turboplex 0.3.3

High Performance Test Engine - Rust + Python

TurboPlex (tpx) — Fast Test Orchestration for Python (Rust core)

English | Español

TurboPlex is a hybrid Rust + Python test runner that accelerates large suites with:

• fast test discovery
• parallel execution
• SHA-based caching
• structured JSON reports for IDEs and AI agents

TL;DR

pip install turboplex
tpx --path tests/

Contents

• Why TurboPlex
• Installation
• Quick Start
• Modes
• Windows / venv (TPX_PYTHON_EXE)
• Skipping tests (pytest.skip)
• Reports
• Cache
• MCP server (IDE integration)
• Configuration
• Benchmarks
• Troubleshooting
• Development
• License

Why TurboPlex?

TurboPlex focuses on making “run tests” fast and machine-consumable:

Feature	What you get
Fast discovery	AST-based collection avoids heavy imports where possible
Parallel execution	Isolated Python subprocess per test (or pytest in compat mode)
Smart caching	Cache pass results keyed by file hash + runtime fingerprint
Structured output	Stable JSON schema for tooling, MCP, and analysis
Watch mode	Rerun suite on file save

Installation

From PyPI:

pip install turboplex
tpx --help

From source (dev):

git clone https://github.com/IzumiKeita/turboplex.git
cd turboplex
pip install -e .
tpx --help

Quick Start

Run a directory:

tpx --path tests/

Run multiple paths:

tpx --path tests/ --path tests/integration/

Auto-discover:

tpx

Watch mode:

tpx --watch --path tests/

Modes

Native mode (default)

Best for suites where:

• imports can be expensive (DB bootstraps, migrations, Pydantic config)
• you want faster iteration with caching and structured results

Pytest compatibility mode (--compat)

Use this when you need full pytest semantics/plugins:

tpx --compat --path tests/

Light collect (--light)

Skips conftest.py import during discovery. Useful when conftest.py performs heavy work on import.

tpx --compat --light

Environment alternative:

export TPX_MCP_LIGHT_COLLECT=1  # Linux/Mac
set TPX_MCP_LIGHT_COLLECT=1     # Windows

Windows / venv (TPX_PYTHON_EXE)

TurboPlex resolves the Python interpreter using this precedence:

1. TPX_PYTHON_EXE (if set and path exists)
2. python.interpreter from turbo_config.toml (if set and path exists)
3. Auto-detected .venv/venv/.env/env near the project
4. System python

Example (force a secondary venv):

$env:TPX_PYTHON_EXE = (Resolve-Path .\.venv_alt\Scripts\python.exe).Path
tpx --path tests/

CLI prints a confirmation line:

• Using TPX_PYTHON_EXE: C:\path\to\.venv_alt\Scripts\python.exe

Skipping tests (pytest.skip)

In native mode, pytest.skip(...) is recognized and reported as a skip (not a failure):

• CLI prints SKIP ... lines
• summary includes skipped
• JSON includes skipped: true and skip_reason

Example summary:

Results: 120 passed, 0 failed, 5 skipped (24000ms)

Reports

TurboPlex generates machine-readable artifacts:

• .tplex_report.json and timestamped .tplex_report_%Y%m%d_%H%M%S.json
• turboplex_full_report.json (JSONL) with richer error context for large suites

Cache

Cache lives in .turboplex_cache/ and is invalidated when:

• test files change (hash)
• runtime fingerprint changes (Python version, deps hash, PYTHONPATH, flags)

MCP server (IDE integration)

Start the MCP server:

tpx mcp

Notes:

• --out-json avoids stdout pollution for JSON-RPC toolchains.
• subprocesses force UTF-8 (PYTHONUTF8=1, PYTHONIOENCODING=utf-8, PYTHONUNBUFFERED=1).

Common MCP env vars:

• TPX_PYTHON_EXE
• TPX_MCP_LIGHT_COLLECT=1
• TPX_MCP_DEBUG=1
• TPX_MCP_STDOUT_MODE=redirect|failfast
• TPX_MCP_TEST_TIMEOUT_S (default 120)
• TPX_MCP_TURBOPLEX_COLLECT_TIMEOUT_S (default 120)
• TPX_MCP_TURBOPLEX_RUN_TIMEOUT_S (default 60)
• TPX_MCP_PYTEST_COLLECT_TIMEOUT_S / TPX_PYTEST_COLLECT_TIMEOUT_S (default 120)
• TPX_MCP_PYTEST_RUN_TIMEOUT_S / TPX_PYTEST_RUN_TIMEOUT_S (default 60)
• TPX_MCP_HEARTBEAT_S (default 1)
• TPX_MCP_TERMINATE_GRACE_S (default 2)
• TPX_MCP_DRAIN_MAX_CHARS (default 2000000)
• TPX_MCP_LOGS_MAX_CHARS (default 20000)

Error contract (ok=false):

• data.error.code in timeout | subprocess_failed | invalid_input | not_found | internal_error
• data.error.message is always human-readable
• data.error.details may include phase, returncode, timeout_s, and truncated stderr/stdout metadata

Tool response shape (discover, run, get_report):

• top-level: schemaVersion, tool, ok, runId, mode, summary, logs, data
• run.summary also includes workers_used, timeouts, subprocess_failures
• DB-first additions:
  • data.results[].db_metrics.write_count
  • data.results[].db_dirty
  • data.results[].db_dirty_summary
  • run.summary.db_write_count_total
  • run.summary.db_dirty_tests

Integration coverage added today:

• tests/test_mcp_db_integration.py validates MCP run DB metrics with SQLite writes.
• strict dirty policy verified:
  • TPX_DB_STRICT_DIRTY=0 -> run can pass while reporting db_dirty.
  • TPX_DB_STRICT_DIRTY=1 -> run fails with db_error.code=db_dirty_state.
• subprocess-only variant added with xfail guard on Windows for occasional native crash 0xC0000005 (Access Violation).

DB hardening env vars:

• TPX_DB_STRICT_DIRTY=0|1 (default 0, fail test only when dirty + strict enabled)
• TPX_DB_METRICS_ENABLED=0|1 (default 1)
• TPX_DB_ISOLATION_MODE=auto|schema|database|transaction (default auto)
• TPX_DB_WORKER_PREFIX=tpx_w
• TPX_DB_DIRTY_TRACK_MAX_TABLES=12

Example MCP config:

{
  "mcpServers": {
    "tpx": {
      "command": "tpx",
      "args": ["mcp"],
      "env": {
        "TPX_PYTHON_EXE": "/path/to/venv/bin/python",
        "TPX_MCP_LIGHT_COLLECT": "1",
        "TPX_MCP_DEBUG": "1"
      }
    }
  }
}

Configuration

turbo_config.toml example:

[execution]
max_workers = 8
default_timeout_ms = 30000
cache_enabled = true

[python]
enabled = true
interpreter = "python"
module = "turboplex_py"
test_paths = ["tests"]
project_path = "."

Benchmarks

Production suite (~200 tests):

Tool	Time	per test
pytest	~340s	~1.7s
tpx (cold)	~180s	~0.9s
tpx (cached)	~25s	~0.13s

Troubleshooting

Common issues:

• “Fixture not found” in native mode
  • Use --compat if you rely on complex pytest fixtures/plugins
• Slow discovery
  • Use --light (or TPX_MCP_LIGHT_COLLECT=1) to avoid heavy conftest imports
• “ModuleNotFoundError” during run
  • Ensure your project is on PYTHONPATH or configured via turbo_config.toml

Development

Build wheels with maturin:

python -m pip install maturin
python -m maturin build --release -o dist
python -m maturin sdist -o dist

License

MIT License — see LICENSE