v0.6.0 — Reasoning engine

ClaudeDrift audits your project's Claude artifacts — CLAUDE.md, skills, agents, commands, hooks, .mcp.json — against the actual codebase, finds where they've drifted, and offers to fix them. Run on demand with /claude-drift:drift-check.

⚡ This release: reasoning, not regex

ClaudeDrift no longer uses a deterministic script. Validated across ~40 real repos, a regex/heuristic scanner was an endless false-positive catch-up — every new repo needed another rule. The engine is now Claude's own reasoning, via subagents, which judges understanding the artifact instead of pattern-matching it.

A regex can't tell a real path from an illustrative example, a sub-package/dependency path, a created-output path, prose where a slash means "and" (CUDA/Metal), or a renamed command in a generic template. Reasoning can.

Architecture

• claude-flow-mapper (new) — determines the project's Claude flow: discovers every artifact, reads the codebase to understand the real stack/layout/workflow, notes dependency-install state, and ranks each artifact by drift-likelihood so the deep audit is prioritized, not N blind agents.
• drift-auditor (rewritten) — reasons about one artifact vs the real code (reference and context drift), returning evidenced findings with exact old → new fixes.
• drift-check skill (rewritten) — orchestrates map → audit → synthesize → apply, purely with subagents.

Removed

• scripts/discover.mjs (~1000-line scanner) and its 53-case test suite.
• Script-only modes: --ci / --baseline / --preflight / --report / --merge.

Proven on real repos

• PlatformPlatform (39 artifacts): found a hallucinated MCP tool name (tabs_context_mcp) in an agent — a token that looks exactly like a real tool — by cross-referencing .mcp.json + settings + the repo, while correctly leaving the genuine capability alone. The flow-mapper triaged 39 artifacts down to 3 deep audits.
• quant.cpp: caught a gitignored-and-absent refs/ tree cited as required reading by every agent, plus measured context drift (WASM 192KB → ~295KB, header 15K → 17.7K LOC) — things no scanner can know.
• ragnasync: surfaced a real Linux-CI-breaking case-mismatch while not flagging a prose false positive (deploy/pulse).

Install

/plugin marketplace add marky291/ClaudeDrift
/plugin install claude-drift

Or try locally: claude --plugin-dir /path/to/ClaudeDrift

Footprint: nothing runs until you invoke the command — no MCP server, no always-on hooks. It spends model tokens proportional to your artifact count, only when you run it.