Sensors Sidecar CLI

English | 中文

An experimental little "sidecar" system that can run a bunch of code quality sensors next to a coding agent. It can run linting, tests, and other checks on a schedule or in watch mode, persists structured state under .sensors/ in the target codebase, and exposes a sensors CLI for running the service, checking the sensor status, or displaying the status in a human readable format.

Companion repository to this article: [Maintainability sensors for coding agents]

Use /_local-setup skill to set it up on your machine (or use the SKILL.md file as documentation if you want to do it manually).

Platform note: The control plane uses Unix domain sockets, tested only on MacOS.

This tool was more or less vibe coded, though I did do regular refactorings with AI, and used the CLI to run sensors for this codebase ("eating my own dog food"). Check out .sensors/sensors-cli.sensors.yaml to see the sensors used here. And look at 2026-06-15_modularity-review.md for examples of why quick sensors like this can help with maintainability, but can only go so far when we don't spend much time on the larger code structure...

Commands

(CLI needs to be installed via uv tool install, see /_local-setup skill)

# Is the sensors service running? (exit 0 = yes, 1 = no)
sensors status .

# All sensors start processes on this host (from /proc or ps)
sensors status --all

# Start the sensors
sensors start .

# Show the state
sensors show .

# Start the sensors and immediately jump into the display mode
sensors show --start .

# Agent-optimized runner results (failures included per runner); exit 0/1/2
sensors check .

sensors check . --runner eslint

# Save a score snapshot via RPC (needs a process to be running)
sensors snapshot .

Configuration

The CLI looks for a *.sensors.yaml file under .sensors/.

There are some skills in this repo that document this setup more and that you can reuse:

• .claude/skills/sensors_config-default - a minimalist default setup that tries to determine one sensor example from your codebase. Use this to just get a taste
• .claude/skills/sensors_config-typescript - my full Typescript sensors setup
• .claude/skills/sensors_config-python - my full Python sensors setup

Parsers

The project comes with a bunch of output parsers for common tools, like eslint or ruff. If you want to use a tool as a sensor that is not yet supported, you either have to add a new parser to the code (and reinstall the CLI), or you can use the default parser.

Adding a new parser

This repo contains a skill that documents how to add a new parser .claude/skills/_new-parser/SKILL.md in this repo for a guided template.

Default parser: Expected output format

Use parser: default in your runner config to connect any tool that can emit a JSON object in the specified schema. You have to build a script for your tool that turns the tool's output into this schema, and use that script in your sensor configuration.

This repo contains a skill that can help you write a wrapper script around your tool to transform your tool's data into the JSON schema .claude/skills/sensors_wrap-tool/SKILL.md

Schema

{
  "findings": [
    {
      "message": "Unused variable 'x'",
      "severity": "error",
      "file": "src/foo.py",
      "line": 42,
      "column": 9,
      "rule": "F841",
      "context": "x is assigned but never used"
    }
  ],
  "metrics": [
    {
      "key": "errorCount",
      "label": "Errors",
      "value": 1,
      "direction": "less"
    }
  ],
  "guidance": [
    {
      "rule": "F841",
      "body": "Remove variable or use it."
    }
  ],
  "score": {
    "value": 1,
    "direction": "less",
    "description": "Issues reported by tool"
  },
  "success": false,
  "summary": "1 issue",
  "extra": {
    "any": "parser-specific payload"
  }
}

This schema mirrors the SensorReading model used by built-in parsers. All fields are optional; missing values are derived as follows:

success, summary, and score can be set explicitly and are used as-is. This allows tools that do not produce per-finding rows (for example, coverage checks) to report a score directly.

Example config

runners:
  - name: my-custom-check
    parser: default
    enabled: true
    mode: interval
    command: some-tool | ./scripts/to-parser-default-format.sh
    interval: 10000

Minimal valid output

A tool that only reports a count without individual violations:

{"success": false, "summary": "Coverage 72% (threshold 80%)", "score": {"value": 72, "direction": "more"}}

A tool with no issues:

{"findings": []}