dcosnet
/
MCP-Drift-State-Tracker
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
MCP-Drift-State-Tracker/BLOG.md

8.8 KiB
Executable File
Raw Blame History

Integrating MCP Drift State Tracker with Your AI Coding Stack

Integration guides for wiring the tracker into production tooling. All configs derive from controller_config.json and mcp_config.json — these reflect the actual service topology the tracker targets.

1. Odysseus

Local AI coding orchestration proxy on port 7000 (controller_config.json). Odysseus sits between the LLM backend (Ollama) and the coding frontend, managing tool calls, context windows, and session state.

Drift tracking is critical here. Odysseus manages long coding sessions where the LLM iterates on files across dozens of turns. By turn 30, the model drops imports, leaves TODO stubs, and produces structurally incomplete code. The tracker operates as a post-edit audit gate.

Config

{
  "mcpServers": {
    "mcp-drift-state-tracker": {
      "command": "node",
      "args": ["${MCP_TRACKER_PATH}/dist/index.js"],
      "env": {
        "MCP_CONTROLLER_PORT": "7000",
        "MCP_CONTROLLER_MODEL": "ollama/llama3"
      }
    }
  }
}

Workflow

Invoke scan_workspace on the target directory after each significant edit cycle. When the drift score exceeds the configured threshold, Odysseus re-prompts the LLM to correct the flagged files. This creates a closed-loop quality gate: the model writes code, the tracker audits it, Odysseus feeds the audit results back into the next generation pass.

Recommended System Prompt Addition

After editing more than 5 files in a session, run get_drift_report and fix every file with a drift score above 0.3 before concluding.

2. Hermes

Orchestration proxy on port 8000. Hermes handles focused, single-task coding with strict tool governance.

Config

{
  "mcpServers": {
    "mcp-drift-state-tracker": {
      "command": "node",
      "args": ["${MCP_TRACKER_PATH}/dist/index.js"],
      "env": {
        "MCP_CONTROLLER_PORT": "8000",
        "MCP_CONTROLLER_MODEL": "ollama/llama3"
      }
    }
  }
}

Workflow

Register scan_file as a post-write hook in Hermes's tool registry. Every time Hermes instructs the LLM to write or modify a file, it immediately follows with scan_file on that path. When stubs or missing imports are detected, Hermes re-issues the write instruction with the audit feedback appended to the context.

This is lighter-weight than scan_workspace and aligns with Hermes's task-oriented design. Reserve scan_workspace for session-end review.

3. Open WebUI

Browser-based LLM interface on port 8080. Provides Ollama with workspace management, model selection, and MCP server support via its Functions/Tools system.

Config

Navigate to Admin Settings > Tools > MCP Servers and add:

Field Value
Name mcp-drift-state-tracker
Command node
Args ${MCP_TRACKER_PATH}/dist/index.js

Or add to Open WebUI's functions.yaml:

mcp_servers:
  mcp-drift-state-tracker:
    command: node
    args:
      - "${MCP_TRACKER_PATH}/dist/index.js"

Workflow

Open WebUI exposes the tracker directly in the chat interface. After the LLM completes a multi-file change, the user types "check my code for drift" — this triggers scan_workspace via the MCP tool and returns a structured report in the chat.

Recommended System Prompt Addition

After completing any multi-file code change, automatically run scan_workspace on the project directory and present the drift report.

4. Claude Desktop / Claude Code

Anthropic's desktop and CLI clients. MCP servers are first-class citizens.

Config

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-drift-state-tracker": {
      "command": "node",
      "args": ["${MCP_TRACKER_PATH}/dist/index.js"]
    }
  }
}

Workflow

Claude maintains strong context in single-session coding. Drift appears in long refactoring sessions regardless. Run scan_file after each file write and scan_workspace at session end. Claude interprets drift reports and corrects flagged stubs on the first retry pass.

5. Cursor / Windsurf / VS Code (Continue.dev)

IDE-native AI coding assistants. Cursor and Windsurf provide built-in MCP support. Continue.dev adds MCP integration to vanilla VS Code.

Config (Cursor / Windsurf)

Add to the IDE's MCP settings (.cursor/mcp.json or the MCP settings panel):

{
  "mcpServers": {
    "mcp-drift-state-tracker": {
      "command": "node",
      "args": ["${MCP_TRACKER_PATH}/dist/index.js"]
    }
  }
}

Config (Continue.dev)

{
  "mcp": {
    "servers": {
      "mcp-drift-state-tracker": {
        "command": "node",
        "args": ["${MCP_TRACKER_PATH}/dist/index.js"]
      }
    }
  }
}

Workflow

IDE integrations provide the tightest feedback loop. When the AI assistant writes a file, configure a post-save hook to run scan_file on the current buffer. Drift results appear inline or in a side panel. Incomplete code is catched before it enters version control.

6. Dify Engine

Workflow automation platform for LLM applications on port 5001. Dify builds multi-step AI pipelines with conditional branching, tool calls, and human-in-the-loop stages.

Integration Method

Dify does not consume MCP servers natively. Wrap the tracker as an external tool via Dify's API Tool node.

  1. Host the tracker as a persistent process: 
    node ${MCP_TRACKER_PATH}/dist/index.js
  2. In Dify, create an HTTP Request tool node targeting the tracker's MCP transport endpoint. Stdio-based servers require an adapter such as mcp-proxy to expose them over HTTP.
  3. Alternatively, define a Dify Function Tool that shells out to the tracker: 
    import subprocess
import json

def scan_workspace(path: str) -> dict:
    result = subprocess.run(
        ["node", "${MCP_TRACKER_PATH}/dist/index.js"],
        input=json.dumps({"tool": "scan_workspace", "arguments": {"path": path}}),
        capture_output=True,
        text=True
    )
    return json.loads(result.stdout)

Workflow

Place the drift check as a conditional gate between the Code Generation step and the Output / Commit step. When scan_workspace returns a drift score above the configured threshold, route back to the generation step with the audit feedback injected into the prompt. This creates an automated quality loop without human intervention.

7. Aider

Terminal-based AI pair programmer with MCP support.

Config

CLI flag:

aider --mcp-server "mcp-drift-state-tracker:node:${MCP_TRACKER_PATH}/dist/index.js"

Or in .aider.conf.yml:

mcp-servers:
  mcp-drift-state-tracker:
    command: node
    args: ["${MCP_TRACKER_PATH}/dist/index.js"]

Workflow

Aider operates in long terminal sessions with large repos and is prone to context erosion. The tracker catches semantic incompleteness that static linters miss — a function that compiles but contains a pass body, for example. Run scan_file on every file Aider touches. Run scan_workspace every 10 commits to catch drift in files the model edited earlier in the session.

8. Cline (VS Code Extension)

Autonomous AI coding agent for VS Code. Creates files, runs commands, and manages entire development tasks.

Config

{
  "mcpServers": {
    "mcp-drift-state-tracker": {
      "command": "node",
      "args": ["${MCP_TRACKER_PATH}/dist/index.js"]
    }
  }
}

Workflow

Cline's autonomy makes drift tracking essential. A single task like "implement user authentication across the backend" creates 8-10 files in one run. Configure Cline to run scan_workspace as its final step before marking a task complete. Files with high drift scores are flagged for a correction pass.

Scan Strategy

Trigger Tool Rationale
After every file edit scan_file Immediate feedback, lowest latency
After multi-file refactor scan_workspace Catches cross-file drift
Session end get_drift_report Summary for human review
New session start get_drift_report Establishes baseline from prior state
Pre-commit scan_workspace Quality gate on every commit

Extending Language Profiles

Add project-specific stub patterns to the relevant language profile's stubs array in language_profiles.json:

".ts": {
  "stubs": ["pass", "TODO", "FIXME", "throw new Error(\"not implemented\")", "// @ts-ignore"]
}

Relationship to Static Analysis

The tracker does not replace ESLint, Ruff, Clippy, or any static analysis tool. It targets a distinct failure class: semantic incompleteness that compiles but represents degraded LLM output. Run static analysis for correctness. Run drift tracking for completeness. Both are necessary.

All integration configs reference the service ports and model names from controller_config.json. Set MCP_TRACKER_PATH to the absolute path of this repository and all configs resolve correctly.