Monitor and steer Claude Code agent trajectories using attractor-flow dynamics.
npx skills add https://github.com/SharathSPhD/attractor-flow --skill attractorflow
Please help me install this Agent Skill. Skill: attractorflow Source: https://github.com/SharathSPhD/attractor-flow/tree/main/skills/attractor-flow Command: npx skills add https://github.com/SharathSPhD/attractor-flow --skill attractorflow
Automates monitoring and steering of Claude Code agent trajectories using attractor-flow dynamics to prevent loops, stalls, or divergent behavior.
Initialize AttractorFlow in your Claude Code project and start recording state after each agent step to monitor regimes and trigger interventions.
High-intent search queries and answers about installing and using this skill.
Prevent multi-agent Claude Code tasks from looping by classifying trajectory regimes with Lyapunov exponents and applying orchestration actions like DECOMPOSE_TASK or INJECT_PERTURBATION to actively steer stalled agents.
Attractor-flow dynamics is a monitoring approach that uses Lyapunov exponents to classify real-time trajectory regimes in Claude Code agents, triggering defined orchestration actions like SPAWN_EXPLORER or BREAK_SYMMETRY when agents drift or stall.
Decompose a Claude Code task into parallel subagents by monitoring trajectory health after each agent step and triggering the DECOMPOSE_TASK orchestration action when the Lyapunov regime classification indicates the task requires splitting.
Yes, you can use an MCP server for real-time diagnostics in Claude Code multi-agent tasks, as AttractorFlow integrates MCP server tooling to monitor trajectory health and trigger orchestration actions during long-running iterations.
Available orchestration actions to steer Claude Code agent trajectories include CONTINUE, REDUCE_TEMPERATURE, INJECT_PERTURBATION, SPAWN_EXPLORER, BREAK_SYMMETRY, RESTORE_CHECKPOINT, DECOMPOSE_TASK, NUDGE, and HALT.
Avoid using Lyapunov-based regime classification for agent orchestration in short-running or single-step Claude Code tasks that do not risk looping, drifting, or stalling, as the trajectory monitoring overhead provides no intervention value.
EvezArtDetect UAP-like phase transitions and strange attractors in time-series data.
danielorzeata83-a11yClassify market regimes and forecast probabilities using Markov models.
carl10086Inspect DOM, console logs, and network requests via Chrome DevTools MCP.
ClaudioCeppi83Coordinate multiple specialized agents for code analysis and feature implementation.
