agent-toolbelt

Instructions

Implementation Reference

• mcp-server/src/index.ts:1158-1198 (handler)

  The 'pack_context_window' tool is registered in mcp-server/src/index.ts, and its handler calls the Agent Toolbelt API endpoint 'context-window-packer'.

  server.registerTool(
    "pack_context_window",
    {
      title: "Context Window Packer",
      description:
        "Pack content chunks into a token budget for an LLM context window. " +
        "Selects the best subset of chunks that fits within the token limit using priority, greedy, or balanced strategies. " +
        "Use when you have more content than fits in the context window.",
      inputSchema: {
        chunks: z.array(z.object({
          text: z.string().describe("Content of this chunk"),
          label: z.string().optional().describe("Optional identifier"),
          priority: z.number().min(0).max(10).default(5).describe("Importance 0–10"),
          metadata: z.record(z.unknown()).optional(),
        })).describe("Content chunks to pack"),
        tokenBudget: z.number().int().describe("Maximum tokens allowed"),
        model: z.string().default("gpt-4o").describe("Target model for tokenization"),
        strategy: z.enum(["priority", "greedy", "balanced"]).default("priority").describe("Packing strategy"),
        separator: z.string().default("\n\n").describe("Text between chunks"),
        systemPrompt: z.string().optional().describe("System prompt to reserve tokens for"),
        reserveForOutput: z.number().int().min(0).default(0).describe("Tokens to reserve for output"),
      },
    },
    async ({ chunks, tokenBudget, model, strategy, separator, systemPrompt, reserveForOutput }) => {
      const result = await callToolApi("context-window-packer", { chunks, tokenBudget, model, strategy, separator, systemPrompt, reserveForOutput });
      const data = result as any;
      const r = data.result;
  
      const lines = [
        `**Packed ${r.stats.chunksPacked}/${r.stats.chunksTotal} chunks** | ${r.stats.tokensUsed}/${r.stats.effectiveBudget} tokens (${r.stats.utilizationPercent}% utilized)`,
        `**Strategy:** ${r.strategy} | **Model:** ${r.model}`,
        r.stats.systemPromptTokens > 0 ? `**System prompt:** ${r.stats.systemPromptTokens} tokens reserved` : "",
        r.stats.reservedForOutput > 0 ? `**Output reservation:** ${r.stats.reservedForOutput} tokens` : "",
        "",
        r.packed.length > 0 ? `**Packed chunks:**\n${r.packed.map((p: any) => `- ${p.label || `chunk[${p.originalIndex}]`} (priority ${p.priority}, ${p.tokens} tokens)`).join("\n")}` : "",
        r.excluded.length > 0 ? `\n**Excluded chunks:**\n${r.excluded.map((e: any) => `- ${e.label || `chunk[${e.originalIndex}]`} (${e.tokens} tokens, ${e.reason})`).join("\n")}` : "",
      ].filter(Boolean);
  
      return { content: [{ type: "text" as const, text: lines.join("\n") }] };
    }
  );