MCP 服务器

MCP Drift State Tracker

MCP server that scans multi-language codebases to detect stubs, missing imports, and structural incompleteness, assigning drift scores to quantify LLM context erosion.

README

MCP Drift State Tracker

MCP server that enforces code completeness, intercepts context erosion, and neutralizes LLM output degradation across multi-language repository workspaces.

Pure TypeScript. JSON-driven language profiles. Zero Python. Zero runtime bloat.

Author

Jeremy Anderson — dcos.net

What It Does

LLMs degrade output quality as context windows fill: they drop imports, leave stub bodies, omit function implementations, and produce structurally incomplete code. This server provides a real-time audit layer over any MCP-compatible coding client, flagging those patterns before they persist in the codebase.

Core Capabilities

Multi-language parsing — 20+ languages supported via data-driven JSON profiles. No per-language parser classes.
Stub and placeholder detection — flags pass, TODO, FIXME, raise NotImplementedError, undefined, return null, and language-specific stub patterns.
Import and dependency tracking — verifies that declared imports correspond to symbols used in the file.
Structural completeness verification — confirms that declared classes, functions, and interfaces have non-empty bodies.
Drift scoring — assigns a numeric drift score to files and sessions to quantify context erosion.
State persistence — writes audit state to .mcp_drift_state.json at the repository root for cross-session continuity.

Architecture

The parsing engine is entirely data-driven. A single index.ts runtime loads language_profiles.json at startup, compiles all regex patterns once, and dispatches files to the correct parsing strategy via a flat extension-to-profile dispatch table. No cascading conditionals. No nested control flow.

index.ts                          MCP server runtime
├── language_profiles.json        language definitions, dispatch table, aliases
├── controller_config.json        downstream AI service endpoints
├── mcp_config.json               client integration reference config
└── Makefile                      standardized build and deploy targets

Language Support

Category	Extensions	Strategy
C-style / Brace	`.js` `.jsx` `.ts` `.tsx` `.c` `.cpp` `.h` `.hpp` `.java` `.cs` `.kt` `.kts` `.swift` `.dart` `.pike` `.pmod` `.r`	Brace-depth counting via `cStyleDefault`
Brace (custom)	`.rs` `.go` `.php` `.hack` `.sh` `.zsh` `.ps1`	Per-language regex overrides in `braceProfiles`
Block-end keywords	`.jl` `.lua` `.ex` `.exs` `.adb` `.ads` `.au3` `.vbs`	Start/end keyword pairing with nest-depth increment rules
Sequential labels	`.bat` `.cmd` `.sql`	Label-based or statement-delimited flow
Indentation-sensitive	`.mojo` `.hs` `.elm`	Signature detection with indent-level body verification

Duplicate profiles (.sh/.zsh/.ps1, .ex/.exs, .adb/.ads, .bat/.cmd, .php/.hack) resolve through the _aliases map in language_profiles.json. Adding a new language requires appending one entry to the dispatch table and restarting the server.

MCP Tools

Tool	Purpose
`scan_file`	Scan a single source file for drift indicators
`scan_workspace`	Recursively scan a workspace directory, respecting `globalIgnoreDirs`
`get_drift_report`	Retrieve accumulated drift state and scores
`reset_drift_state`	Clear drift state to begin a fresh audit session

Requirements

Node.js >= 20.0.0
npm (bundled with Node.js)

Installation

git clone https://git.dcos.net/dcosnet/mcp-drift-state-tracker.git
cd mcp-drift-state-tracker
make

Running

make dev       # development, no compile step
make prod      # compile and run
make inspect   # interactive MCP tool testing via Inspector

Configuration

`language_profiles.json`

All language-specific regex patterns, stub detectors, file extension routing, and ignore rules. Validated against the inline $schema. Key structure:

_dispatch — flat lookup table mapping every supported extension to its profile category and profile key. The runtime resolves extensions in O(1) without cascading conditionals.
_aliases — maps duplicate extensions to their canonical profile key.
_stubDefaults — common stub patterns applied to profiles that omit an explicit stubs array.
globalIgnoreDirs — directories skipped during workspace scans.
stateFileName — per-repo state file name.
cStyleDefault — fallback regex set for brace-matching languages.
braceProfiles / braceExtensions — custom overrides and extension lists.
blockEndProfiles — start/end keyword patterns.
sequentialProfiles — label-based flow definitions.
indentationProfiles — indent-sensitive language definitions.

`controller_config.json`

Downstream AI service endpoints. All ports are integers. All service keys are lowercase slugs. The base_dir follows FHS conventions.

Service	Port	Purpose
ollama	11434	Local inference engine
open-webui	8080	Browser-based LLM chat interface
aider-web	8501	Web frontend for Aider
hermes	8000	Orchestration proxy
odysseus	7000	Orchestration proxy
dify-engine	5001	Workflow automation platform