Summary

SmileLikeYe/agent-chief tackles information overload by acting as an intelligent "chief of staff" for user attention. It filters agent outputs, alerts, and feeds, ensuring only critical interruptions reach the user, while efficiently dispatching and verifying agent tasks. This local-first, privacy-focused system empowers engineers and power users to regain control over digital noise.

How It Works

A three-stage worthiness engine (hard rules, similarity, LLM judge) prioritizes low-cost filtering. A scene engine dynamically adjusts interrupt thresholds based on user context (e.g., deep work, sleep). Events are then actioned as interruptions, dispatched to agents (with verification), curated into memory, or dropped. The system emphasizes explainability via editable Markdown policies and backs all claims with runnable, offline evaluations.

Quick Start & Requirements

Run uvx agent-chief demo for an offline demonstration. Install permanently with pip install agent-chief or uv tool install agent-chief. Initialize with uvx agent-chief init and run the service using chief run. Requires Python 3.12+.

Highlighted Details

• Three-Stage Worthiness Engine: Filters events via hard rules (µs), similarity (ms), and on-demand LLM judgment.
• Scene-Aware Timing: Adapts interrupt thresholds based on user context (clock, calendar, focus).
• Shadow Mode: Simulates actions without interrupting initially, allowing users to grade decisions and build trust.
• Editable Policy: Learned preferences distill nightly into POLICY.md for direct user editing.
• Verified Dispatch: Agent tasks undergo verification before confirmation.
• Local-First: Operates offline using SQLite/Markdown under ~/.chief, no cloud/telemetry.
• Evaluated Claims: All claims backed by 326 offline tests and runnable evaluations.

Maintenance & Community

Roadmap (ROADMAP.md) and contribution guidelines (CONTRIBUTING.md) are provided. Specific community channels (Discord/Slack) are not detailed.

Licensing & Compatibility

MIT license permits broad use, including commercial applications. Local-first, privacy-centric design enhances compatibility with security-conscious or offline environments.

Limitations & Caveats

v1 omits features like a web UI, cloud sync, and direct Slack/Discord delivery. LLM judge requires backend configuration. "Shadow mode" is a mandatory trust-building phase.