CPFS Runner — Program without Programming with AI

Program without programming with AI. Let me show you how.

CPFS Runner catches the three things AI coding assistants do wrong — heuristic shortcuts that won't generalize, oversteps into shared code, and spinning on failed logic — and stops them by mechanism, not memory. You direct the work. AI writes the code. You verify the evidence.

The three problems CPFS Runner solves

1. AI writes heuristic shortcuts that pass one example but won't generalize

Your AI writes a fix that works on the sample you showed it — but it's a demo-tuned shortcut: a ratio gate (if area < 8.5), a path-literal branch (if name.includes("MAX4")), a magic number. It passes today and breaks on the next upload.

CPFS Runner's §13 heuristic scanner reads every added diff line and flags numeric thresholds in conditionals, path-literal branches, single-sample verification, and workaround comments — before you ship. It also scans prior code for the same patterns so you know what might break.

2. AI oversteps into shared code other features depend on

Your AI edits a function. Three other features — already passed and shipped — call that function. The change looks clean. You ship it. Two of those features break.

CPFS Runner's blast-radius detection tracks every symbol every task touched. When a new change overlaps a prior passed task's symbols — directly or via a one-hop caller — it tells you to re-verify the affected task before the change ships. Frozen retest recipes re-run automatically and report pass/fail.

3. AI spins on the same failed logic across sessions

Your AI tries an approach. It fails. Next session, context is gone — the AI tries the same approach again. And again. You watch it spin.

CPFS Runner parses DO NOT REPEAT from the log and injects it into every prompt. The AI sees what already failed and is forced to change track. When all approaches are exhausted, an AI EXHAUSTED declaration fires an escalation event so you can step in with new direction.

You're the Director

CPFS Runner changes the relationship. You don't read diffs, hunt for scope violations, or catch the AI retrying last week's failed fix. You say what to fix and how to verify. The AI codes. CPFS Runner enforces the guardrails — scope, verification, regression, no-repeat — by technical limitation, not by hoping the AI remembers. You confirm pass when the evidence is real. That's what makes it possible to program without programming: the guardrails ensure the AI actually proved its work before it claims done.

No cloud. No account. No telemetry. Everything runs locally in your repo.

Proof: stl2cad.com was built with CPFS Runner — zero coding, under 30 days

stl2cad.com is a live paid service that converts mechanical STL uploads to STEP CAD files — detecting flat faces, holes, cylinders, and edges from a triangle mesh and reconstructing them as editable STEP geometry. It handles brackets, housings, plates, mounts, and fixtures. It is not a general-purpose mesh-to-CAD engine: artistic meshes, organic sculpts, and 3D scans are out of scope, files are capped at ~250,000 triangles (larger files are simplified on upload), and curved regions export as boundary outlines you finish in CAD rather than full parametric surfaces.

It was written in under 30 days. With zero coding. The owner directed the AI to build it using the CPFS protocol — every feature logged, every change scoped, every pass verified against real STL test parts. The owner never wrote or read a line of code. CPFS Runner enforced the discipline so the AI couldn't fake "done."

That's not a demo. That's a production product with paying customers — built by a non-programmer directing AI, with CPFS Runner as the guardrail. See it live at stl2cad.com.

CPFS Runner was also programmed using CPFS Runner. The extension enforced its own development discipline — every feature logged, every change scoped, every pass verified. The tool built itself. That's the deepest proof it works: we didn't ship a discipline tool that we wrote without discipline.

Install

After install, reload the window. You should see a CPFS item in the status bar.

Quick start

1. Open your project folder in Cursor / VS Code.
2. Ctrl+Shift+P → CPFS: Enable Workspace — creates logs/development_history/ and .cpfs/config.json. (After the first time, CPFS auto-enables on workspace open.)
3. Confirm it works — run CPFS: Run Install Self-Test (see below).
4. Work in chat — say what to fix and how to verify in one message. CPFS matches or creates the feature log; the status bar shows the active feature + attempt #.
5. Validate — CPFS: Validate Attempt runs your test command and records the real outcome.
6. Confirm pass — CPFS: Confirm Pass marks the feature COMPLETE (only after verification is agreed and checks pass).

Confirm your install: the self-test

CPFS: Run Install Self-Test runs a one-click check that the extension is wired up: workspace folder, CPFS enabled, config, logs folder, zones, and your test command.

• PASS → a green toast and a report in the CPFS Self-Test output panel. You're ready.
• FAIL → open the CPFS Self-Test output panel; each failing line says what's wrong. Fix the items marked FAIL, then run the self-test again.
• To run it again any time: Ctrl+Shift+P → CPFS: Run Install Self-Test.

On a fresh install, the self-test runs automatically once and shows you the result.

Where to read more

• CPFS protocol (free/open): https://ragbox.llc/tutorials/cpfs.html
• In-extension reference manual (every command, setting, file): run CPFS: Open Reference Manual
• In-extension quick start: run CPFS: Open Quick Start
• Dashboard: CPFS: Open Dashboard — hero next-action, timeline, verify status, stats.

Settings that matter

Privacy

All data is written only to .cpfs/ and logs/development_history/ under your opened project folder. No network calls are made for CPFS features. Your validation commands run locally on your machine.

License

CPFS Runner is proprietary commercial software. Use requires a paid subscription (or a written pre-release evaluation authorization). See LICENSE.

The free CPFS methodology on ragbox.llc is separate — it describes how to work; it does not grant rights to copy or redistribute CPFS Runner.