Overview
GLLA, distributed as pi-goal-list-loop-audit, is a Pi package for supervising substantial work across multiple agent turns. It preserves objectives, progress, pauses, and completion claims on disk, while a separate auditor checks results against agreed acceptance criteria.
The package is intended for work where continuity and verification justify additional coordination, including cross-system features, migrations, repository audits, documentation overhauls, backlogs, and specification-driven projects. It is not a model or hosted coding service.
Key Features
- Keeps goals and acceptance criteria durable across turns and interruptions.
- Supports bounded recovery when providers, sessions, or workers fail.
- Rejects unsupported completion claims and returns rejected work to the agent.
- Tracks confirmed task plans as milestones without treating them as proof of correctness.
- Provides visible pause, resume, status, verification, audit, cancellation, and archive controls.
Work Modes
A goal manages one verifiable outcome such as a feature, fix, migration, report, or audit. A list manages several independently auditable results and normally advances after each accepted item is archived.
Loop mode supports repeated improvement until a configured bound, metric plateau, or explicit stop. Respec mode develops a project against an evolving specification and verifies adopted requirements independently before archival.
Supervision and Completion
The workflow begins by preserving an objective and observable acceptance criteria. The agent then researches and implements through supervised turns, retaining state through interruptions.
Finishing requires an evidence-backed claim followed by detached verification; simply saying the work is done does not close the goal. Important scope changes remain visible, and unfinished tracked tasks prevent completion unless an explicit deferral has been recorded.
Package Safety
GLLA runs within the user’s Pi session and relies on its tools and authenticated providers. Like other Pi packages, it can execute code and influence agent behavior, so its source should be reviewed before installation; Pi’s project trust controls resource loading but does not sandbox tool calls.