Skip to content
memoDwarves Foundation

Reports/Architecture

MCP playbook code flow

A detailed look into the mcp-playbook's internal code flow, data handling for chat logs and prompts, and interactions with GitHub repositories like prompt-db and prompt-log. This technical deep-dive helps developers understand the implementation details for debugging and extending functionality.

By monotykamary13 May 20254 min read

tl;dr

The MCP Playbook server processes tool calls via src/index.ts, routing them to handlers in src/handlers/. These handlers interact with the local file system and GitHub (e.g., prompt-db, prompt-log). Chat log syncing involves editor-specific parsers and uploads to a structured path in prompt-log, while prompt syncing targets prompt-db.

This document provides a detailed walkthrough of the mcp-playbook server's internal code execution, its data flow, particularly concerning chat log and prompt synchronization, and its interactions with key GitHub repositories.

Core request handling and tool execution

The fundamental operation of the mcp-playbook server revolves around receiving Model Context Protocol (MCP) requests and dispatching them to appropriate handlers for execution.

  1. Entry point (src/index.ts):

    • This file initializes the MCP server and defines handlers for various request types (CallToolRequestSchema, ListToolsRequestSchema, etc.).
    • When a CallToolRequest is received, src/index.ts extracts the toolName and arguments.
    • A switch statement on toolName routes the request to the corresponding function in src/handlers.ts (which aggregates handlers from the src/handlers/ directory).

  2. Tool handlers (src/handlers/):

    • Each tool (e.g., create_spec, sync_prompt, save_and_upload_chat_log) has a dedicated handler file in this directory (e.g., handleCreateSpec.ts).
    • These handlers contain the core logic for the tool, utilizing arguments passed from src/index.ts.
    • Interactions include:
      • File system operations within the target_project_dir.
      • API calls to GitHub.

Detailed data flow diagrams

The interaction between the LLM, MCP Playbook server, and external systems can be broken down into several key areas:

MCP server core logic and documentation tools

This diagram shows the initial request handling and the flow for tools that primarily interact with the local file system for documentation purposes.

graph TD
    A["LLM / MCP Client"] --> B["AI Conversation & Tool Calls"]
    B --> F["MCP Playbook Server (src/index.ts)"]

    subgraph "MCP Playbook Server Logic - Core"
        F --> ROUTER["Switch(toolName)"]
        ROUTER --> H_INIT["handleInitPlaybook"]
        ROUTER --> H_DOCS["handleCreateSpec/Adr/Changelog"]
        H_DOCS --> FS_DOCS["Local FS (target_project_dir/docs/...)\n(ADRs, Specs, Changelogs)"]
        %% Removed other handlers for focus %%
    end

Chat log synchronization flow

This diagram details the path for the save_and_upload_chat_log tool, from parsing editor-specific logs to uploading them to dwarvesf/prompt-log.

graph TD
    F["MCP Playbook Server (src/index.ts)"] --> ROUTER["Switch(toolName)"]
    ROUTER --> H_SAVE_LOG["handleSaveAndUploadChatLog (src/handlers/handleSaveAndUploadChatLog.ts)"]

    subgraph "Chat Log Processing"
        H_SAVE_LOG --> PARSERS["Editor Parsers (src/handlers/parser/) (Cline, Cursor, Zed)"]
        PARSERS --> FS_CHAT_LOCAL["Local FS Save (target_project_dir/.chat/chat_log_user_proj_ts.md)"]
        FS_CHAT_LOCAL --> H_SAVE_LOG_UPLOAD["Upload Logic in handleSaveAndUploadChatLog"]
    end

    subgraph "GitHub Repository: prompt-log"
      H_SAVE_LOG_UPLOAD --> G_PLOG["dwarvesf/prompt-log"]
      G_PLOG --> G_PLOG_STRUCTURE["project-logs/projectName/userId/.chat/chat_log_ts.md"]
      G_PLOG --> G_PLOG_UNSTRUCTURED["unstructured/ (Potentially for other log types)"]
    end

Prompt and runbook synchronization flow

This diagram illustrates how prompts are synced to dwarvesf/prompt-db and how runbook-related tools interact with dwarvesf/runbook.

graph TD
    F["MCP Playbook Server (src/index.ts)"] --> ROUTER["Switch(toolName)"]

    subgraph "Prompt & Runbook Tool Logic"
        ROUTER --> H_SYNC_PROMPT["handleSyncPrompt (src/handlers/handleSyncPrompt.ts)"]
        ROUTER --> H_SEARCH_P["handleSearchPrompts"]
        ROUTER --> H_SUGGEST_RB["handleSuggestRunbook"]
        ROUTER --> H_SEARCH_RB["handleSearchRunbook"]
    end

    subgraph "GitHub Repository: prompt-db"
        H_SYNC_PROMPT --> G_PDB_SYNC_TARGET["dwarvesf/prompt-db"]
        G_PDB_SYNC_TARGET --> G_PDB_SYNCED[".synced_prompts/projectName/promptName.md"]

        H_SEARCH_P --> G_PDB_SEARCH_TARGET["dwarvesf/prompt-db"]
        G_PDB_SEARCH_TARGET --> G_PDB_CATEGORIES["Categorized Prompts (coding/, general/, etc.)"]
    end

    subgraph "GitHub Repository: runbook"
        H_SUGGEST_RB --> G_RB["dwarvesf/runbook"]
        H_SEARCH_RB --> G_RB
        G_RB --> G_RB_CONTENT["Runbook Content (e.g., technical-patterns/, operational-state-reporting/)"]
    end

    %% Connection for context from file exploration for sync_prompt (optional to include here or assume from main text)
    %% B["AI Conversation & Tool Calls"] --> C{"Explore Files in target_project_dir"}
    %% C -->|Files with Prompts| D["Prompts in Files"]
    %% D -->|Provides Context for sync_prompt| H_SYNC_PROMPT

Chat log synchronization (save_and_upload_chat_log)

The save_and_upload_chat_log tool has a specific flow for handling conversation histories from different AI-assisted coding editors.

  1. Invocation: The tool is called with targetProjectDir, userId, and editorType (e.g., 'cline', 'cursor', 'zed').
  2. Parsing (src/handlers/parser/):
    • Based on editorType, handleSaveAndUploadChatLog.ts invokes a specific parser:
      • clineChatParser.ts: Reads Cline's SQLite database (~/.local/share/cline/history.db or platform equivalent) to extract messages, timestamps, and potentially associated project paths.
      • cursorChatParser.ts: Likely parses Cursor's internal storage format (often JSON or similar flat files within its application support directory, e.g., ~/.cursor/conversations/ or ~/Library/Application Support/Cursor/conversations/) to get conversation turns, code snippets, and project context.
      • zedChatParser.ts: Reads Zed's conversation log files, which might be stored in a structured format (e.g., JSONL) within Zed's application support directory (e.g., ~/Library/Application Support/Zed/conversations/), extracting messages and associated context.
    • Each parser aims to extract a structured representation of the conversation, including messages, roles (user/assistant), timestamps, and ideally, the project name or path the conversation was associated with.
    • parserUtils.ts may contain shared utility functions for these parsers, such as common path resolution or data cleaning.
  3. Formatting: The retrieved history is then formatted into a standardized Markdown representation by an editor-specific formatting function (e.g., formatCursorHistory).
  4. Local save:
    • The formatted Markdown content is saved locally within the target_project_dir at: .chat/chat_log_<safeUserId>_<safeProjectName>_<timestamp>.md.
    • <safeProjectName> is derived by the parser; if it cannot be determined, it defaults to "unknown-project".
  5. GitHub upload:
    • The same Markdown content is uploaded to the dwarvesf/prompt-log repository.
    • The path in the repository is: project-logs/<safeProjectName>/<safeUserId>/.chat/chat_log_<timestamp>.md.
    • This structured path allows for organized storage and retrieval based on project and user.

This detailed flow ensures that conversation logs are captured from various sources, standardized, and archived in a discoverable manner.

Prompt synchronization (sync_prompt)

The sync_prompt tool facilitates saving and versioning LLM prompts.

  1. Invocation: Called with projectName, promptName, and promptContent.
  2. GitHub interaction:
    • The tool directly interacts with the dwarvesf/prompt-db GitHub repository.
    • It creates or updates a file at the path: .synced_prompts/<projectName>/<promptName>.md.
    • The content of this file is the promptContent.

This allows for a centralized and version-controlled database of prompts used across different projects and by various team members. The search_prompts tool, in contrast, searches the categorized directories like coding/, general/, etc., within prompt-db, intentionally excluding the .synced_prompts/ directory which acts more as a direct, raw sync target.

This detailed code and data flow provides clarity on how mcp-playbook operates internally and manages the crucial task of knowledge capture and dissemination in an AI-augmented development environment.

  • Next: Back to documentation index