mcpctl/src/mcpd/src/bootstrap/system-project.ts

/**
 * Bootstrap the mcpctl-system project and its system prompts.
 *
 * This runs on every mcpd startup and uses upserts to be idempotent.
 * System prompts are editable by users but will be re-created if deleted.
 */

import type { PrismaClient } from '@prisma/client';

/** Well-known project name for system prompts. */
export const SYSTEM_PROJECT_NAME = 'mcpctl-system';

/** Well-known email for the system user. */
const SYSTEM_USER_EMAIL = 'system@mcpctl.local';

interface SystemPromptDef {
  name: string;
  priority: number;
  content: string;
  /** Template variables that must be present when editing (e.g., '{{maxTokens}}'). */
  requiredVars?: string[];
}

const SYSTEM_PROMPTS: SystemPromptDef[] = [
  {
    name: 'gate-instructions',
    priority: 10,
    content: `This project uses a gated session. Before you can access tools, you must start a session by calling begin_session.

Call begin_session immediately using the arguments it requires (check its input schema). If it accepts a description, briefly describe the user's task. If it accepts tags, provide 3-7 keywords relevant to the user's request.

The available tools and prompts are listed below. After calling begin_session, you will receive relevant project context and full access to all tools.`,
  },
  {
    name: 'gate-encouragement',
    priority: 10,
    content: `If any of the listed prompts seem relevant to your work, or if you encounter unfamiliar patterns, conventions, or constraints during implementation, use read_prompts({ tags: [...] }) to retrieve them.

It is better to check and not need it than to proceed without important context. The project maintainers have documented common pitfalls, architecture decisions, and required patterns — taking 10 seconds to retrieve a prompt can save hours of rework.`,
  },
  {
    name: 'gate-intercept-preamble',
    priority: 10,
    content: `The following project context was automatically retrieved based on your tool call. You bypassed the begin_session step, so this context was matched using keywords extracted from your tool invocation.

Review this context carefully — it may contain important guidelines, constraints, or patterns relevant to your work. If you need more context, use read_prompts({ tags: [...] }) at any time.`,
  },
  {
    name: 'gate-session-active',
    priority: 10,
    content: `The session is now active with full tool access. Proceed with the user's original request using the tools listed above.`,
  },
  {
    name: 'session-greeting',
    priority: 10,
    content: `Welcome to this project. To get started, call begin_session with the arguments it requires.

Examples:
  begin_session({ tags: ["zigbee", "pairing", "mqtt"] })
  begin_session({ description: "I want to pair a new Zigbee device" })

This will load relevant project context, policies, and guidelines tailored to your work.`,
  },

  // ── LLM pipeline prompts (priority 5, editable) ──

  {
    name: 'llm-response-filter',
    priority: 5,
    content: `You are a data filtering assistant. Your job is to extract only the relevant information from MCP tool responses.

Rules:
- Remove redundant or verbose fields that aren't useful to the user's query
- Keep essential identifiers, names, statuses, and key metrics
- Preserve error messages and warnings in full
- If the response is already concise, return it unchanged
- Output valid JSON only, no markdown or explanations
- If you cannot parse the input, return it unchanged`,
  },
  {
    name: 'llm-request-optimization',
    priority: 5,
    content: `You are a query optimization assistant. Your job is to optimize MCP tool call parameters.

Rules:
- Add appropriate filters or limits if the query is too broad
- Keep the original intent of the request
- Output valid JSON with the optimized parameters only, no markdown or explanations
- If no optimization is needed, return the original parameters unchanged`,
  },
  {
    name: 'llm-pagination-index',
    priority: 5,
    content: `You are a document indexing assistant. Given a large tool response split into pages, generate a concise summary for each page describing what data it contains.

Rules:
- For each page, write 1-2 sentences describing the key content
- Be specific: mention entity names, IDs, counts, or key fields visible on that page
- If it's JSON, describe the structure and notable entries
- If it's text, describe the topics covered
- Output valid JSON only: an array of objects with "page" (1-based number) and "summary" (string)
- Example output: [{"page": 1, "summary": "Configuration nodes and global settings (inject, debug, function nodes 1-15)"}, {"page": 2, "summary": "HTTP request nodes and API integrations (nodes 16-40)"}]`,
  },
  {
    name: 'llm-gate-context-selector',
    priority: 5,
    content: `You are a context selection assistant. Given a developer's task keywords and a list of available project prompts, select which prompts are relevant to their work. Return a JSON object with "selectedNames" (array of prompt names) and "reasoning" (brief explanation). Priority 10 prompts must always be included.`,
  },
  {
    name: 'llm-summarize',
    priority: 5,
    requiredVars: ['{{maxTokens}}'],
    content: `Summarize the following in about {{maxTokens}} tokens. Preserve all items marked MUST, REQUIRED, or CRITICAL verbatim. Be specific — mention names, IDs, counts, key values.`,
  },
  {
    name: 'llm-paginate-titles',
    priority: 5,
    requiredVars: ['{{pageCount}}'],
    content: `Generate exactly {{pageCount}} short descriptive titles (max 60 chars each) for the following {{pageCount}} pages. Return ONLY a JSON array of {{pageCount}} strings. No markdown, no explanation.`,
  },
];

/**
 * Ensure the mcpctl-system project and its system prompts exist.
 * Uses upserts so this is safe to call on every startup.
 */
export async function bootstrapSystemProject(prisma: PrismaClient): Promise<void> {
  // Ensure a system user exists (needed as project owner)
  const systemUser = await prisma.user.upsert({
    where: { email: SYSTEM_USER_EMAIL },
    create: {
      email: SYSTEM_USER_EMAIL,
      name: 'System',
      passwordHash: '!locked', // Cannot login — not a real password hash
      role: 'USER',
    },
    update: {},
  });

  // Upsert the system project
  const project = await prisma.project.upsert({
    where: { name: SYSTEM_PROJECT_NAME },
    create: {
      name: SYSTEM_PROJECT_NAME,
      description: 'System prompts for mcpctl gating and session management',
      prompt: '',
      gated: false,
      ownerId: systemUser.id,
    },
    update: {}, // Don't overwrite user edits to the project itself
  });

  // Upsert each system prompt (re-create if deleted, don't overwrite content if edited)
  for (const def of SYSTEM_PROMPTS) {
    const existing = await prisma.prompt.findFirst({
      where: { name: def.name, projectId: project.id },
    });

    if (!existing) {
      await prisma.prompt.create({
        data: {
          name: def.name,
          content: def.content,
          priority: def.priority,
          projectId: project.id,
        },
      });
    }
    // If the prompt exists, don't overwrite — user may have edited it
  }
}

/** Get the names of all system prompts (for delete protection). */
export function getSystemPromptNames(): string[] {
  return SYSTEM_PROMPTS.map((p) => p.name);
}

/** Get the required template variables for a system prompt (e.g., ['{{maxTokens}}']). */
export function getSystemPromptRequiredVars(name: string): string[] | undefined {
  const def = SYSTEM_PROMPTS.find((p) => p.name === name);
  return def?.requiredVars;
}

/** Get the default content for a system prompt (for reset-on-delete). */
export function getSystemPromptDefault(name: string): string | undefined {
  const def = SYSTEM_PROMPTS.find((p) => p.name === name);
  return def?.content;
}