateMessageTokens, t as COMPACTION_PREFIX } from "../../../compaction-helpers-DvcZnvQ1.js";
+import { m as estimateStringTokens, p as estimateMessageTokens, t as COMPACTION_PREFIX } from "../../../compaction-helpers-fJyf8j4m.js";
 import { z } from "zod";
 //#region src/experimental/memory/session/search.ts
 /**
@@ -446,7 +446,7 @@ var ContextBlocks = class {
 	refreshSnapshot() {
 		return this.captureSnapshot();
 	}
-	captureSnapshot() {
+	renderPrompt() {
 		const parts = [];
 		const sep = "═".repeat(46);
 		for (const block of this.blocks.values()) {
@@ -463,7 +463,10 @@ var ContextBlocks = class {
 			else header += " [writable]";
 			parts.push(`${sep}\n${header}\n${sep}\n${block.content}`);
 		}
-		this.snapshot = parts.join("\n\n");
+		return parts.join("\n\n");
+	}
+	captureSnapshot() {
+		this.snapshot = this.renderPrompt();
 		return this.snapshot;
 	}
 	/**
@@ -512,6 +515,23 @@ var ContextBlocks = class {
 		return prompt;
 	}
 	/**
+	* Return the prompt text used for token estimation without persisting a new
+	* frozen prompt to the prompt store.
+	*
+	* This still reads an existing cached prompt when present, so estimates match
+	* the prompt that inference would reuse. If no cached prompt exists, it loads
+	* providers and renders the current blocks without freezing the snapshot.
+	*/
+	async getSystemPromptForEstimate() {
+		if (this.snapshot !== null) return this.snapshot;
+		if (this.promptStore) {
+			const stored = await this.promptStore.get();
+			if (stored !== null) return stored;
+		}
+		if (!this.loaded) await this.load();
+		return this.renderPrompt();
+	}
+	/**
 	* Force reload blocks from providers, re-render the system prompt,
 	* and persist to the store. Use this after block content has changed
 	* or to invalidate the cached prompt.
@@ -1004,6 +1024,8 @@ var Session = class Session {
 		this._ready = false;
 		this.storage = storage;
 		this.context = new ContextBlocks(options?.context ?? [], options?.promptStore);
+		this._tokenCounter = options?.tokenCounter;
+		this._compactionErrorHandler = options?.onCompactionError;
 		this._ready = true;
 	}
 	/**
@@ -1072,9 +1094,24 @@ var Session = class Session {
 	/**
 	* Auto-compact when estimated token count exceeds the threshold.
 	* Checked after each `appendMessage`. Requires `onCompaction()`.
+	*
+	* By default this uses a Workers-safe heuristic over stored messages plus
+	* the Session-managed frozen system prompt. Provide `tokenCounter` when you
+	* have model-reported usage or a tokenizer and need a stricter budget.
 	*/
-	compactAfter(tokenThreshold) {
+	compactAfter(tokenThreshold, options) {
 		this._tokenThreshold = tokenThreshold;
+		if (options?.tokenCounter) this._tokenCounter = options.tokenCounter;
+		return this;
+	}
+	/**
+	* Handle failures from the automatic `compactAfter()` trigger.
+	*
+	* Manual `compact()` still reports errors through the existing session error
+	* broadcast path.
+	*/
+	onCompactionError(handler) {
+		this._compactionErrorHandler = handler;
 		return this;
 	}
 	/**
@@ -1223,8 +1260,42 @@ var Session = class Session {
 			...data
 		}));
 	}
+	_shouldEstimateTokens() {
+		return Boolean(this._broadcaster || this._tokenThreshold != null && this._compactionFn);
+	}
+	async _estimateTokenCount() {
+		const messages = await this.getHistory();
+		const systemPrompt = await this.context.getSystemPromptForEstimate();
+		if (this._tokenCounter) {
+			if (!this.context.isLoaded()) await this.context.load();
+			const contextBlocks = this.context.getBlocks();
+			const estimate = await this._tokenCounter({
+				messages,
+				systemPrompt,
+				contextBlocks
+			});
+			return Number.isFinite(estimate) ? Math.max(0, Math.ceil(estimate)) : 0;
+		}
+		return estimateMessageTokens(messages) + estimateStringTokens(systemPrompt);
+	}
+	async _handleAutoCompactionError(error) {
+		const message = error instanceof Error ? error.message : String(error);
+		if (this._compactionErrorHandler) try {
+			await this._compactionErrorHandler(error);
+		} catch (handlerError) {
+			const handlerMessage = handlerError instanceof Error ? handlerError.message : String(handlerError);
+			console.warn(`Session auto-compaction error handler failed: ${handlerMessage}`);
+		}
+		else console.warn(`Session auto-compaction failed: ${message}`);
+		this._emitError(message);
+	}
 	async _emitStatus(phase, extra) {
-		const tokenEstimate = estimateMessageTokens(await this.getHistory());
+		let tokenEstimate = 0;
+		if (this._shouldEstimateTokens()) try {
+			tokenEstimate = await this._estimateTokenCount();
+		} catch (err) {
+			await this._handleAutoCompactionError(err);
+		}
 		this._broadcast("cf_agent_session", {
 			phase,
 			tokenEstimate,
@@ -1256,7 +1327,9 @@ var Session = class Session {
 		let compacted = false;
 		if (this._tokenThreshold != null && this._compactionFn && tokenEstimate > this._tokenThreshold) try {
 			compacted = Boolean(await this.compact());
-		} catch {}
+		} catch (err) {
+			await this._handleAutoCompactionError(err);
+		}
 		if (!compacted) await this._notifyMessagesChanged({
 			type: "append",
 			message,
@@ -1468,6 +1541,8 @@ var SessionManager = class SessionManager {
 		mgr._pending = [];
 		mgr._compactionFn = null;
 		mgr._tokenThreshold = void 0;
+		mgr._tokenCounter = void 0;
+		mgr._compactionErrorHandler = void 0;
 		mgr._sessions = /* @__PURE__ */ new Map();
 		mgr._tableReady = false;
 		mgr._ready = false;
@@ -1496,8 +1571,16 @@ var SessionManager = class SessionManager {
 	* Auto-compact when estimated token count exceeds the threshold.
 	* Propagated to all sessions. Requires `onCompaction()`.
 	*/
-	compactAfter(tokenThreshold) {
+	compactAfter(tokenThreshold, options) {
 		this._tokenThreshold = tokenThreshold;
+		if (options?.tokenCounter) this._tokenCounter = options.tokenCounter;
+		return this;
+	}
+	/**
+	* Handle failures from automatic compaction in managed sessions.
+	*/
+	onCompactionError(handler) {
+		this._compactionErrorHandler = handler;
 		return this;
 	}
 	/**
@@ -1576,7 +1659,8 @@ var SessionManager = class SessionManager {
 				provider: this._createHistoryProvider()
 			});
 			if (this._compactionFn) s.onCompaction(this._compactionFn);
-			if (this._tokenThreshold != null) s.compactAfter(this._tokenThreshold);
+			if (this._tokenThreshold != null) s.compactAfter(this._tokenThreshold, { tokenCounter: this._tokenCounter });
+			if (this._compactionErrorHandler) s.onCompactionError(this._compactionErrorHandler);
 			session = s;
 			this._sessions.set(sessionId, session);
 		}