Skillget

smarttune-cli

mcp
Guvenlik Denetimi
Gecti
Health Gecti
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 21 GitHub stars
Code Gecti
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

SmartTune CLI — Multi-platform flight log analysis & tuning advisor (ArduPilot + Betaflight + PX4)

README.md

SmartTune

Offline flight log analysis, built agent-first from day one
One command from raw log to tunable PID/FFT/MagFit parameters — no special flights needed

v2.3.1 License Python 3.9+ Tests

ArduPilot · Betaflight · PX4
Quick Start · For Agents · Commands · Output Formats · Architecture

Pip-install SmartTune, point it at a flight log, and your agent comes back with exact parameter deltas — not vague "try lowering P."
Every output is structured JSON or Markdown. No TUI, no blocking prompts, no ANSI escape codes to strip. The CLI was designed from day one to be invoked by an LLM agent as a subprocess tool — OpenClaw, Hermes, QwenPaw, Codex, Claude Code. Any agent that can shell out to subprocess.run() can use it.

Under the hood: ArduPilot step-response analysis replicates WebTools PIDReview.js via Wiener deconvolution. Betaflight blackbox logs are parsed by a 1000+ line pure-Python decoder (no C extensions, no Node.js). All three platforms output to one FlightData dataclass so analyzers work identically across APM/BF/PX4. The 6-layer knowledge base is plain JSON — agents can read rules and propose new ones by editing ~/.smarttune/knowledge/.

Install

pip install git+https://github.com/raylanlin/smarttune-cli.git

# With all platform extras
pip install "git+https://github.com/raylanlin/smarttune-cli.git#egg=smarttune[all]"

# Development
git clone https://github.com/raylanlin/smarttune-cli.git
cd smarttune-cli
pip install -e ".[dev,all]"

Requires Python 3.9+

For Agents

SmartTune was designed specifically for LLM agent tool-calling workflows. Every aspect of the CLI follows agent-friendly principles:

Principle Implementation
Deterministic output No interactive prompts, no TUI, no progress bars when piped to files. Same input → same output.
Structured by default JSON output via MCP server (smarttune_analyze_log). Markdown/HTML via CLI --report md|html. No parsing fragile ANSI-escaped terminal dumps.
Self-describing stune platforms lists available adapters. Error codes are standardized (E10xx–E50xx). Exit codes are meaningful.
Fail-fast & isolated Single-module failure doesn't abort the full analysis. Each module gets its own try/except block.
Config-free Zero config files needed. Everything is flags or auto-detected. No env vars required.
Offline-first No network calls. No API keys. No rate limits. Safe for isolated/air-gapped environments.
Machine-recommendable Tuning suggestions include confidence scores and reasoning, not just parameter values. Agents can weigh multiple recommendations.

What agents can learn through SmartTune

SmartTune isn't just a tool agents call — it's how agents learn the craft of flight controller tuning:

Skill How SmartTune teaches it
PID tuning intuition Step-response analysis with confidence scores. Agents learn which overshoot/rise-time patterns call for higher Kp vs. damping.
Frequency-domain reasoning FFT spectra with peak detection. Agents learn to distinguish vibration sources (prop/ motor/ frame resonance) from the spectrum shape.
Filter design logic Notch and low-pass filter recommendations with Bode plots. Agents see the tradeoff between filtering and phase lag.
Platform differences ArduPilot vs Betaflight parameter conventions. ParamRef maps between them — agents learn to translate tuning knowledge across platforms.
Safety awareness All recommendations are capped at ±20%. Agents learn conservative tuning by default.
Rule-based reasoning The 6-layer knowledge base is plain JSON. Agents can read, understand, and even propose rule changes by writing to their user layer.

What agents can do with SmartTune

  • Batch analysis — analyze hundreds of logs with a loop; JSON output per file
  • Auto-tuning — feed recommendations back into a flight controller via MAVLink or CLI
  • Fleet monitoring — aggregate vibration/PID metrics across multiple aircraft
  • CI/CD integration — run stune analyze as part of a pre-flight validation pipeline
  • Collaborative diagnosis — have the agent compare logs from before/after a crash

MCP Server (Model Context Protocol)

SmartTune includes a read-only MCP server that lets LLM agents call analysis tools directly — no shell, no subprocess, no arbitrary file writes.

Install with MCP support:

pip install -e ".[all,mcp]"

Run the MCP server:

smarttune-mcp          # stdio transport (for agent frameworks)
# or
python -m smarttune.mcp_server

Available MCP tools:

Tool Purpose
smarttune_list_platforms List supported platforms, extensions, and capabilities
smarttune_log_quality Parse a log and return quality score, data availability, validation issues
smarttune_analyze_log Run full analysis (PID + FFT + MagFit + hardware) and return structured JSON or Markdown

All tools are annotated readOnlyHint=True, destructiveHint=False, idempotentHint=True.

Security boundary:

  • No shell execution — library calls only
  • No arbitrary file writes — results are returned inline
  • No parameter mutation — no MAVLink writes, no firmware flashing
  • Path validation — allowed roots, extensions (.bin, .log, .bbl, .bfl, .ulg), file size limits, symlink resolution
  • Configurable via environment variables:
export SMARTTUNE_MCP_ALLOWED_ROOTS="/path/a:/path/b"
export SMARTTUNE_MCP_MAX_FILE_MB="300"

OpenClaw / Claude Desktop configuration:

{
  "mcp": {
    "servers": {
      "SmartTune": {
        "command": "smarttune-mcp",
        "args": [],
        "env": {
          "SMARTTUNE_MCP_ALLOWED_ROOTS": "/home/user/.openclaw/workspace/files/inbox:/home/user/.openclaw/workspace/files/output:/tmp",
          "SMARTTUNE_MCP_MAX_FILE_MB": "300"
        }
      }
    }
  }
}

Quick Start

# Full analysis (auto-detect platform)
stune analyze -i flight.bin

# With charts (human-friendly)
stune analyze -i flight.bbl --visual

# Per-module deep dive
stune pid -i flight.bin -a roll --visual
stune fft -i flight.bin --visual
stune magfit -i flight.bin
stune sysid -i flight.bin -a pitch
stune hardware -i flight.bin

# Export to Markdown report
stune analyze -i flight.bin --report md -o report.md

# Export to HTML report
stune analyze -i flight.bin --report html -o report.html

# List supported platforms
stune platforms

# For JSON output, use the MCP server (see "For Agents" above)

Output Formats

SmartTune supports multiple output formats, each designed for a specific consumption mode:

Format Use Case Example
Terminal Human inspection in the shell stune analyze -i flight.bin
JSON Agent/script consumption MCP: smarttune_analyze_log(log_path="flight.bin")
Markdown Reports, READMEs, documentation stune analyze -i flight.bin --report md -o report.md
HTML Visual reports with embedded charts stune analyze -i flight.bin --report html -o report.html

Note: JSON output is available through the MCP server. The CLI currently supports --report md and --report html. A --format json CLI flag is planned for a future release.

JSON output example

{
  "platform": "ArduPilot",
  "timestamp": "2026-05-03T22:30:00",
  "pid": {
    "roll": {
      "rating": "GOOD",
      "confidence": 0.87,
      "kp": {"current": 0.12, "recommended": 0.14, "reason": "Slight oscillation at 8 Hz"},
      "ki": {"current": 0.05, "recommended": 0.05, "reason": "No steady-state error"},
      "max_overshoot_pct": 8.2,
      "rise_time_ms": 85,
      "settling_time_ms": 210
    }
  },
  "fft": {
    "vibration": {
      "level_rms": 2.1,
      "grade": "EXCELLENT"
    },
    "peaks": [
      {"freq_hz": 47.5, "magnitude_db": -12.3, "source": "propeller"}
    ]
  }
}

Commands

stune analyze

Full-spectrum analysis: PID + FFT + MagFit + hardware — all in one pass.

stune analyze -i flight.bin                           # Auto-detect
stune analyze -i flight.bbl --platform betaflight      # Force platform
stune analyze -i flight.bin --visual                   # With charts
stune analyze -i flight.bin --report md -o report.md   # Markdown export
stune analyze -i flight.bin --report html -o report.html  # HTML export

stune pid

PID step-response analysis with per-axis tuning recommendations.

stune pid -i flight.bin                                # All axes
stune pid -i flight.bin -a roll                        # Single axis
stune pid -i flight.bbl --visual                       # Betaflight

stune fft

Frequency-domain vibration analysis with notch filter suggestions.

stune fft -i flight.bin                                # Full spectrum
stune fft -i flight.bin --visual                       # With spectrum plot

stune magfit

Magnetometer calibration — hard/soft iron offset, coverage, field strength.

stune magfit -i flight.bin                             # ArduPilot only

stune sysid

ARX system identification — natural frequency and damping ratio.

stune sysid -i flight.bin -a roll                      # Single axis
stune sysid -i flight.bin -a pitch --na 4 --nb 3       # Custom order

stune hardware

Full hardware configuration report: firmware version, sensors, battery, parameters.

stune hardware -i flight.bin
stune hardware -i flight.bbl

stune filter

Filter chain analysis with Bode plots.

stune filter -i flight.bin --gyro-filter 40 --visual
stune filter -i flight.bin --auto                     # Auto-derive from params

stune platforms

List all available platform adapters and their capabilities.

stune platforms

Supported Platforms

Platform Log Format Parser Status
ArduPilot .bin / .log (DataFlash) pymavlink ✅ Full support
Betaflight .bbl / .bfl (Blackbox) Pure Python ✅ Full support
PX4 .ulg (ULog) pyulog 🔲 Coming in v2.x

Auto-Detection

SmartTune identifies your log format from file headers — no --platform flag needed:

Bytes Platform
0xA3 0x95 ArduPilot DataFlash
H Product:Blackbox Betaflight Blackbox
ULog magic PX4 ULog

Architecture

┌─────────────────────────────────────────────┐
│  CLI Layer                                   │
│  stune analyze / pid / fft / ...            │
│  --report md / html                          │
└──────────────────┬──────────────────────────┘
                   │
┌──────────────────┤  ┌───────────────────────┐
│                  │  │  MCP Server (stdio)    │
│                  │  │  smarttune-mcp         │
│                  │  │  JSON / Markdown out   │
│                  │  │  Read-only · No shell  │
│                  │  └───────────┬───────────┘
│                  │              │
│  ┌───────────────▼──────────────▼────────────┐
│  │  Services Layer (shared)                   │
│  │  services/analysis.py · services/serialize │
│  └───────────────┬──────────────────────────┘
│                  │
┌──────────────────▼──────────────────────────┐
│  Platform Adapter Layer                      │
│  ArduPilot · Betaflight · PX4               │
│  Parsers → FlightData (unified IR)          │
└──────────────────┬──────────────────────────┘
                   │
┌──────────────────▼──────────────────────────┐
│  Analysis Engine (platform-aware)            │
│  PID / FFT / SysID / MagFit / Filter / HW   │
│  Per-platform modules:                       │
│    ardupilot/  → WebTools-aligned FFT        │
│    betaflight/ → Wiener deconvolution FFT    │
│    px4/        → stubs                       │
│  BF: Feedforward · RPM Filter · D-term      │
│  Protocol-based interface constraints        │
└──────────────────┬──────────────────────────┘
                   │ AnalysisResult + ParamRef
┌──────────────────▼──────────────────────────┐
│  Knowledge Base (6-layer deep merge)         │
│  common → platform → user → Pro             │
│  JSON-based rules — inspectable & editable   │
└──────────────────┬──────────────────────────┘
                   │
┌──────────────────▼──────────────────────────┐
│  Output Layer                                │
│  Terminal (Rich) / JSON / Markdown / HTML    │
│  ParamRef → platform-native parameter names  │
└─────────────────────────────────────────────┘

Knowledge Base

A 6-layer deep-merge rule engine powers all tuning recommendations. Each layer overrides the previous:

# Layer Location Editable
1 Common physics rules smarttune/knowledge/rules/common/ ❌ Built-in
2 Platform rules smarttune/knowledge/rules/{platform}/ ❌ Built-in
3 User common ~/.smarttune/knowledge/common/ ✅
4 User platform ~/.smarttune/knowledge/{platform}/ ✅
5 Pro common smarttune-knowledge-pro (optional) 🔒
6 Pro platform smarttune-knowledge-pro (optional) 🔒

Rules are standard JSON files. Add a file, restart the command, and the engine picks it up. No compilation, no database, no setup.

Development

# Install with dev dependencies
pip install -e ".[dev,all]"

# Run tests
pytest tests/ -v                              # 96 tests, 1.5s
pytest tests/test_bbl_parser.py -v            # BBL parser only
pytest tests/test_betaflight_analyzers.py -v  # BF-specific analyzers

# Lint
ruff check smarttune/
black --check smarttune/

Adding a New Platform

  1. Create adapter — Subclass PlatformAdapter, implement parse(), detect(), map_param_to_platform()
  2. Add knowledge rules — Drop JSON files in smarttune/knowledge/rules/{platform}/
  3. Register — Use @register decorator
@register
class MyPlatform(PlatformAdapter):
    name = "myplatform"
    ...

stune platforms will auto-discover it.

Agent Stack Integration

SmartTune is designed to work with any LLM agent framework. Here's how it fits:

Framework Integration
OpenClaw smarttune-mcp as an MCP server — structured JSON output, read-only, no config needed
Claude Code / Codex MCP server or shell tool call — stune analyze -i log.bin --report md
Hermes Agent Deterministic output, safe for agent-in-the-loop tuning workflows
Custom agents pip-installable, importable Python API via smarttune.services.analysis

Agents call stune, get structured tuning recommendations, and can act on them. No TUI to navigate, no prompts to answer, no fragile screen-scraping.

Examples

Terminal Output

Platform: Betaflight
╭──────────────────────────────────────────────────────────────────╮
│ PID Step Response Analysis                                       │
╰──────────────────────────────────────────────────────────────────╯

  PITCH: MARGINAL  (steps: 4)
  ROLL:  MARGINAL  (steps: 1)
  YAW:   MARGINAL  (steps: 1)
  Overall: MARGINAL

╭──────────────────────────────────────────────────────────────────╮
│ FFT Vibration Analysis                                           │
╰──────────────────────────────────────────────────────────────────╯
  Vibration: MARGINAL (10.0 m/s²)
  Freq (Hz)    Amplitude (dB)    Source
       93.7             -46.5    motor
    → gyro_notch1_hz: 93.7
    → gyro_lowpass_hz: 40
    → acc_lpf_hz: 10

✓ Analysis complete!

PID Step Response

ArduPilot (DataFlash .bin log):

ArduPilot PID step response

Betaflight (Blackbox .bbl log):

Betaflight PID step response

Agent Analysis Report

When an AI agent analyzes a flight log through SmartTune, it produces a structured diagnostic report like this:

ArduPilot Flight Log Analysis Report
Log: 2026-04-26 13-46-44.bin | Duration: 995s | Platform: ArduPilot

1. PID Step Response Analysis

Axis Rating Rise Time Overshoot Settling Oscillations
Roll MARGINAL -1ms 0.0% 510ms 8
Pitch MARGINAL -1ms 0.0% 510ms 4
Yaw MARGINAL -1ms -1.0% -1ms -

Roll Axis Recommendations:

Parameter Current → New Change Reason
ATC_RAT_RLL_D 0.0036 → 0.0040 +10% Reduce oscillation (8 cycles)
ATC_RAT_RLL_I 0.115 → 0.144 +25% Eliminate steady-state error (99.8%)
ATC_RAT_RLL_P 0.115 → 0.104 -10% Reduce oscillation

Pitch Axis Recommendations:

Parameter Current → New Change Reason
ATC_RAT_PIT_I 0.115 → 0.144 +25% Eliminate steady-state error (99.8%)
ATC_RAT_PIT_D 0.0036 → 0.0040 +10% Reduce oscillation (4 cycles)
ATC_RAT_PIT_P 0.115 → 0.104 -10% Reduce oscillation

Yaw Axis: No changes needed — parameters already acceptable.

2. FFT Vibration Analysis

Rating: EXCELLENT (0.5 m/s²)

Current filter settings:

Parameter Value
INS_GYRO_FILTER 60 Hz
INS_ACCEL_FILTER 10 Hz
Notch filters None enabled

Vibration levels are excellent. No additional filtering required.

3. Magnetometer Calibration

Fitness: 567.98 mGauss — BAD

Issues detected:

Issue Threshold Actual
Hard iron offset (max |OFS|) 600 625
Soft iron anomaly (DIA_X/Y/Z) — 0.300
Motor interference (max |MOT|) 100.0 200.0
Flight coverage — No attitude variation

Recommendations:

  • Remove hard iron interference sources (speakers, magnets)
  • Optimize soft iron layout (battery/motor placement)
  • Recalibrate with proper flight pattern: yaw > 300°, pitch/roll > ±30°

Summary

Module Status Action
Vibration ✅ Excellent Hardware is solid
PID ⚠️ Marginal Increase I and D gains on Roll/Pitch, reduce P slightly
Compass ❌ Bad Recalibrate before precision flight

The agent interprets SmartTune's structured JSON output, adds context, and produces a human-readable summary — bridging the gap between raw data and actionable tuning advice.

For Humans

Yes, the terminal output is also beautiful. Rich-powered tables, progress bars, color-coded diagnostics — everything you'd expect from a modern CLI. But the architecture underneath is agent-first.

# Human-friendly terminal output (default)
stune analyze -i flight.bin

# Machine-parseable via MCP server
# smarttune_analyze_log(log_path="flight.bin", response_format="json")

Roadmap

Phase Content Status
v1.x ArduPilot full support ✅
v2.0 Phase 1 Multi-platform architecture ✅
v2.0 Phase 2 Betaflight BBL parser + analytics ✅
v2.1 Platform-specific analyzers + Protocol constraints ✅
v2.2 Full English docs, CLI --help, OpenClaw SKILL.md ✅
v2.x PX4 ULog adapter 🔲
v3.0 Tool-calling manifest, plugin system, web UI 🔲

Author

Raylan LIN — @raylanlin

Built and maintained by a pilot who builds ArduPilot firmware (ParallelFC, self-learning PID, STM32H7 custom FC boards) and teaches his AI agent to tune better than he does.

License

MIT — see LICENSE for details.

smarttune-knowledge-pro is a separate closed-source tuning knowledge base with proprietary tuning rules and industry experience.

For commercial collaboration — custom tuning knowledge bases, fleet-level expertise integration, or enterprise tuning rule development — reach out at [email protected].

Yorumlar (0)

Sonuc bulunamadi