pi-behavior-monitors

Behavior monitors for pi that watch agent activity, classify against pattern libraries, steer corrections, and write structured findings to JSON files.

Monitors are JSON files (.monitor.json) with typed blocks: classify (LLM side-channel), patterns (JSON library), actions (steer + write to JSON), and scope (main/subagent/workflow targeting).

Install

pi install npm:@davidorex/pi-behavior-monitors

Or install all three extensions at once: pi install npm:@davidorex/pi-project-workflows

On first run, if no monitors exist in your project, example monitors are seeded into .pi/monitors/. Edit or delete them to customize.

Bundled Example Monitors

• fragility — detects when the agent leaves broken state behind (errors it noticed but didn't fix, TODO comments instead of solutions, empty catch blocks). Writes findings to .workflow/gaps.json.
• hedge — detects when the agent deviates from what the user actually said (rephrasing questions, assuming intent, deflecting with counter-questions)
• work-quality — on-demand audit of work quality (trial-and-error, not reading before editing, fixing symptoms instead of root causes). Invoked via /work-quality. Writes findings to .workflow/gaps.json.

File Structure

Each monitor is a triad of JSON files:

.pi/monitors/
├── fragility.monitor.json       # Definition (classify + patterns + actions + scope)
├── fragility.patterns.json      # Known patterns (grows automatically)
└── fragility.instructions.json  # User corrections (optional)

Writing Your Own

Create a .monitor.json file in .pi/monitors/ conforming to schemas/monitor.schema.json. Ask the LLM to read the pi-behavior-monitors skill for the full schema and examples.

Commands

| Command | Description | |---------|-------------| | /monitors | List all monitors, scope, and state | | /<name> | Show monitor patterns and instructions | | /<name> <text> | Add an instruction to calibrate the monitor |

How It Works

1. A monitor fires on a configured event (e.g., after each assistant message)
2. It checks scope (main context, subagent, workflow) and activation conditions
3. It collects relevant conversation context (tool results, assistant text, etc.)
4. A side-channel LLM call classifies the context against the JSON pattern library
5. Based on the verdict, the monitor executes actions:
   • steer: inject a correction message into the conversation (main scope only)
   • write: append structured findings to a JSON file (any scope)
   • learn: add new patterns to the library automatically
6. Downstream workflows can consume the JSON findings (e.g., gaps.json → verify step → gate)

Schemas

• schemas/monitor.schema.json — monitor definition format
• schemas/monitor-pattern.schema.json — pattern library entry format

Development

Part of the pi-project-workflows monorepo. All three packages (pi-project, pi-workflows, pi-behavior-monitors) are versioned in lockstep at 0.2.0.

npm run build compiles TypeScript to dist/ via tsc. The package ships dist/, not src/. Tests use vitest run (npm test).