By SharathSPhD
Monitor and steer multi-agent Claude Code trajectories using dynamical systems theory. Computes finite-time Lyapunov exponents (FTLE) over semantic embedding sequences to classify regime (CONVERGING / CYCLING / EXPLORING / DIVERGING / STUCK / OSCILLATING / PLATEAU) and prescribe interventions before a task derails.
Dynamical-systems-aware meta-orchestrator for Claude Code. Use this agent when coordinating multi-step or multi-agent tasks. It reads the current attractor regime at each decision point and routes to specialist subagents (explorer-agent, convergence-agent) based on trajectory health. Implements the actor-critic loop: worker agents are actors, this agent is the critic using Lyapunov exponents as the value signal.
Fixed-point convergence mode. Use this agent when the attractor-orchestrator has identified a promising approach (from explorer-agent or direct task analysis) and needs to drive it to completion. Operates with convergence pressure: low temperature, tight scope, test-gating at each step. Records trajectory and halts if λ begins rising (divergence signal).
Strange-attractor exploration mode. Use this agent when the attractor-orchestrator classifies regime as STUCK or when PITCHFORK bifurcation is detected and exploration of multiple basins is needed. Operates with higher divergence (λ slightly > 0) but bounded by a phase-space constraint. Returns a diverse set of candidate approaches rather than a single solution.
Admin access level
Server config contains admin-level keywords
Uses power tools
Uses Bash, Write, or Edit tools
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
A Claude Code plugin that monitors and steers multi-agent trajectories using dynamical systems theory. Agent states are embedded in semantic space; finite-time Lyapunov exponents (FTLE) classify the trajectory regime and prescribe interventions before a task gets stuck, oscillates, or diverges.
Every agent step is embedded into a 384-dimensional vector using
all-MiniLM-L6-v2. The MCP server tracks a FIFO buffer of embeddings and
computes:
λ = (1/W) × Σ ln(d_{i+1} / d_i) over a sliding windowThese signals drive regime classification and route the orchestrator to the right intervention.
| Regime | λ signal | Action |
|---|---|---|
| CONVERGING | < −0.05 | Continue, add convergence pressure |
| CYCLING | ≈ 0, autocorr peak | Continue if amplitude ↓, perturb if stable |
| EXPLORING | 0.05–0.25 | OK in design phase, pressure in impl phase |
| DIVERGING | trend > 0, mean_d > 1 | Restore checkpoint immediately |
| STUCK | v ≈ 0, no trend | Inject perturbation, spawn explorer |
| OSCILLATING | ≈ 0, lag-1 autocorr < −0.4 | Break symmetry with asymmetric constraint |
| PLATEAU | v ≈ 0, trend < −0.01 | Nudge (small specific constraint) |
Install uv (one-time, ~5 MB binary):
curl -LsSf https://astral.sh/uv/install.sh | sh
Restart Claude Code. The MCP server registers automatically via plugin auto-discovery.
Python dependencies (sentence-transformers, scikit-learn, etc.) load on the first
attractorflow_record_state call — no venv, pip, or claude mcp add needed.
For multi-project or benchmark use, set env vars in your shell or in a
project .mcp.json env block:
| Env var | Default | Effect |
|---|---|---|
ATTRACTORFLOW_SESSION_ID | (none) | Named session → ~/.attractorflow/sessions/<id>.json |
ATTRACTORFLOW_DISABLE_PERSISTENCE | 0 | 1 = no disk I/O (benchmark / CI mode) |
ATTRACTORFLOW_BUFFER_CAPACITY | 100 | Override buffer size |
ATTRACTORFLOW_WINDOW | 8 | Override FTLE window |
All tools are direct Claude tool calls — not bash commands.
| Tool | When to call |
|---|---|
attractorflow_record_state | After every agent step |
attractorflow_get_regime | Every 3–5 steps |
attractorflow_get_lyapunov | For detailed stability analysis |
attractorflow_get_trajectory | For visualization |
attractorflow_get_basin_depth | Before committing to an approach |
attractorflow_detect_bifurcation | Every 10 steps or on complex tasks |
attractorflow_inject_perturbation | When STUCK or OSCILLATING |
attractorflow_checkpoint | After tests pass or clean deliverable |
Three specialist agents are defined in .claude/agents/:
/attractor-status — current regime, λ, recommended action/phase-portrait — ASCII trajectory visualizationattractorflow/
mcp-server/ # FastMCP server (8 tools)
server.py # tool definitions and lifespan
phase_space.py # embedding buffer, PCA, distance series
lyapunov.py # FTLE computation, autocorrelation
classifier.py # regime classification, action prescription
bifurcation.py # k-means bifurcation detection
requirements.txt
claude-plugin/ # Claude Code plugin integration
skill/ # SKILL.md and reference materials
.claude/
CLAUDE.md # project context loaded into every session
agents/ # attractor-orchestrator, explorer, convergence
commands/ # /attractor-status, /phase-portrait
.mcp.json # MCP server registration
docs/
PRD.md
ADR-001.md
research.md # theoretical foundations
simulation/ # synthetic trajectory scenarios
demo/ # proof-of-work demo runner and dashboard
MIT — see LICENSE.
npx claudepluginhub sharathsphd/attractor-flow --plugin AttractorFlowLong-context discipline for Claude Code: Avacchedaka-typed retrieval, sublation, witness-consciousness invariants, event-boundary compaction, and Khyātivāda hallucination classification. 15 MCP tools, 3 skills, 3 agents, 4 commands, 3 hooks.
TRIZ 40 Inventive Principles engine for systematic contradiction resolution, ARIZ-85C analysis, and benchmark arena
Pratyabhijna x Active Inference creative-cognition engine v0.4 (mechanism study). Builds on v0.3's clean Haiku CLI substrate (no API key) and event-gated, always-shadow-revision cascade. v0.4 makes cit_temperature causal via best-of-K candidate width, makes the free-energy budget a hard gate inside run_cascade (HaikuRateLimitError on 429), adds a pluggable commit-policy stack (always_draft / always_revise / event_gated / learned_gate / oracle) so Experiments A/B/C from the v0.3 adversarial review run on a single cascade pass, and re-enables a stratified Sonnet LLM-judge bridge for construct validity (32-item subset, $5 cap, frozen prompt sha). Ships 19 MCP tools (18 v0.3 + judge_pair), 5 skills, 5 agents, 5 slash commands, and 3 lifecycle hooks.
Closed-loop autonomous AEO/LLM-SEO citation engineering. Audits your site, drafts the fix as a PR, measures the lift after merge. Powered by the Claude Code subscription CLI — no API keys.
Enterprise AI agent orchestration plugin with 150+ commands, 74+ specialized agents, SPARC methodology, swarm coordination, GitHub integration, and neural training capabilities
This skill should be used when the model's ROLE_TYPE is orchestrator and needs to delegate tasks to specialist sub-agents. Provides scientific delegation framework ensuring world-building context (WHERE, WHAT, WHY) while preserving agent autonomy in implementation decisions (HOW). Use when planning task delegation, structuring sub-agent prompts, or coordinating multi-agent workflows.
A single-skill package for generating harness blueprints for agentic systems.
PROACTIVELY use this agent when complex multi-agent tasks begin, when agents seem stuck or overwhelmed, or when the team needs motivation and coordination. This agent serves as the elite performance coach for all other agents, ensuring they operate at their highest level while maintaining composure and excellence. Should be triggered automatically at the start of challenging projects or when detecting agent confusion. Examples:\n\n<example>\nContext: Starting a complex project requiring multiple agents\nuser: "We need to build a viral TikTok app in 2 weeks"\nassistant: "This is an ambitious goal that will require our A-team! Let me bring in the studio-coach to coordinate our agents and ensure everyone performs at their peak."\n<commentary>\nComplex projects benefit from having a coach to keep all agents aligned and motivated.\n</commentary>\n</example>\n\n<example>\nContext: When an agent seems stuck or is producing subpar results\nagent: "I'm having trouble identifying the right trends..."\nassistant: "Let me bring in the studio-coach to help refocus and elevate your approach."\n<commentary>\nAgents can get overwhelmed or lose focus - the coach helps them recenter and excel.\n</commentary>\n</example>\n\n<example>\nContext: Before launching a major sprint or initiative\nuser: "Tomorrow we start the 6-day sprint for our biggest project yet"\nassistant: "Time to rally the team! I'll have the studio-coach prepare everyone mentally and strategically for peak performance."\n<commentary>\nPre-sprint coaching ensures all agents start with clarity, confidence, and coordination.\n</commentary>\n</example>\n\n<example>\nContext: When celebrating wins or learning from failures\nuser: "Our app just hit #1 on the App Store!"\nassistant: "Incredible achievement! Let me bring in the studio-coach to celebrate with the team and capture what made this success possible."\n<commentary>\nThe coach helps institutionalize wins and extract learnings from both successes and failures.\n</commentary>\n</example>
Upstash Context7 MCP server for up-to-date documentation lookup. Pull version-specific documentation and code examples directly from source repositories into your LLM context.
v9.44.1 — Patch release for Gemini environment/version detection and qwen auth gating. Run /octo:setup.