Robot Ops — Development skill for Claude Code

robot-ops-copilot

Five Claude-powered reliability agents for robotics fleets: sensor-fusion sanity checks, ROS 2 uptime tracking, edge-to-cloud latency profiling, automated root cause analysis, and predictive maintenance for actuators.

Every agent runs offline against a built-in synthetic scenario, so you can clone the repo and see all five working before you point them at a robot or spend a token.

git clone https://github.com//robot-ops-copilot.git
cd robot-ops-copilot
pip install -e ".[dev]"
robotops fusion --offline     # no API key needed

The design rule

Deterministic code decides. Claude explains.

Each agent splits into two halves. Plain Python owns detection: filters, thresholds, percentiles, trend fits. It is testable, fast, and never hallucinates. Claude reads the resulting summary and supplies what statistics cannot — the physical cause, the cross-signal correlation, the recommended fix in engineering language.

This is not an aesthetic preference. A language model has no place inside a state estimator's update loop or a watchdog's decision path, and anything safety-critical must hold when the API is unreachable. So every agent produces its findings with or without a model, and the `source` field on each finding records which half produced it: `rule` or `claude`.

The five agents

Command	What it watches	Deterministic half	Claude's half
robotops fusion	IMU, wheel odometry, LiDAR	EKF innovations, cross-sensor p90 voting	Names the physical cause, proposes R adjustments
robotops diagnostics	ROS 2 /diagnostics	Uptime, MTBF, stale heartbeats, CPU peaks	Correlates failures across nodes
robotops latency	Teleop round trips	Per-stage p50/p95/p99, jitter, budget breaches	Attributes the bottleneck to a cause
robotops rca	E-stops and crashes	30 s rolling buffer, peak extraction	Writes the RCA: proximate vs root cause
robotops maintenance	Motor telemetry