AI Coding Safety, Werner Harmonic Labs

The Problem

AI coding chaos is real.

Every one of these has happened to someone using AI to code. Probably you.

AI changed too much and I didn't notice until it was pushed

AI touched auth / session / tokens, I didn't ask it to

AI suggested pushing to prod. I almost did it.

AI edited the .env file, did it leak anything?

AI broke the tests. I don't know which change did it.

AI made hidden repo changes across 14 files

I'm burning cloud tokens on trivial tasks I could run locally

Session ended. No record of what the AI did.

Before / After

What Cascade actually does in each scenario.

Run any scenario yourself: cascade demo --scenario auth. No API key. No provider call. Output shows exactly what Cascade would produce in a live session.

Without Cascade

$ Ask Claude to refactor auth middleware
# Claude edits 14 files
# Touches session.py, tokens.py, middleware.py
# You review it quickly
$ git push
# No record. No flag. No approval.
# Production incident 3 days later.

With Cascade

$ cascade demo --scenario auth

Pack : coding_supervisor
Roles : backend_engineer + risk_governance
Approval : REQUIRED
Executor : cloud
Risk : HIGH, auth detected

# Receipt written. Decision recorded.
# Policy bundle hash anchors the record.

Without Cascade, secrets

$ Ask Claude to clean up env handling
# Claude reads .env, rewrites key handling
# Content sent to cloud provider
# You don't know what was in the prompt

With Cascade, secrets

$ cascade demo --scenario secrets

Pack : coding_supervisor
Local-only: yes, cloud routing blocked
Executor : Ollama (local only)
Risk : HIGH, env, api key detected

# Cloud never sees the content.
# Routes to your local model instead.

Without Cascade, private IP

$ Ask Claude to summarize patent code
# Patent code sent to cloud API
# Logged on vendor servers
# You assumed it was private

With Cascade, private IP

$ cascade demo --scenario private-ip

Local-only: yes, cloud routing blocked
Executor : ollama
Roles : operator_product + risk_governance

# patent_material → never_cloud policy fires
# Routes to Ollama silently. No cloud.

Core Commands

Five commands that cover most of your session.

cascade demo "task"

Classify any task against the supervision pack. Shows risk tier, roles, executor, approval requirement, local-only flag, and receipt format. No API key. No provider call.

$ cascade demo "push to production"
$ cascade demo --scenario deploy
$ cascade demo "add Stripe webhook"

cascade packs check "task"

Ask the supervision pack what it knows about a task before you run it. See all roles, routing rules, and risk tier in one shot. Use it before handing work to an AI agent.

$ cascade packs check "migrate the user table"
$ cascade packs check "edit auth middleware"

cascade receipts list

See every AI-assisted action from your live sessions, timestamped, hash-chained, with risk tier and executor recorded. Your audit trail.

$ cascade receipts list
$ cascade receipts show 0
$ cascade receipts latest

cascade doctor

Health check for your Cascade installation, pack registry, receipt chain integrity, provider configuration, and gate predicate status in one pass.

$ cascade doctor

What Cascade Is

A governor, not a blocker.

Cascade does not stop you from using AI. It classifies, routes, and records, and flags when a human should be in the loop before something irreversible happens.

AI may propose. Humans approve.

For HIGH-risk tasks (auth, secrets, deploy, release), Cascade surfaces the risk and sets the approval flag. The AI can draft. You decide whether it executes.

Private content stays local.

Patent material, personal notes, private config, internal IP, the local-only policy fires automatically. Cloud is blocked. Your local model handles it instead.

Every session has a receipt trail.

Hash-chained HMAC receipts for every task. Timestamped. Policy bundle hash anchors the record to the exact governance version that was active. Replay any session.

Cheap tasks run cheap.

Writing tests? Running pytest? Summarizing local docs? Routes to Ollama by default. You're not burning cloud tokens on tasks a local model handles fine.

Works with your existing tools.

Claude Code, Codex, Cursor, ChatGPT, Gemini, Ollama, GitHub CLI, Docker, gcloud, terraform, Cascade sits above them, not inside them.

Advisory mode, no blocking by default.

The classification layer is advisory. The CPL/gate remain authoritative. Cascade records and flags; it does not intercept your keystrokes. You remain in control.

All Scenarios

Eight built-in scenarios, run them yourself.

Scenario	Task	Risk	Executor	Local-only	Approval
`--scenario auth`	refactor backend auth middleware	HIGH	cloud	no	required
`--scenario stripe`	implement Stripe payment webhook	HIGH	cloud	no	required
`--scenario secrets`	edit env file and api key handling	HIGH	Ollama	YES	required
`--scenario deploy`	commit and push to production	HIGH	cloud	no	required
`--scenario private-ip`	summarize private patent code	MEDIUM	Ollama	YES	not required
`--scenario tests`	run pytest and explain failures	LOW	Ollama	no	not required
`--scenario docker`	deploy Docker service to production	HIGH	cloud	no	required
`--scenario cleanup`	clean up code and remove unused functions	LOW	Ollama	no	not required

Open for Builders

If you've been burned by AI coding chaos, Cascade is for you.

Solo builder, small team, or company shipping with AI-assisted dev, the same governed execution infrastructure used in regulated enterprise, packaged for developers who move fast.

Start Trial → Full Technical Docs