SURGE.BABY

Pi Flow

2026

A TypeScript workflow tool that turns vague agent tasks into traceable plans, execution loops, and multi-model review.

Pi Flow is a TypeScript workflow tool for Pi CLI and AI-agent delivery. It turns vague requests into clarification-first plans, then moves through /goal or /flow execution loops before a review loop closes the work.

I designed it around a zero-tool-injection pattern: it does not extend the agent runtime, but uses prompts, role separation, and traceable reports to keep delivery accountable. Planner, executor, and reviewer roles can use different model assignments, so one model is not forced to plan, execute, and self-review alone.

The project includes /goal, /flow, and /review commands, local HTML progress reports, bilingual prompt assets, and smoke tests for planning, flow, review, reporting, and language behavior.

It is published as @surgee/pi-flow on npm and is open on GitHub. The public case study describes the product and engineering trade-offs without claiming unverified delivery metrics.