DeployProof

Deterministic pre-push verification for AI-assisted codebases: AST mutation testing, credential scanning, sandbox-escape detection, mock-usage alerts, swallowed-exception checks, and dependency hallucination defense.

Why This Exists

AI-assisted development introduces subtle failure modes that standard linters and coverage tools miss: test suites with high line coverage but near-zero mutation scores, hardcoded credentials generated in passing, symlinks that deceive approval prompts into escaping the repository sandbox, silently swallowed exceptions, and package names hallucinated by LLMs that don't exist on PyPI. DeployProof catches these at the pre-push stage, before they reach CI or production.

Privacy & Security Guarantee: DeployProof runs 100% locally on your machine. It makes zero outbound network calls, except for querying the official PyPI registry (JSON API) to verify that newly introduced dependencies exist and are not hallucinated. DeployProof sends no source code, telemetry, test results, or secret findings to any external server.

Install

pip install deployproof

Requires Python 3.10+.

How to Use DeployProof

DeployProof provides two distinct modes of operation. We believe in total transparency about execution costs:

1. Diff-Scoped Pre-Push Gate (deployproof check) — Primary Fast Workflow

Evaluates only the files modified in your active session or git diff (1–3 files typically). Because only newly edited AST nodes are mutated, it executes in 2 to 5 seconds in local developer loops, pre-commit hooks, and pre-push gates.

# Fast daily check: verifies modified files in your current working tree / git diff (2-5s)
deployproof check

# Output structured JSON for automation or IDE tooling
deployproof check --json

# Enforce strict gates on newly introduced mocks or swallowed errors
deployproof check --strict-mocks --strict-error-handling

2. Full Repository Audit Mode (deployproof check --full-repo) — Thorough Codebase Audits

Evaluates every tracked Python file across the entire repository. Because mutation testing generates hundreds or thousands of mutants and runs the test suite against each one in isolated multi-worker sandboxes, full repository scans take real, honest compute time:

# Run a full-repository audit across all tracked files
deployproof check --full-repo

# Customize the parallel worker process count (defaults to auto-detected CPU count capped at 8)
deployproof check --full-repo --workers 8

Real-World Timing Expectations:

Parallel Multi-Worker Sandboxing (--workers N)

DeployProof includes a built-in multi-process execution engine (ProcessPoolExecutor) for parallel mutation testing. You can supply --workers <N> to both diff-scoped checks and full-repo audits:

# Parallelize a large uncommitted diff across 8 worker processes
deployproof check --workers 8

# Full repository audit distributed across 8 worker processes
deployproof check --full-repo --workers 8

How Parallel Sandboxing Works:

1. Snapshot Creation: DeployProof takes an initial atomic snapshot of your project into a clean temporary directory.
2. PID-Keyed Sandboxes: Each worker process receives its own dedicated filesystem sandbox (worker_<PID>), with an independent pytest cache (--override-ini=cache_dir=...) and separate temporary directory (--basetemp=...).
3. Zero Mutation Leaks: Mutants are generated and executed inside individual worker sandboxes in parallel. The working tree is untouched, and signal handlers (SIGINT, SIGTERM, SIGBREAK) ensure sandboxes are cleanly purged upon completion or interruption.

Advantages & When to Use:

• Large Diff / Full Repo Speedup: Near-linear execution scaling across CPU cores for batches with 50+ mutants, cutting 30-minute sweeps down to 5–8 minutes.
• Process & Cache Isolation: Eliminates cross-test state leakage, shared database locking, and .pytest_cache collisions between concurrent workers.

Trade-offs & When NOT to Use:

• Small Daily Diffs (1–3 files): Do NOT use --workers for small 2-line edits. Spawning isolated sandboxes and copying file trees incurs ~1–2 seconds of snapshot overhead. Sequential in-place mutation executes in 2–5 seconds with zero snapshot overhead.
• Disk Space & I/O Overhead: Running $N$ workers copies the repository snapshot $N$ times into temporary storage ($N \times \text{repo size}$ in tempfile.gettempdir()). On disk-constrained environments, use fewer workers (e.g. --workers 2 or --workers 4).
• Subprocess / Port Collisions: If your test suite binds to fixed network ports (e.g. localhost:8080) without dynamic port selection, parallel workers running tests concurrently may trigger port conflicts. Use isolated ports or run sequentially in such environments.

Output includes a section for each check — symlink scan, secrets scan, dependency scan, mock detection, control flow analysis, and mutation score — with a pass/fail line at the bottom.

Exit Codes:

• 0 — PASSED: All verification checks passed and mutation score meets threshold.
• 1 — FAILED: Code quality or security gate triggered (mutation score below threshold, untested files, hardcoded secrets, sandbox-escape symlinks, hallucinated packages, or strict flags).
• 2 — ERROR: Test environment failure (test suite failed to collect or execute before mutation testing began due to missing dependencies or broken imports).

Example Walkthrough

Given a newly written function calculator.py:

def calculate_discount(price: float, rate: float) -> float:
    if rate > 0.5:
        return price * 0.5
    return price * (1.0 - rate)

With an AI-generated test that achieves 100% line coverage by only asserting standard discounts (rate = 0.2):

def test_calculate_discount_basic():
    assert calculate_discount(100.0, 0.2) == 80.0

Running deployproof check mutates AST operators and detects that boundary conditions and threshold caps are untested:

$ deployproof check

DeployProof - LOCAL PRE-CHECK (approximate) - not the verified score
====================================================================

Target Scope (1 file evaluated):
  * calculator.py

Symlink & Sandbox Escape Scan (CWE-61/CWE-451):
  Clean: No symlinks or sandbox-escape traversal links detected across 1 session file.

Secrets & Credentials Pre-Push Scan:
  Clean: No hardcoded secrets or tracked .env files detected across 1 session file.

Dependency & Slopsquatting Scan (PyPI Registry & Age Analysis):
  Clean: No new external packages introduced across 1 session file.

Mock Usage Introduced (flagged for review):
  Clean: No modified test files in scope.

Control Flow & Error Handling (flagged for review):
  Clean: No bare excepts, swallowed exceptions, or unreachable code detected across 1 session file.

Local Pre-Check Mutation Verification:
  Score:  57.1% (4/7 mutants killed)
  Status: FAILED (score 57.1% below 80.0%) (threshold: 80.0%)
  Time:   2.27s

Skipped Constructs: None (No known unsupported constructs detected)

Surviving Mutants (3 unverified changes):

  [1] calculator.py:2
      Mutation: Replace numeric constant '0.5' with '1.5'
      Original: if rate > 0.5:
      Mutated:  if rate > 1.5:

  [2] calculator.py:3
      Mutation: Replace numeric constant '0.5' with '1.5'
      Original: return price * 0.5
      Mutated:  return price * 1.5

  [3] calculator.py:3
      Mutation: Replace binary operator '*' with '/'
      Original: return price * 0.5
      Mutated:  return price / 0.5

====================================================================
Notice: Local pre-check only. Full verified score runs in CI on push (via mutmut).
Pre-check FAILED: Score 57.1% is below threshold 80.0% (3 surviving mutants).

Adding tests for threshold cap (rate = 0.8) and exact boundary (rate = 0.5) kills all mutants:

Local Pre-Check Mutation Verification:
  Score:  100.0% (7/7 mutants killed)
  Status: PASSED (threshold: 80.0%)
  Time:   2.31s

Surviving Mutants: None (All generated mutants caught by test suite)

====================================================================
Notice: Local pre-check only. Full verified score runs in CI on push (via mutmut).
Pre-check clean: 100% of tested basic mutations caught.

CLI Options & Flags

Note: deployproof check --wsl (Windows only) is newer and less battle-tested than the core checks — file an issue if you hit something.

Machine-Readable Output (--json)

DeployProof provides a stable structured JSON schema for CI/CD pipelines, IDEs, and automation:

deployproof check --json

JSON Output Schema

{
  "version": "1.0.0",
  "status": "passed",
  "summary": {
    "target_files_count": 1,
    "mutation_score": 100.0,
    "threshold": 80.0,
    "secrets_found": 0,
    "symlink_escapes_found": 0,
    "dependency_findings": {
      "high_risk": 0,
      "medium_risk": 0,
      "ok": 1,
      "unknown": 0,
      "unscanned": 0
    },
    "mock_usages_found": 0,
    "control_flow_findings": 0,
    "strict_mocks_active": false,
    "strict_mocks_triggered": false,
    "strict_error_handling_active": false,
    "strict_error_handling_triggered": false
  },
  "scope": {
    "target_files": [
      {
        "file": "app.py",
        "loc": 45,
        "is_large": false
      }
    ]
  },
  "mutation_testing": {
    "score": 100.0,
    "threshold": 80.0,
    "total_mutants": 6,
    "killed_mutants": 6,
    "survived_mutants_count": 0,
    "duration_seconds": 1.2,
    "surviving_mutants": [],
    "skipped_constructs": [],
    "untested_files": []
  },
  "secrets": {
    "clean": true,
    "files_scanned": 1,
    "findings": []
  },
  "symlinks": {
    "clean": true,
    "files_scanned": 1,
    "findings": []
  },
  "dependencies": {
    "clean": true,
    "total_scanned": 1,
    "findings": [],
    "unscanned_sources": []
  },
  "mocks": {
    "clean": true,
    "strict_gate_triggered": false,
    "findings": []
  },
  "control_flow": {
    "clean": true,
    "strict_gate_triggered": false,
    "findings": []
  }
}

What It Checks

• Mutation Score — Mutates AST operators (>=, ==, and, or, *, numeric constants, comparisons) in modified files and runs your test suite against each mutant. Reports surviving mutants and a percentage score. Does not use line coverage. Features atomic file restoration protected by SIGINT/SIGTERM/SIGBREAK signal handlers to ensure interrupted runs never leave mutated code on disk.
• Secrets and Credentials — Scans modified files for hardcoded API keys (OpenAI, Anthropic, AWS, GitHub, Stripe, private keys) and tracked .env files using pattern matching and entropy analysis.
• Symlink and Sandbox Escape — Resolves symbolic links and flags any whose target escapes the repository root (CWE-61 / CWE-451). Catches the class of path-traversal trick used in the GhostApproval disclosure (Wiz Research, July 2026).
• Dependency and Slopsquatting — For each new import, dynamic import (importlib.import_module, __import__), or manifest entry (including recursive -r includes) introduced in the diff, queries the PyPI JSON API and checks registration age. Packages that don't exist (HTTP 404) are flagged HIGH RISK; packages registered within the last 30 days are flagged MEDIUM RISK.
• Mock Usage Detection — Scans test diffs for newly introduced imports or fixture uses of unittest.mock, mocker, and monkeypatch, flagging them for human review with an optional --strict-mocks hard gate.
• Control Flow and Error Handling — AST-based detector for bare except: without re-raise, silently swallowed broad exceptions (except Exception: that only pass or log/print without re-raising or returning error indicators), and dead/unreachable code following unconditional return, raise, break, or continue, with an optional --strict-error-handling hard gate.

What This Doesn't Do

• Test-only diffs are not yet caught. If a diff modifies or weakens assertions in a test file without changing the corresponding source file, DeployProof currently sees zero modified source lines and passes with 0 mutants evaluated. This is a known gap — see INVESTIGATION_blastradius.md for the reverse-mapping approach being evaluated to close it. Until this lands, DeployProof does not protect against test suites being weakened directly.
• Diff-scoping vs Full Audits: Standard deployproof check is intentionally scoped strictly to files modified in the active git diff (for 2–5s speed). To audit every file in the entire repository, explicitly pass the --full-repo flag.
• Python only. Mutation testing and import extraction currently support Python files only. Other languages are not scanned.
• No auto-fix. DeployProof reports findings; it does not modify your code, rewrite imports, or suggest patches.
• No IDE plugin yet. There is no VS Code extension or JetBrains plugin. The CLI is the interface. IDE integration is on the roadmap.

See It Catch Real Bugs

Clone this repository and run the standalone stress-test suite to see DeployProof evaluate 11 planted edge cases:

python stress_fixtures/run_stress_tests.py

Fixtures cover weak test suites, zero-test orphan modules, planted OpenAI/AWS credentials, GhostApproval sandbox-escape traps, swallowed exceptions/dead code, and mock-masked broken implementations.

Status & Roadmap

• Current (v1.0.0): Diff-scoped AST mutation testing (with recursive discovery and SIGINT/SIGTERM/SIGBREAK signal-safe disk restoration), Full Repository Audit Mode (--full-repo) with isolated parallel multi-worker sandboxes and AST import-graph test discovery, baseline test-collection failure isolation with distinct exit code 2, entropy-driven value-based secrets scanner, GhostApproval symlink sandbox-escape detector, PyPI dependency hallucination / slopsquatting scanner (with import-to-distribution translation, recursive -r requirements scanning, and dynamic import detection via importlib / __import__), mock-introduction detector (--strict-mocks), control-flow / swallowed-exception scanner (--strict-error-handling), 11/11 launch-day stress test suite, 111 unit tests, live unbuffered progress streaming, and machine-readable --json output.
• Next: Reverse test-to-source dependency mapping (see FUTURE_SCOPE.md), SARIF 2.1.0 PR annotations, and multi-language mutation rule packs.

Contributing

Issues and pull requests are welcome. Open an issue first for significant changes so the approach can be discussed before implementation.

License

MIT. See LICENSE.

Created by SVS Praveen · Portfolio · LinkedIn