@supyagent/sdk 0.1.36 → 0.1.38

package/dist/context.d.ts

@@ -0,0 +1,204 @@
+import { UIMessage } from 'ai';
+
+interface ContextManagerOptions {
+    /**
+     * Maximum context window size in tokens for the model.
+     * Used to calculate soft/hard thresholds.
+     * @default 128_000
+     */
+    maxTokens?: number;
+    /**
+     * Soft threshold as a fraction of maxTokens (0–1).
+     * When exceeded after a response completes, triggers background summarization.
+     * @default 0.75
+     */
+    softThreshold?: number;
+    /**
+     * Hard threshold as a fraction of maxTokens (0–1).
+     * When exceeded, blocks before the next LLM call and compactifies synchronously.
+     * @default 0.90
+     */
+    hardThreshold?: number;
+    /**
+     * Tokens to reserve for the model's response output.
+     * Subtracted from the context budget when checking thresholds.
+     * @default 4096
+     */
+    responseReserve?: number;
+    /**
+     * Minimum number of recent messages to always keep (never summarized away).
+     * @default 4
+     */
+    minRecentMessages?: number;
+    /**
+     * Custom prompt for the default summarizer.
+     * Overrides the built-in summarization system prompt.
+     */
+    summaryPrompt?: string;
+    /**
+     * Full override for compactification.
+     * When provided, replaces the default summarizer entirely.
+     * Receives all messages and must return the compacted message list.
+     */
+    compactify?: (messages: UIMessage[]) => Promise<UIMessage[]>;
+    /**
+     * The model to use for summarization (same type as streamText's `model` param).
+     * Required unless a custom `compactify` function is provided.
+     */
+    summaryModel?: any;
+    /**
+     * Custom token estimator function.
+     * Given a UIMessage[], returns an estimated token count.
+     * @default Character-based heuristic (~4 chars per token)
+     */
+    estimateTokens?: (messages: UIMessage[]) => number;
+}
+interface ContextState {
+    /** Total input tokens consumed across all LLM calls in this chat */
+    totalInputTokens: number;
+    /** Total output tokens consumed across all LLM calls in this chat */
+    totalOutputTokens: number;
+    /** Estimated current context size (tokens for the next LLM call) */
+    estimatedContextSize: number;
+    /** The configured maximum context window */
+    maxTokens: number;
+    /** Usage ratio (estimatedContextSize / effectiveBudget), clamped 0–1 */
+    usageRatio: number;
+    /** Whether the soft threshold has been exceeded */
+    softThresholdExceeded: boolean;
+    /** Whether the hard threshold has been exceeded */
+    hardThresholdExceeded: boolean;
+    /** Number of context-summary messages found in the chat */
+    summaryCount: number;
+}
+interface ContextSummaryMetadata {
+    type: "context-summary";
+    /** Number of messages that were summarized */
+    messagesSummarized: number;
+    /** Estimated tokens of the original messages that were summarized */
+    originalTokens: number;
+    /** Estimated tokens of the summary */
+    summaryTokens: number;
+    /** ISO timestamp of when the summary was created */
+    createdAt: string;
+}
+interface ContextMessageMetadata {
+    context?: {
+        /** Input tokens used by this specific LLM call */
+        inputTokens?: number;
+        /** Output tokens used by this specific LLM call */
+        outputTokens?: number;
+        /** Cumulative total tokens consumed in this chat so far */
+        totalTokens?: number;
+        /** Current usage ratio at completion (0–1) */
+        usageRatio?: number;
+    };
+}
+interface ContextManager {
+    /** Get the current context state. */
+    getState(): ContextState;
+    /**
+     * Record token usage from a completed LLM step.
+     * Called from onStepFinish or onFinish callbacks.
+     */
+    recordUsage(usage: {
+        inputTokens?: number;
+        outputTokens?: number;
+    }): void;
+    /**
+     * Update the estimated context size based on current messages.
+     * Should be called whenever the message list changes.
+     */
+    updateEstimate(messages: UIMessage[]): void;
+    /**
+     * Prepare messages for the LLM call.
+     * - Finds the last context-summary message
+     * - Drops all messages before it
+     * - Injects the summary text into the system prompt
+     * - Returns the trimmed messages and updated system prompt
+     *
+     * Does NOT mutate the input array.
+     */
+    prepareMessages(messages: UIMessage[], systemPrompt: string): Promise<{
+        messages: UIMessage[];
+        systemPrompt: string;
+    }>;
+    /**
+     * Returns true if estimated context exceeds the hard threshold.
+     * Indicates compactification should block before the next LLM call.
+     */
+    shouldCompactify(messages: UIMessage[]): boolean;
+    /**
+     * Returns true if estimated context exceeds the soft threshold
+     * but not the hard threshold.
+     * Indicates background summarization should be triggered.
+     */
+    shouldSummarize(messages: UIMessage[]): boolean;
+    /**
+     * Perform compactification (synchronous in the request flow).
+     * Returns a new message array with a summary message inserted
+     * and older messages removed. Does NOT mutate the input.
+     */
+    compactify(messages: UIMessage[]): Promise<UIMessage[]>;
+    /**
+     * Get context metadata to attach to the streamed response.
+     * Used with toUIMessageStreamResponse's messageMetadata callback.
+     */
+    getMessageMetadata(): ContextMessageMetadata;
+    /** Reset internal state (e.g. when starting a new chat). */
+    reset(): void;
+}
+
+/**
+ * Prepare messages for an LLM call by trimming everything before the last
+ * context-summary message and injecting the summary into the system prompt.
+ *
+ * - If no summary exists, all messages pass through unchanged.
+ * - If a summary exists at index `i`, messages `[0..i]` are dropped and the
+ *   summary text is appended to the system prompt.
+ *
+ * Neither the input array nor any individual message is mutated.
+ */
+declare function prepareMessages(messages: UIMessage[], systemPrompt: string): {
+    messages: UIMessage[];
+    systemPrompt: string;
+};
+/**
+ * Find the index of the last message that is a context summary.
+ * Returns -1 if none found.
+ */
+declare function findLastSummaryIndex(messages: UIMessage[]): number;
+/** Count how many context-summary messages exist in the array. */
+declare function countSummaries(messages: UIMessage[]): number;
+
+/**
+ * Create a context manager that tracks token usage, detects threshold
+ * breaches, and can compactify message history via summarisation.
+ *
+ * The manager is stateless across HTTP requests — it derives cumulative
+ * totals from message metadata written by previous requests and
+ * accumulates usage recorded during the current request.
+ */
+declare function createContextManager(options?: ContextManagerOptions): ContextManager;
+
+/**
+ * Estimate the token count of a UIMessage array using a character-based heuristic.
+ *
+ * This is intentionally model-agnostic (~4 chars per token) and errs slightly
+ * on the conservative side. Pass a custom `estimateTokens` to
+ * `createContextManager` if you need model-specific accuracy.
+ */
+declare function estimateTokens(messages: UIMessage[]): number;
+
+/**
+ * Summarise a list of UIMessages into a single prose block
+ * using an LLM call via the AI SDK's `generateText`.
+ */
+declare function summarize(messages: UIMessage[], options: {
+    model: any;
+    prompt?: string;
+    /** Set to true when compacting mid-tool-chain */
+    midChain?: boolean;
+}): Promise<string>;
+
+export { type ContextManager, type ContextManagerOptions, type ContextMessageMetadata, type ContextState, type ContextSummaryMetadata, countSummaries, createContextManager, estimateTokens, findLastSummaryIndex, prepareMessages, summarize };

package/dist/context.js

@@ -0,0 +1,290 @@
+// src/context/token-estimator.ts
+var CHARS_PER_TOKEN = 4;
+var MESSAGE_OVERHEAD = 4;
+var CONVERSATION_OVERHEAD = 3;
+function estimateTokens(messages) {
+  let totalChars = 0;
+  for (const msg of messages) {
+    totalChars += MESSAGE_OVERHEAD * CHARS_PER_TOKEN;
+    for (const part of msg.parts) {
+      switch (part.type) {
+        case "text":
+          totalChars += (part.text ?? "").length;
+          break;
+        case "tool-invocation": {
+          const inv = part;
+          if (inv.toolName) totalChars += inv.toolName.length;
+          if (inv.input) totalChars += JSON.stringify(inv.input).length;
+          if (inv.output) totalChars += JSON.stringify(inv.output).length;
+          break;
+        }
+        case "file": {
+          const file = part;
+          totalChars += (file.url ?? "").length;
+          break;
+        }
+        default:
+          try {
+            totalChars += JSON.stringify(part).length;
+          } catch {
+            totalChars += 50;
+          }
+      }
+    }
+  }
+  return Math.ceil(totalChars / CHARS_PER_TOKEN) + CONVERSATION_OVERHEAD;
+}
+
+// src/context/message-preparation.ts
+var SUMMARY_PREAMBLE = "\n\n[Previous conversation summary]\nThe following is a summary of the conversation history that has been compacted to save context space:\n\n";
+function prepareMessages(messages, systemPrompt) {
+  const summaryIdx = findLastSummaryIndex(messages);
+  if (summaryIdx === -1) {
+    return { messages, systemPrompt };
+  }
+  const summaryMessage = messages[summaryIdx];
+  const summaryText = extractSummaryText(summaryMessage);
+  const trimmedMessages = messages.slice(summaryIdx + 1);
+  return {
+    messages: trimmedMessages,
+    systemPrompt: systemPrompt + SUMMARY_PREAMBLE + summaryText
+  };
+}
+function findLastSummaryIndex(messages) {
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const meta = messages[i].metadata;
+    if (meta?.type === "context-summary") {
+      return i;
+    }
+  }
+  return -1;
+}
+function countSummaries(messages) {
+  let count = 0;
+  for (const msg of messages) {
+    const meta = msg.metadata;
+    if (meta?.type === "context-summary") count++;
+  }
+  return count;
+}
+function extractSummaryText(message) {
+  for (const part of message.parts) {
+    if (part.type === "text") {
+      return part.text ?? "";
+    }
+  }
+  return "";
+}
+
+// src/context/summarizer.ts
+import { generateText } from "ai";
+var DEFAULT_SUMMARY_PROMPT = `You are a conversation summariser. Produce a concise summary of the conversation below.
+
+Focus on:
+1. Key topics discussed and decisions made
+2. Tasks completed and their outcomes (including important IDs, names, or values)
+3. Pending tasks or open questions
+4. Any in-progress multi-step work the assistant was performing
+
+Be concise but preserve all actionable information. Output only the summary text, no preamble.`;
+var MID_CHAIN_ADDENDUM = `
+
+IMPORTANT: The assistant was in the middle of executing a multi-step task when this summary was requested. Make sure to clearly note what step it was on and what remains to be done, so the task can be resumed seamlessly.`;
+async function summarize(messages, options) {
+  const conversationText = formatConversation(messages);
+  let systemPrompt = options.prompt ?? DEFAULT_SUMMARY_PROMPT;
+  if (options.midChain) {
+    systemPrompt += MID_CHAIN_ADDENDUM;
+  }
+  const { text } = await generateText({
+    model: options.model,
+    system: systemPrompt,
+    prompt: conversationText
+  });
+  return text;
+}
+function formatConversation(messages) {
+  const lines = [];
+  for (const msg of messages) {
+    const role = msg.role.toUpperCase();
+    const textParts = [];
+    const toolParts = [];
+    for (const part of msg.parts) {
+      if (part.type === "text") {
+        const text = part.text;
+        if (text?.trim()) textParts.push(text.trim());
+      } else if (part.type === "tool-invocation") {
+        const inv = part;
+        const name = inv.toolName ?? "unknown";
+        const state = inv.state ?? "";
+        if (state === "output-available" && inv.output != null) {
+          const outputStr = typeof inv.output === "string" ? inv.output : JSON.stringify(inv.output, null, 0);
+          const truncated = outputStr.length > 2e3 ? outputStr.slice(0, 2e3) + "... (truncated)" : outputStr;
+          toolParts.push(`[Tool: ${name} \u2192 ${truncated}]`);
+        } else {
+          toolParts.push(`[Tool: ${name} (${state || "pending"})]`);
+        }
+      }
+    }
+    const content = [...textParts, ...toolParts].join("\n");
+    if (content.trim()) {
+      lines.push(`${role}: ${content}`);
+    }
+  }
+  return lines.join("\n\n");
+}
+
+// src/context/context-manager.ts
+var DEFAULT_MAX_TOKENS = 128e3;
+var DEFAULT_SOFT_THRESHOLD = 0.75;
+var DEFAULT_HARD_THRESHOLD = 0.9;
+var DEFAULT_RESPONSE_RESERVE = 4096;
+var DEFAULT_MIN_RECENT = 4;
+function generateId() {
+  return `ctx-${Date.now()}-${Math.random().toString(36).slice(2, 9)}`;
+}
+function createContextManager(options) {
+  const maxTokens = options?.maxTokens ?? DEFAULT_MAX_TOKENS;
+  const softThreshold = options?.softThreshold ?? DEFAULT_SOFT_THRESHOLD;
+  const hardThreshold = options?.hardThreshold ?? DEFAULT_HARD_THRESHOLD;
+  const responseReserve = options?.responseReserve ?? DEFAULT_RESPONSE_RESERVE;
+  const minRecentMessages = options?.minRecentMessages ?? DEFAULT_MIN_RECENT;
+  const estimateTokens2 = options?.estimateTokens ?? estimateTokens;
+  const effectiveBudget = maxTokens - responseReserve;
+  let requestInputTokens = 0;
+  let requestOutputTokens = 0;
+  let estimatedContextSize = 0;
+  function getCumulativeTotals(messages) {
+    for (let i = messages.length - 1; i >= 0; i--) {
+      const meta = messages[i].metadata;
+      if (meta?.context) {
+        return {
+          totalTokens: meta.context.totalTokens ?? 0,
+          inputTokens: meta.context.inputTokens ?? 0,
+          outputTokens: meta.context.outputTokens ?? 0
+        };
+      }
+    }
+    return { totalTokens: 0, inputTokens: 0, outputTokens: 0 };
+  }
+  function computeUsageRatio() {
+    if (effectiveBudget <= 0) return 0;
+    return Math.min(Math.max(estimatedContextSize / effectiveBudget, 0), 1);
+  }
+  const manager = {
+    getState() {
+      const ratio = computeUsageRatio();
+      return {
+        totalInputTokens: requestInputTokens,
+        totalOutputTokens: requestOutputTokens,
+        estimatedContextSize,
+        maxTokens,
+        usageRatio: ratio,
+        softThresholdExceeded: ratio > softThreshold,
+        hardThresholdExceeded: ratio > hardThreshold,
+        summaryCount: 0
+        // updated by updateEstimate
+      };
+    },
+    recordUsage(usage) {
+      requestInputTokens += usage.inputTokens ?? 0;
+      requestOutputTokens += usage.outputTokens ?? 0;
+      if (usage.inputTokens != null && usage.inputTokens > 0) {
+        estimatedContextSize = usage.inputTokens;
+      }
+    },
+    updateEstimate(messages) {
+      const { messages: trimmed } = prepareMessages(messages, "");
+      estimatedContextSize = estimateTokens2(trimmed);
+    },
+    async prepareMessages(messages, systemPrompt) {
+      return prepareMessages(messages, systemPrompt);
+    },
+    shouldCompactify(messages) {
+      const { messages: trimmed } = prepareMessages(messages, "");
+      const estimate = estimateTokens2(trimmed);
+      return estimate / effectiveBudget > hardThreshold;
+    },
+    shouldSummarize(messages) {
+      const { messages: trimmed } = prepareMessages(messages, "");
+      const estimate = estimateTokens2(trimmed);
+      const ratio = estimate / effectiveBudget;
+      return ratio > softThreshold && ratio <= hardThreshold;
+    },
+    async compactify(messages) {
+      if (options?.compactify) {
+        return options.compactify(messages);
+      }
+      if (!options?.summaryModel) {
+        throw new Error(
+          "createContextManager: either `summaryModel` or a custom `compactify` function is required to perform compactification."
+        );
+      }
+      if (messages.length <= minRecentMessages) {
+        return messages;
+      }
+      const splitIdx = messages.length - minRecentMessages;
+      const messagesToSummarize = messages.slice(0, splitIdx);
+      const recentMessages = messages.slice(splitIdx);
+      const lastMsg = messages[messages.length - 1];
+      const midChain = lastMsg?.role === "assistant" && lastMsg.parts.some(
+        (p) => p.type === "tool-invocation" && p.state !== "output-available"
+      );
+      const summaryText = await summarize(messagesToSummarize, {
+        model: options.summaryModel,
+        prompt: options.summaryPrompt,
+        midChain
+      });
+      const originalTokens = estimateTokens2(messagesToSummarize);
+      const summaryMessage = {
+        id: generateId(),
+        role: "assistant",
+        parts: [{ type: "text", text: summaryText }],
+        metadata: {
+          type: "context-summary",
+          messagesSummarized: messagesToSummarize.length,
+          originalTokens,
+          summaryTokens: estimateTokens2([
+            {
+              id: "tmp",
+              role: "assistant",
+              parts: [{ type: "text", text: summaryText }]
+            }
+          ]),
+          createdAt: (/* @__PURE__ */ new Date()).toISOString()
+        }
+      };
+      return [summaryMessage, ...recentMessages];
+    },
+    getMessageMetadata() {
+      return {
+        context: {
+          inputTokens: requestInputTokens,
+          outputTokens: requestOutputTokens,
+          totalTokens: requestInputTokens + requestOutputTokens,
+          usageRatio: computeUsageRatio()
+        }
+      };
+    },
+    reset() {
+      requestInputTokens = 0;
+      requestOutputTokens = 0;
+      estimatedContextSize = 0;
+    }
+  };
+  const originalGetState = manager.getState.bind(manager);
+  manager.getState = function getStateWithSummaryCount() {
+    const state = originalGetState();
+    return state;
+  };
+  return manager;
+}
+export {
+  countSummaries,
+  createContextManager,
+  estimateTokens,
+  findLastSummaryIndex,
+  prepareMessages,
+  summarize
+};
+//# sourceMappingURL=context.js.map

package/dist/context.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/context/token-estimator.ts","../src/context/message-preparation.ts","../src/context/summarizer.ts","../src/context/context-manager.ts"],"sourcesContent":["import type { UIMessage } from \"ai\";\n\n/** Approximate characters per token for mixed English text and code. */\nconst CHARS_PER_TOKEN = 4;\n\n/** Overhead tokens added per message (role, framing). */\nconst MESSAGE_OVERHEAD = 4;\n\n/** Base overhead for the conversation itself. */\nconst CONVERSATION_OVERHEAD = 3;\n\n/**\n * Estimate the token count of a UIMessage array using a character-based heuristic.\n *\n * This is intentionally model-agnostic (~4 chars per token) and errs slightly\n * on the conservative side. Pass a custom `estimateTokens` to\n * `createContextManager` if you need model-specific accuracy.\n */\nexport function estimateTokens(messages: UIMessage[]): number {\n  let totalChars = 0;\n\n  for (const msg of messages) {\n    totalChars += MESSAGE_OVERHEAD * CHARS_PER_TOKEN; // per-message overhead\n\n    for (const part of msg.parts) {\n      switch (part.type) {\n        case \"text\":\n          totalChars += ((part as { text: string }).text ?? \"\").length;\n          break;\n\n        case \"tool-invocation\": {\n          // Estimate from the serialised input + output\n          const inv = part as {\n            toolName?: string;\n            input?: unknown;\n            output?: unknown;\n          };\n          if (inv.toolName) totalChars += inv.toolName.length;\n          if (inv.input) totalChars += JSON.stringify(inv.input).length;\n          if (inv.output) totalChars += JSON.stringify(inv.output).length;\n          break;\n        }\n\n        case \"file\": {\n          // Only count the URL length (actual file content is handled by the provider)\n          const file = part as { url?: string };\n          totalChars += (file.url ?? \"\").length;\n          break;\n        }\n\n        default:\n          // Catch-all for unknown part types\n          try {\n            totalChars += JSON.stringify(part).length;\n          } catch {\n            totalChars += 50; // small fallback\n          }\n      }\n    }\n  }\n\n  return Math.ceil(totalChars / CHARS_PER_TOKEN) + CONVERSATION_OVERHEAD;\n}\n","import type { UIMessage } from \"ai\";\nimport type { ContextSummaryMetadata } from \"./types.js\";\n\nconst SUMMARY_PREAMBLE =\n  \"\\n\\n[Previous conversation summary]\\nThe following is a summary of the conversation history that has been compacted to save context space:\\n\\n\";\n\n/**\n * Prepare messages for an LLM call by trimming everything before the last\n * context-summary message and injecting the summary into the system prompt.\n *\n * - If no summary exists, all messages pass through unchanged.\n * - If a summary exists at index `i`, messages `[0..i]` are dropped and the\n *   summary text is appended to the system prompt.\n *\n * Neither the input array nor any individual message is mutated.\n */\nexport function prepareMessages(\n  messages: UIMessage[],\n  systemPrompt: string\n): { messages: UIMessage[]; systemPrompt: string } {\n  const summaryIdx = findLastSummaryIndex(messages);\n\n  if (summaryIdx === -1) {\n    return { messages, systemPrompt };\n  }\n\n  const summaryMessage = messages[summaryIdx];\n  const summaryText = extractSummaryText(summaryMessage);\n  const trimmedMessages = messages.slice(summaryIdx + 1);\n\n  return {\n    messages: trimmedMessages,\n    systemPrompt: systemPrompt + SUMMARY_PREAMBLE + summaryText,\n  };\n}\n\n/**\n * Find the index of the last message that is a context summary.\n * Returns -1 if none found.\n */\nexport function findLastSummaryIndex(messages: UIMessage[]): number {\n  for (let i = messages.length - 1; i >= 0; i--) {\n    const meta = messages[i].metadata as ContextSummaryMetadata | undefined;\n    if (meta?.type === \"context-summary\") {\n      return i;\n    }\n  }\n  return -1;\n}\n\n/** Count how many context-summary messages exist in the array. */\nexport function countSummaries(messages: UIMessage[]): number {\n  let count = 0;\n  for (const msg of messages) {\n    const meta = msg.metadata as ContextSummaryMetadata | undefined;\n    if (meta?.type === \"context-summary\") count++;\n  }\n  return count;\n}\n\n/** Extract the text content from a summary message. */\nfunction extractSummaryText(message: UIMessage): string {\n  for (const part of message.parts) {\n    if (part.type === \"text\") {\n      return (part as { text: string }).text ?? \"\";\n    }\n  }\n  return \"\";\n}\n","import { generateText, type UIMessage } from \"ai\";\n\nconst DEFAULT_SUMMARY_PROMPT = `You are a conversation summariser. Produce a concise summary of the conversation below.\n\nFocus on:\n1. Key topics discussed and decisions made\n2. Tasks completed and their outcomes (including important IDs, names, or values)\n3. Pending tasks or open questions\n4. Any in-progress multi-step work the assistant was performing\n\nBe concise but preserve all actionable information. Output only the summary text, no preamble.`;\n\nconst MID_CHAIN_ADDENDUM = `\\n\\nIMPORTANT: The assistant was in the middle of executing a multi-step task when this summary was requested. Make sure to clearly note what step it was on and what remains to be done, so the task can be resumed seamlessly.`;\n\n/**\n * Summarise a list of UIMessages into a single prose block\n * using an LLM call via the AI SDK's `generateText`.\n */\nexport async function summarize(\n  messages: UIMessage[],\n  options: {\n    // eslint-disable-next-line @typescript-eslint/no-explicit-any\n    model: any;\n    prompt?: string;\n    /** Set to true when compacting mid-tool-chain */\n    midChain?: boolean;\n  }\n): Promise<string> {\n  const conversationText = formatConversation(messages);\n\n  let systemPrompt = options.prompt ?? DEFAULT_SUMMARY_PROMPT;\n  if (options.midChain) {\n    systemPrompt += MID_CHAIN_ADDENDUM;\n  }\n\n  const { text } = await generateText({\n    model: options.model,\n    system: systemPrompt,\n    prompt: conversationText,\n  });\n\n  return text;\n}\n\n/** Convert UIMessage[] into a plain-text conversation transcript for the summariser. */\nfunction formatConversation(messages: UIMessage[]): string {\n  const lines: string[] = [];\n\n  for (const msg of messages) {\n    const role = msg.role.toUpperCase();\n    const textParts: string[] = [];\n    const toolParts: string[] = [];\n\n    for (const part of msg.parts) {\n      if (part.type === \"text\") {\n        const text = (part as { text: string }).text;\n        if (text?.trim()) textParts.push(text.trim());\n      } else if (part.type === \"tool-invocation\") {\n        const inv = part as { toolName?: string; output?: unknown; state?: string };\n        const name = inv.toolName ?? \"unknown\";\n        const state = inv.state ?? \"\";\n        if (state === \"output-available\" && inv.output != null) {\n          const outputStr =\n            typeof inv.output === \"string\"\n              ? inv.output\n              : JSON.stringify(inv.output, null, 0);\n          // Truncate very large outputs to keep the summariser prompt reasonable\n          const truncated =\n            outputStr.length > 2000\n              ? outputStr.slice(0, 2000) + \"... (truncated)\"\n              : outputStr;\n          toolParts.push(`[Tool: ${name} → ${truncated}]`);\n        } else {\n          toolParts.push(`[Tool: ${name} (${state || \"pending\"})]`);\n        }\n      }\n    }\n\n    const content = [...textParts, ...toolParts].join(\"\\n\");\n    if (content.trim()) {\n      lines.push(`${role}: ${content}`);\n    }\n  }\n\n  return lines.join(\"\\n\\n\");\n}\n","import type { UIMessage } from \"ai\";\nimport type {\n  ContextManager,\n  ContextManagerOptions,\n  ContextMessageMetadata,\n  ContextState,\n  ContextSummaryMetadata,\n} from \"./types.js\";\nimport { estimateTokens as defaultEstimateTokens } from \"./token-estimator.js\";\nimport {\n  prepareMessages as doPrepareMessages,\n  findLastSummaryIndex,\n  countSummaries,\n} from \"./message-preparation.js\";\nimport { summarize } from \"./summarizer.js\";\n\nconst DEFAULT_MAX_TOKENS = 128_000;\nconst DEFAULT_SOFT_THRESHOLD = 0.75;\nconst DEFAULT_HARD_THRESHOLD = 0.9;\nconst DEFAULT_RESPONSE_RESERVE = 4096;\nconst DEFAULT_MIN_RECENT = 4;\n\n/** Generate a simple unique ID for summary messages. */\nfunction generateId(): string {\n  return `ctx-${Date.now()}-${Math.random().toString(36).slice(2, 9)}`;\n}\n\n/**\n * Create a context manager that tracks token usage, detects threshold\n * breaches, and can compactify message history via summarisation.\n *\n * The manager is stateless across HTTP requests — it derives cumulative\n * totals from message metadata written by previous requests and\n * accumulates usage recorded during the current request.\n */\nexport function createContextManager(\n  options?: ContextManagerOptions\n): ContextManager {\n  const maxTokens = options?.maxTokens ?? DEFAULT_MAX_TOKENS;\n  const softThreshold = options?.softThreshold ?? DEFAULT_SOFT_THRESHOLD;\n  const hardThreshold = options?.hardThreshold ?? DEFAULT_HARD_THRESHOLD;\n  const responseReserve = options?.responseReserve ?? DEFAULT_RESPONSE_RESERVE;\n  const minRecentMessages = options?.minRecentMessages ?? DEFAULT_MIN_RECENT;\n  const estimateTokens = options?.estimateTokens ?? defaultEstimateTokens;\n\n  // Effective budget = maxTokens minus what we reserve for the response\n  const effectiveBudget = maxTokens - responseReserve;\n\n  // ── Mutable per-request state ──────────────────────────────────────\n\n  /** Input tokens recorded during this request (from onStepFinish). */\n  let requestInputTokens = 0;\n  /** Output tokens recorded during this request. */\n  let requestOutputTokens = 0;\n  /** Latest estimated context size in tokens. */\n  let estimatedContextSize = 0;\n\n  // ── Helpers ────────────────────────────────────────────────────────\n\n  /** Read cumulative totals from the last assistant message that has context metadata. */\n  function getCumulativeTotals(messages: UIMessage[]): {\n    totalTokens: number;\n    inputTokens: number;\n    outputTokens: number;\n  } {\n    for (let i = messages.length - 1; i >= 0; i--) {\n      const meta = messages[i].metadata as ContextMessageMetadata | undefined;\n      if (meta?.context) {\n        return {\n          totalTokens: meta.context.totalTokens ?? 0,\n          inputTokens: meta.context.inputTokens ?? 0,\n          outputTokens: meta.context.outputTokens ?? 0,\n        };\n      }\n    }\n    return { totalTokens: 0, inputTokens: 0, outputTokens: 0 };\n  }\n\n  function computeUsageRatio(): number {\n    if (effectiveBudget <= 0) return 0;\n    return Math.min(Math.max(estimatedContextSize / effectiveBudget, 0), 1);\n  }\n\n  // ── ContextManager implementation ──────────────────────────────────\n\n  const manager: ContextManager = {\n    getState(): ContextState {\n      const ratio = computeUsageRatio();\n      return {\n        totalInputTokens: requestInputTokens,\n        totalOutputTokens: requestOutputTokens,\n        estimatedContextSize,\n        maxTokens,\n        usageRatio: ratio,\n        softThresholdExceeded: ratio > softThreshold,\n        hardThresholdExceeded: ratio > hardThreshold,\n        summaryCount: 0, // updated by updateEstimate\n      };\n    },\n\n    recordUsage(usage) {\n      requestInputTokens += usage.inputTokens ?? 0;\n      requestOutputTokens += usage.outputTokens ?? 0;\n      // Use actual input tokens as the new context size estimate\n      // (they reflect the real token count the provider saw)\n      if (usage.inputTokens != null && usage.inputTokens > 0) {\n        estimatedContextSize = usage.inputTokens;\n      }\n    },\n\n    updateEstimate(messages) {\n      // After prepareMessages, estimate what the LLM would see\n      const { messages: trimmed } = doPrepareMessages(messages, \"\");\n      estimatedContextSize = estimateTokens(trimmed);\n    },\n\n    async prepareMessages(messages, systemPrompt) {\n      return doPrepareMessages(messages, systemPrompt);\n    },\n\n    shouldCompactify(messages) {\n      // Re-estimate in case messages changed since last updateEstimate\n      const { messages: trimmed } = doPrepareMessages(messages, \"\");\n      const estimate = estimateTokens(trimmed);\n      return estimate / effectiveBudget > hardThreshold;\n    },\n\n    shouldSummarize(messages) {\n      const { messages: trimmed } = doPrepareMessages(messages, \"\");\n      const estimate = estimateTokens(trimmed);\n      const ratio = estimate / effectiveBudget;\n      return ratio > softThreshold && ratio <= hardThreshold;\n    },\n\n    async compactify(messages) {\n      // If a custom compactify is provided, delegate entirely\n      if (options?.compactify) {\n        return options.compactify(messages);\n      }\n\n      if (!options?.summaryModel) {\n        throw new Error(\n          \"createContextManager: either `summaryModel` or a custom `compactify` function is required to perform compactification.\"\n        );\n      }\n\n      // Don't compact if there aren't enough messages\n      if (messages.length <= minRecentMessages) {\n        return messages;\n      }\n\n      // Split: everything before the cutpoint gets summarised\n      const splitIdx = messages.length - minRecentMessages;\n      const messagesToSummarize = messages.slice(0, splitIdx);\n      const recentMessages = messages.slice(splitIdx);\n\n      // Detect if we're mid-chain (last message is assistant with tool invocations\n      // that might indicate an in-progress workflow)\n      const lastMsg = messages[messages.length - 1];\n      const midChain =\n        lastMsg?.role === \"assistant\" &&\n        lastMsg.parts.some(\n          (p) =>\n            p.type === \"tool-invocation\" &&\n            (p as { state?: string }).state !== \"output-available\"\n        );\n\n      const summaryText = await summarize(messagesToSummarize, {\n        model: options.summaryModel,\n        prompt: options.summaryPrompt,\n        midChain,\n      });\n\n      const originalTokens = estimateTokens(messagesToSummarize);\n      const summaryMessage: UIMessage = {\n        id: generateId(),\n        role: \"assistant\",\n        parts: [{ type: \"text\", text: summaryText }],\n        metadata: {\n          type: \"context-summary\",\n          messagesSummarized: messagesToSummarize.length,\n          originalTokens,\n          summaryTokens: estimateTokens([\n            {\n              id: \"tmp\",\n              role: \"assistant\",\n              parts: [{ type: \"text\", text: summaryText }],\n            },\n          ]),\n          createdAt: new Date().toISOString(),\n        } satisfies ContextSummaryMetadata,\n      };\n\n      return [summaryMessage, ...recentMessages];\n    },\n\n    getMessageMetadata(): ContextMessageMetadata {\n      return {\n        context: {\n          inputTokens: requestInputTokens,\n          outputTokens: requestOutputTokens,\n          totalTokens: requestInputTokens + requestOutputTokens,\n          usageRatio: computeUsageRatio(),\n        },\n      };\n    },\n\n    reset() {\n      requestInputTokens = 0;\n      requestOutputTokens = 0;\n      estimatedContextSize = 0;\n    },\n  };\n\n  // Patch getState to include summaryCount dynamically\n  const originalGetState = manager.getState.bind(manager);\n  manager.getState = function getStateWithSummaryCount() {\n    const state = originalGetState();\n    // summaryCount is derived; caller should use countSummaries if needed\n    return state;\n  };\n\n  return manager;\n}\n\nexport { countSummaries };\n"],"mappings":";AAGA,IAAM,kBAAkB;AAGxB,IAAM,mBAAmB;AAGzB,IAAM,wBAAwB;AASvB,SAAS,eAAe,UAA+B;AAC5D,MAAI,aAAa;AAEjB,aAAW,OAAO,UAAU;AAC1B,kBAAc,mBAAmB;AAEjC,eAAW,QAAQ,IAAI,OAAO;AAC5B,cAAQ,KAAK,MAAM;AAAA,QACjB,KAAK;AACH,yBAAgB,KAA0B,QAAQ,IAAI;AACtD;AAAA,QAEF,KAAK,mBAAmB;AAEtB,gBAAM,MAAM;AAKZ,cAAI,IAAI,SAAU,eAAc,IAAI,SAAS;AAC7C,cAAI,IAAI,MAAO,eAAc,KAAK,UAAU,IAAI,KAAK,EAAE;AACvD,cAAI,IAAI,OAAQ,eAAc,KAAK,UAAU,IAAI,MAAM,EAAE;AACzD;AAAA,QACF;AAAA,QAEA,KAAK,QAAQ;AAEX,gBAAM,OAAO;AACb,yBAAe,KAAK,OAAO,IAAI;AAC/B;AAAA,QACF;AAAA,QAEA;AAEE,cAAI;AACF,0BAAc,KAAK,UAAU,IAAI,EAAE;AAAA,UACrC,QAAQ;AACN,0BAAc;AAAA,UAChB;AAAA,MACJ;AAAA,IACF;AAAA,EACF;AAEA,SAAO,KAAK,KAAK,aAAa,eAAe,IAAI;AACnD;;;AC3DA,IAAM,mBACJ;AAYK,SAAS,gBACd,UACA,cACiD;AACjD,QAAM,aAAa,qBAAqB,QAAQ;AAEhD,MAAI,eAAe,IAAI;AACrB,WAAO,EAAE,UAAU,aAAa;AAAA,EAClC;AAEA,QAAM,iBAAiB,SAAS,UAAU;AAC1C,QAAM,cAAc,mBAAmB,cAAc;AACrD,QAAM,kBAAkB,SAAS,MAAM,aAAa,CAAC;AAErD,SAAO;AAAA,IACL,UAAU;AAAA,IACV,cAAc,eAAe,mBAAmB;AAAA,EAClD;AACF;AAMO,SAAS,qBAAqB,UAA+B;AAClE,WAAS,IAAI,SAAS,SAAS,GAAG,KAAK,GAAG,KAAK;AAC7C,UAAM,OAAO,SAAS,CAAC,EAAE;AACzB,QAAI,MAAM,SAAS,mBAAmB;AACpC,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;AAGO,SAAS,eAAe,UAA+B;AAC5D,MAAI,QAAQ;AACZ,aAAW,OAAO,UAAU;AAC1B,UAAM,OAAO,IAAI;AACjB,QAAI,MAAM,SAAS,kBAAmB;AAAA,EACxC;AACA,SAAO;AACT;AAGA,SAAS,mBAAmB,SAA4B;AACtD,aAAW,QAAQ,QAAQ,OAAO;AAChC,QAAI,KAAK,SAAS,QAAQ;AACxB,aAAQ,KAA0B,QAAQ;AAAA,IAC5C;AAAA,EACF;AACA,SAAO;AACT;;;ACpEA,SAAS,oBAAoC;AAE7C,IAAM,yBAAyB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAU/B,IAAM,qBAAqB;AAAA;AAAA;AAM3B,eAAsB,UACpB,UACA,SAOiB;AACjB,QAAM,mBAAmB,mBAAmB,QAAQ;AAEpD,MAAI,eAAe,QAAQ,UAAU;AACrC,MAAI,QAAQ,UAAU;AACpB,oBAAgB;AAAA,EAClB;AAEA,QAAM,EAAE,KAAK,IAAI,MAAM,aAAa;AAAA,IAClC,OAAO,QAAQ;AAAA,IACf,QAAQ;AAAA,IACR,QAAQ;AAAA,EACV,CAAC;AAED,SAAO;AACT;AAGA,SAAS,mBAAmB,UAA+B;AACzD,QAAM,QAAkB,CAAC;AAEzB,aAAW,OAAO,UAAU;AAC1B,UAAM,OAAO,IAAI,KAAK,YAAY;AAClC,UAAM,YAAsB,CAAC;AAC7B,UAAM,YAAsB,CAAC;AAE7B,eAAW,QAAQ,IAAI,OAAO;AAC5B,UAAI,KAAK,SAAS,QAAQ;AACxB,cAAM,OAAQ,KAA0B;AACxC,YAAI,MAAM,KAAK,EAAG,WAAU,KAAK,KAAK,KAAK,CAAC;AAAA,MAC9C,WAAW,KAAK,SAAS,mBAAmB;AAC1C,cAAM,MAAM;AACZ,cAAM,OAAO,IAAI,YAAY;AAC7B,cAAM,QAAQ,IAAI,SAAS;AAC3B,YAAI,UAAU,sBAAsB,IAAI,UAAU,MAAM;AACtD,gBAAM,YACJ,OAAO,IAAI,WAAW,WAClB,IAAI,SACJ,KAAK,UAAU,IAAI,QAAQ,MAAM,CAAC;AAExC,gBAAM,YACJ,UAAU,SAAS,MACf,UAAU,MAAM,GAAG,GAAI,IAAI,oBAC3B;AACN,oBAAU,KAAK,UAAU,IAAI,WAAM,SAAS,GAAG;AAAA,QACjD,OAAO;AACL,oBAAU,KAAK,UAAU,IAAI,KAAK,SAAS,SAAS,IAAI;AAAA,QAC1D;AAAA,MACF;AAAA,IACF;AAEA,UAAM,UAAU,CAAC,GAAG,WAAW,GAAG,SAAS,EAAE,KAAK,IAAI;AACtD,QAAI,QAAQ,KAAK,GAAG;AAClB,YAAM,KAAK,GAAG,IAAI,KAAK,OAAO,EAAE;AAAA,IAClC;AAAA,EACF;AAEA,SAAO,MAAM,KAAK,MAAM;AAC1B;;;ACrEA,IAAM,qBAAqB;AAC3B,IAAM,yBAAyB;AAC/B,IAAM,yBAAyB;AAC/B,IAAM,2BAA2B;AACjC,IAAM,qBAAqB;AAG3B,SAAS,aAAqB;AAC5B,SAAO,OAAO,KAAK,IAAI,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,CAAC,CAAC;AACpE;AAUO,SAAS,qBACd,SACgB;AAChB,QAAM,YAAY,SAAS,aAAa;AACxC,QAAM,gBAAgB,SAAS,iBAAiB;AAChD,QAAM,gBAAgB,SAAS,iBAAiB;AAChD,QAAM,kBAAkB,SAAS,mBAAmB;AACpD,QAAM,oBAAoB,SAAS,qBAAqB;AACxD,QAAMA,kBAAiB,SAAS,kBAAkB;AAGlD,QAAM,kBAAkB,YAAY;AAKpC,MAAI,qBAAqB;AAEzB,MAAI,sBAAsB;AAE1B,MAAI,uBAAuB;AAK3B,WAAS,oBAAoB,UAI3B;AACA,aAAS,IAAI,SAAS,SAAS,GAAG,KAAK,GAAG,KAAK;AAC7C,YAAM,OAAO,SAAS,CAAC,EAAE;AACzB,UAAI,MAAM,SAAS;AACjB,eAAO;AAAA,UACL,aAAa,KAAK,QAAQ,eAAe;AAAA,UACzC,aAAa,KAAK,QAAQ,eAAe;AAAA,UACzC,cAAc,KAAK,QAAQ,gBAAgB;AAAA,QAC7C;AAAA,MACF;AAAA,IACF;AACA,WAAO,EAAE,aAAa,GAAG,aAAa,GAAG,cAAc,EAAE;AAAA,EAC3D;AAEA,WAAS,oBAA4B;AACnC,QAAI,mBAAmB,EAAG,QAAO;AACjC,WAAO,KAAK,IAAI,KAAK,IAAI,uBAAuB,iBAAiB,CAAC,GAAG,CAAC;AAAA,EACxE;AAIA,QAAM,UAA0B;AAAA,IAC9B,WAAyB;AACvB,YAAM,QAAQ,kBAAkB;AAChC,aAAO;AAAA,QACL,kBAAkB;AAAA,QAClB,mBAAmB;AAAA,QACnB;AAAA,QACA;AAAA,QACA,YAAY;AAAA,QACZ,uBAAuB,QAAQ;AAAA,QAC/B,uBAAuB,QAAQ;AAAA,QAC/B,cAAc;AAAA;AAAA,MAChB;AAAA,IACF;AAAA,IAEA,YAAY,OAAO;AACjB,4BAAsB,MAAM,eAAe;AAC3C,6BAAuB,MAAM,gBAAgB;AAG7C,UAAI,MAAM,eAAe,QAAQ,MAAM,cAAc,GAAG;AACtD,+BAAuB,MAAM;AAAA,MAC/B;AAAA,IACF;AAAA,IAEA,eAAe,UAAU;AAEvB,YAAM,EAAE,UAAU,QAAQ,IAAI,gBAAkB,UAAU,EAAE;AAC5D,6BAAuBA,gBAAe,OAAO;AAAA,IAC/C;AAAA,IAEA,MAAM,gBAAgB,UAAU,cAAc;AAC5C,aAAO,gBAAkB,UAAU,YAAY;AAAA,IACjD;AAAA,IAEA,iBAAiB,UAAU;AAEzB,YAAM,EAAE,UAAU,QAAQ,IAAI,gBAAkB,UAAU,EAAE;AAC5D,YAAM,WAAWA,gBAAe,OAAO;AACvC,aAAO,WAAW,kBAAkB;AAAA,IACtC;AAAA,IAEA,gBAAgB,UAAU;AACxB,YAAM,EAAE,UAAU,QAAQ,IAAI,gBAAkB,UAAU,EAAE;AAC5D,YAAM,WAAWA,gBAAe,OAAO;AACvC,YAAM,QAAQ,WAAW;AACzB,aAAO,QAAQ,iBAAiB,SAAS;AAAA,IAC3C;AAAA,IAEA,MAAM,WAAW,UAAU;AAEzB,UAAI,SAAS,YAAY;AACvB,eAAO,QAAQ,WAAW,QAAQ;AAAA,MACpC;AAEA,UAAI,CAAC,SAAS,cAAc;AAC1B,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAGA,UAAI,SAAS,UAAU,mBAAmB;AACxC,eAAO;AAAA,MACT;AAGA,YAAM,WAAW,SAAS,SAAS;AACnC,YAAM,sBAAsB,SAAS,MAAM,GAAG,QAAQ;AACtD,YAAM,iBAAiB,SAAS,MAAM,QAAQ;AAI9C,YAAM,UAAU,SAAS,SAAS,SAAS,CAAC;AAC5C,YAAM,WACJ,SAAS,SAAS,eAClB,QAAQ,MAAM;AAAA,QACZ,CAAC,MACC,EAAE,SAAS,qBACV,EAAyB,UAAU;AAAA,MACxC;AAEF,YAAM,cAAc,MAAM,UAAU,qBAAqB;AAAA,QACvD,OAAO,QAAQ;AAAA,QACf,QAAQ,QAAQ;AAAA,QAChB;AAAA,MACF,CAAC;AAED,YAAM,iBAAiBA,gBAAe,mBAAmB;AACzD,YAAM,iBAA4B;AAAA,QAChC,IAAI,WAAW;AAAA,QACf,MAAM;AAAA,QACN,OAAO,CAAC,EAAE,MAAM,QAAQ,MAAM,YAAY,CAAC;AAAA,QAC3C,UAAU;AAAA,UACR,MAAM;AAAA,UACN,oBAAoB,oBAAoB;AAAA,UACxC;AAAA,UACA,eAAeA,gBAAe;AAAA,YAC5B;AAAA,cACE,IAAI;AAAA,cACJ,MAAM;AAAA,cACN,OAAO,CAAC,EAAE,MAAM,QAAQ,MAAM,YAAY,CAAC;AAAA,YAC7C;AAAA,UACF,CAAC;AAAA,UACD,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,QACpC;AAAA,MACF;AAEA,aAAO,CAAC,gBAAgB,GAAG,cAAc;AAAA,IAC3C;AAAA,IAEA,qBAA6C;AAC3C,aAAO;AAAA,QACL,SAAS;AAAA,UACP,aAAa;AAAA,UACb,cAAc;AAAA,UACd,aAAa,qBAAqB;AAAA,UAClC,YAAY,kBAAkB;AAAA,QAChC;AAAA,MACF;AAAA,IACF;AAAA,IAEA,QAAQ;AACN,2BAAqB;AACrB,4BAAsB;AACtB,6BAAuB;AAAA,IACzB;AAAA,EACF;AAGA,QAAM,mBAAmB,QAAQ,SAAS,KAAK,OAAO;AACtD,UAAQ,WAAW,SAAS,2BAA2B;AACrD,UAAM,QAAQ,iBAAiB;AAE/B,WAAO;AAAA,EACT;AAEA,SAAO;AACT;","names":["estimateTokens"]}