drift-analyzer 1.1.16

Deterministic architectural drift detection for AI-accelerated Python repositories through cross-file coherence analysis

Drift

Deterministic architecture erosion detection for AI-accelerated codebases

97.3% precision (single-rater) · 15 signals · deterministic · no LLM in pipeline · full study · docs

---

pip install drift-analyzer
drift analyze --repo .

╭─ drift analyze  myproject/ ──────────────────────────────────────────────────╮
│  DRIFT SCORE  0.52  Δ -0.031 ↓ improving  │  87 files  │  AI: 34%  │  2.1s │
╰──────────────────────────────────────────────────────────────────────────────╯

  Module                  Score  Bar                   Findings  Top Signal
  src/api/routes/          0.71  ██████████████░░░░░░       12   PFS 0.85
  src/services/auth/       0.58  ███████████░░░░░░░░░        7   AVS 0.72
  src/db/models/           0.41  ████████░░░░░░░░░░░░        4   MDS 0.61

  ◉ PFS  0.85  Error handling split 4 ways
               → src/api/routes.py:42
               → Next: consolidate into shared error handler

  ◉ AVS  0.72  DB import in API layer
               → src/api/auth.py:18
               → Next: move DB access behind service interface

Start here

Drift finds the architecture erosion AI-generated code silently introduces: pattern fragmentation, boundary violations, near-duplicate utilities, and structural hotspots that pass tests but weaken the codebase.

Designed for Python teams that want fast structural feedback without adding an LLM to the analysis path.

Three good ways to start

• Run it: Quick Start and Configuration
• Evaluate: Example Findings, Trust and Evidence, Stability
• Contribute: CONTRIBUTING.md, DEVELOPER.md, POLICY.md

CI (start report-only, tighten later)

- uses: sauremilk/drift@v1
  with:
    fail-on: none
    upload-sarif: "true"

Vibe-Coding Workflow

Built for AI-assisted sessions where an LLM writes most code and you steer.

drift scan --repo . --max-findings 5      # session start: agent learns baseline
drift diff --uncommitted                   # before commit
drift diff --staged-only                   # index only
drift diff --diff-ref main                 # compare against main
drift check --repo . --fail-on high        # CI gate

Each call returns accept_change: true | false with blocking reasons the agent can act on directly.

MCP integration

Drift can run as an MCP (Model Context Protocol) server so AI agents can call analysis tools directly over stdio.

Install MCP support:

pip install drift-analyzer[mcp]

Start the server:

drift mcp --serve

Minimal VS Code setup in .vscode/mcp.json:

{
  "servers": {
    "drift": {
      "type": "stdio",
      "command": "drift",
      "args": ["mcp", "--serve"]
    }
  }
}

Common agent-native calls:

drift scan --repo .
drift diff --staged-only
drift validate --repo .
drift fix-plan --repo .

See Integrations and API and Outputs for details.

CI

- run: drift check --repo . --fail-on high

Same signals, same deterministic engine — no LLM involved at analysis time.

Why teams use drift

Your linter, type checker, and test suite can tell you whether code is valid. They do not tell you whether the repository is quietly splitting into incompatible patterns across modules.

Drift focuses on that gap:

• Ruff / formatters / type checkers: local correctness and style, not cross-module coherence.
• Semgrep / CodeQL / security scanners: risky flows and policy violations, not architectural consistency.
• Maintainability dashboards: broad quality heuristics, not a drift-specific score with reproducible signal families.

Current public evidence: 15 real-world repositories in the study corpus, 15 scoring signals, and auto-calibration that rebalances weights at runtime. Full study → · Trust & limitations

Use cases

Pattern fragmentation in a connector layer

Problem: A FastAPI service has 4 connectors, each implementing error handling differently — bare except, custom exceptions, retry decorators, and silent fallbacks.

Solution:

drift analyze --repo . --sort-by impact --max-findings 5

Output: PFS finding with score 0.96 — "26 error_handling variants in connectors/" — shows exactly which files diverge and suggests consolidation.

Architecture boundary violation in a monorepo

Problem: A database model file imports directly from the API layer, creating a circular dependency that breaks test isolation.

Solution:

drift check --fail-on high

Output: AVS finding — "DB import in API layer at src/api/auth.py:18" — blocks the CI pipeline until the import direction is fixed.

Duplicate utility code from AI-generated scaffolding

Problem: AI code generation created 6 identical _run_async() helper functions across separate task files instead of finding the existing shared utility.

Solution:

drift analyze --repo . --format json | jq '.findings[] | select(.signal=="MDS")'

Output: MDS findings listing all 6 locations with similarity scores ≥ 0.95, enabling a single extract-to-shared-module refactoring.

Setup and rollout options

Full GitHub Action (recommended: start report-only)

name: Drift

on: [push, pull_request]

jobs:
  drift:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      security-events: write

    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - uses: sauremilk/drift@v1
        with:
          fail-on: none           # report findings without blocking CI
          upload-sarif: "true"    # findings appear as PR annotations

Once the team has reviewed findings for a few sprints, tighten the gate:

      - uses: sauremilk/drift@v1
        with:
          fail-on: high           # block only high-severity findings
          upload-sarif: "true"

CI gate (local)

drift check --fail-on none    # report-only
drift check --fail-on high    # block on high-severity findings

pre-commit hook

The fastest way to add drift to your workflow:

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/sauremilk/drift
    rev: v0.10.2
    hooks:
      - id: drift-check          # blocks on high-severity findings
      # - id: drift-report        # report-only alternative (start here)

Or use a local hook if you already have drift installed:

# .pre-commit-config.yaml
repos:
  - repo: local
    hooks:
      - id: drift
        name: drift
        entry: drift check --fail-on high
        language: system
        pass_filenames: false
        always_run: true

More setup paths:

• Quick Start
• Configuration
• Team Rollout

If you want example findings before integrating, start with docs-site/product/example-findings.md.

What you get

╭─ drift analyze  myproject/ ──────────────────────────────────────────────────╮
│  DRIFT SCORE  0.52  Δ -0.031 ↓ improving  │  87 files  │  AI: 34%  │  2.1s │
╰──────────────────────────────────────────────────────────────────────────────╯
  Trend: 0.551 → 0.548 → 0.520 (3 snapshots)

                        Module Drift Ranking
  Module                        Score  Bar                   Findings  Top Signal
  ─────────────────────────────────────────────────────────────────────────────────
  src/api/routes/                0.71  ██████████████░░░░░░       12   PFS 0.85
  src/services/auth/             0.58  ███████████░░░░░░░░░        7   AVS 0.72
  src/db/models/                 0.41  ████████░░░░░░░░░░░░        4   MDS 0.61

┌──┬────────┬───────┬──────────────────────────────────────┬──────────────────────┐
│  │ Signal │ Score │ Title                                │ Location             │
├──┼────────┼───────┼──────────────────────────────────────┼──────────────────────┤
│◉ │ PFS    │  0.85 │ Error handling split 4 ways          │ src/api/routes.py:42 │
│◉ │ AVS    │  0.72 │ DB import in API layer               │ src/api/auth.py:18   │
│○ │ MDS    │  0.61 │ 3 near-identical validators          │ src/utils/valid.py   │
└──┴────────┴───────┴──────────────────────────────────────┴──────────────────────┘

Drift scores 15 signal families. For the full list, weights, and scoring details, see:

• Signal Reference
• Algorithm Deep Dive
• Scoring Model

How drift compares

Data sourced from STUDY.md §9 and benchmark_results/.

Capability	drift	SonarQube	pylint / mypy	jscpd / CPD
Pattern Fragmentation across modules	Yes	No	No	No
Near-Duplicate Detection	Yes	Partial (text)	No	Yes (text)
Architecture Violation signals	Yes	Partial	No	No
Temporal / change-history signals	Yes	No	No	No
GitHub Code Scanning via SARIF	Yes	Yes	No	No
Zero server setup	Yes	No	Partial	Yes
TypeScript Support	Optional ¹	Yes	No	Yes

¹ Experimental via drift-analyzer[typescript]. Python is the primary target.

Drift is designed to complement linters and security scanners, not replace them. Recommended stack: linter (style) + type checker (types) + drift (coherence) + security scanner (SAST).

Full comparison: STUDY.md §9 — Tool Landscape Comparison

Is drift a good fit?

Drift is a strong fit for:

• Python teams using AI coding tools in repositories where architecture matters
• repositories with 20+ files and recurring refactors across modules
• teams that want deterministic architectural feedback in local runs and CI

Wait or start more cautiously if:

• the repository is tiny and a few findings would dominate the score
• you need bug finding, security review, or type-safety enforcement rather than structural analysis
• Python 3.11+ is not available in your local and CI execution path yet

The safest rollout path is progressive:

1. Start with drift analyze locally and review the top findings.
2. Add drift check --fail-on none in CI as report-only discipline.
3. Gate only on high findings once the team understands the output.
4. Ignore generated or vendor code and tune config only after reviewing real findings in your repo.

Recommended guides:

• Team Rollout
• Finding Triage
• Benchmarking and Trust

Trust and limitations

Public claims safe to repeat today: Drift is deterministic, benchmarked on 15 real-world repositories in the current study corpus, and uses 15 scoring signals with auto-calibration for runtime weight rebalancing and small-repo noise suppression.

What's limited: Benchmark validation is single-rater; not yet independently replicated. Small repos can be noisy. Temporal signals depend on clone depth. The composite score is orientation, not a verdict.

What's next: Independent external validation, multi-rater ground truth, signal-specific confidence intervals.

Drift is designed to earn trust through determinism and reproducibility:

• no LLMs in the detection pipeline
• reproducible CLI and CI output
• signal-specific interpretation instead of score-only messaging
• explicit benchmarking and known-limitations documentation

Interpreting the score

The drift score measures structural entropy, not code quality. Keep these principles in mind:

• Interpret deltas, not snapshots. Use drift trend to track changes over time. A single score in isolation has limited meaning.
• Temporary increases are expected during migrations. Two coexisting patterns (old and new) will raise PFS/MDS signals. This is the migration happening, not a problem.
• Deliberate polymorphism is not erosion. Strategy, Adapter, and Plugin patterns produce structural similarity that MDS flags as duplication. Findings include a deliberate_pattern_risk hint — verify intent before acting.
• The score rewards reduction, not correctness. Deleting code lowers the score just like refactoring does. Do not optimize for a low score — optimize for understood, intentional structure.

For a detailed discussion of epistemological boundaries (what drift can and cannot see), see STUDY.md §14.

Drift vs. erosion: Without layer_boundaries in drift.yaml, drift detects emergent drift — structural patterns that diverge without explicit prohibition. With configured layer_boundaries, drift additionally performs conformance checking against a defined architecture. Both modes are complementary: drift does not replace dedicated architecture conformance frameworks (e.g. PyTestArch for executable layer rules in pytest), but catches cross-file coherence issues those tools do not model.

Start with the strongest, most actionable findings first. If a signal is noisy for your repository shape, tune or de-emphasize it instead of forcing an early hard gate.

Further reading:

• Benchmarking and Trust
• Full Study
• Case Studies

Release status

The PyPI classifier remains Development Status :: 3 - Alpha intentionally.

That is a conservative release signal, not a claim that the core workflow is unusable. The strongest path today is the deterministic Python analysis and report-only CI rollout; some adjacent surfaces remain intentionally marked as experimental.

Current release posture:

• core Python analysis: stable
• CI and SARIF workflow: stable
• TypeScript support: experimental
• embeddings-based parts: optional / experimental
• benchmark methodology: evolving

Full rationale and matrix: Stability and Release Status

Contributing

Drift's biggest blind spots are found by people running it on codebases the maintainers have never seen. Your real-world experience is a direct contribution to signal quality — whether you write code or not.

If Drift surprised you with an unexpected result, that's valuable feedback: open an issue or start a discussion. A well-documented false positive can be more valuable than a new feature.

I want to…	Go here
Ask a usage question	Discussions
Report a false positive / false negative	FP/FN template
Report a bug	Bug report
Suggest a feature	Feature request
Propose a contribution before coding	Contribution proposal
Report a security vulnerability	SECURITY.md — not a public issue

New here? Start contributing

You don't need to understand the whole analyzer to help. Start at the level that fits your time:

1. 15 min: Fix a typo or clarify a docs example → open a PR directly
2. 30 min: Report an unexpected finding with reproduction steps → FP/FN template
3. 1 hour: Add an edge-case test → pick a good first issue
4. 2+ hours: Improve signal logic or finding explanations → see CONTRIBUTING.md

git clone https://github.com/sauremilk/drift.git && cd drift && make install
make test-fast    # confirm everything passes, then start

First contribution? We'll help you scope it. Open a contribution proposal or ask in Discussions if you're unsure where to start.

Typical first contributions:

• Report a false positive or false negative with reproduction steps
• Add a ground-truth fixture for a signal edge case
• Improve a finding's explanation text to be more actionable
• Write a test for an untested edge case
• Clarify docs or add a configuration example

What we value most: reproducibility, explainability, false-alarm reduction. What we deprioritize: new output formats without insight value, comfort features, complexity without analysis improvement.

See CONTRIBUTING.md for the full guide, contributor types, and the contribution ladder. See ROADMAP.md for current priorities.

Documentation map

• Getting Started
• How It Works
• Benchmarking and Trust
• Product Strategy
• Contributor Guide
• Developer Guide

License

MIT. See LICENSE.