00Z — The ZEN Agent

Experimental local, security-first agent harness template for structured AI-agent workflows, with adapter contracts for Claude- and pi-agent-oriented usage.

⚠️ 00Z is experimental. It is intentionally conservative, local-first, and not production software.

Status at a glance

Area	Status
Local validation tools	Implemented
Reasoning contracts	Implemented
Safety/policy gates	Implemented as local checks
Pi/Claude adapter contracts	Present as contracts/scaffolding
Native Pi/Claude runtime	Not bundled
Native LLM provider connector	Not included
Productive project writes	Disabled by default
Dependency installation	Not required for the first demo
GitHub/deploy automation	Not included

In short

00Z is a local-first safety harness for experimenting with AI-agent architecture. It makes prompts, policies, reasoning, memory concepts, adapters, pipelines, and validation boundaries explicit and inspectable.

Use it to study or prototype structured AI-agent workflows without claiming a production runtime, automatic deployment flow, or native LLM provider connector.

Reasoning-first safety design

One of 00Z's core strengths is its explicit reasoning layer. Reasoning is treated as a controlled quality and audit mechanism, not as hidden magic and not as chain-of-thought disclosure.

The repository separates reasoning into inspectable artifacts:

• harness/reasoning/reasoning_catalog.yaml defines safe reasoning methods, evidence rules, budgets, and forbidden outputs.
• harness/reasoning/reasoning_routing.yaml maps task types, risk levels, pipelines, and context budgets to suitable reasoning methods.
• harness/reasoning/reasoning_output.schema.json defines a compact, auditable output contract for reasoning summaries.
• kontext/reasoning/ provides human-readable protocol and routing references.

The design favors:

• evidence-backed answers over uncheckable internal reasoning
• dynamic method selection based on task type and risk
• explicit complexity levels: simple, standard, complex, critical
• bounded outputs with token budgets and evidence limits
• no private scratchpad persistence
• no chain-of-thought disclosure
• HITL points for review, escalation, or blocking decisions

In practice, 00Z aims to make reasoning reviewable, budgeted, and safety-gated while keeping productive writes and runtime activation disabled until explicitly approved.

Install with curl

00Z intentionally does not recommend curl | bash. Instead, download the installer, inspect it, and run it locally. Run it from the directory where you want the 00Z/ folder to be created.

curl -fsSL https://raw.githubusercontent.com/svenbecht/00Z/main/install.sh -o /tmp/00z-install.sh
less /tmp/00z-install.sh
bash /tmp/00z-install.sh
cd 00Z

The installer only downloads and extracts the repository. It does not use sudo, install dependencies, or execute project code automatically.

Optional: install a specific tag or into a custom directory once tags are published:

bash /tmp/00z-install.sh --ref v0.1.0 --ref-type tag --dir ./00Z-v0.1.0

Quick start

Run from the repository root after clone or install:

PYTHONDONTWRITEBYTECODE=1 python3 tools/zen_validate.py --check-only

Short form, if just is installed:

just validate

Typical successful output:

ZEN VALIDATE: PASS
Mode: check-only/no-write
Summary: PASS=... WARN=... FAIL=0

Optional read-only orientation:

PYTHONDONTWRITEBYTECODE=1 python3 tools/zen_onboarding.py --linear --no-color

What 00Z is

00Z is a YAML-/Markdown-first harness template for making AI-agent workflow boundaries explicit:

• declarative policies, schemas, gates, commands, and pipelines
• local validation tools with no destructive defaults
• reasoning contracts with evidence, risk, validation, and HITL fields
• adapter contracts for Pi and Claude oriented workflows
• memory and context-management concepts
• dry-run-first execution and report boundaries

What 00Z is not

00Z is not:

• a finished Agent OS
• a production API/connector platform
• a native LLM provider connector
• a native Pi or Claude runtime
• a runtime that executes live destructive writes by default
• an automatic deployment or GitHub publishing pipeline

00Z does not currently include a native LLM provider connector. Custom LLM provider support would require a future provider-runtime layer with explicit network, secret, validation, and safety boundaries.

Current capabilities

• YAML-first harness with machine-readable schemas
• Reasoning contracts for safe, reviewable reasoning summaries
• Agent and prompt templates in Markdown
• Pi and Claude adapter contracts (native_runtime_implemented: false)
• 5-layer memory concept in Markdown artifacts
• P1/P2/P3 pipeline model for intent, escalation, and review-oriented workflow design
• Local validation tooling with --check-only no-write mode
• Explicit write-boundary design where productive writes stay disabled until separate approval and gates exist

Typical workflow

Read docs/welcome.md
→ run the validation demo
→ inspect readiness/status docs
→ plan changes
→ keep mutation gated
→ run explicit snapshot/handoff after review

Project structure

core/          identity, orchestrator, boot protocol
harness/       policies, schemas, pipelines, commands, reasoning contracts
adapters/      Pi and Claude adapter contract placeholders
agents/        system agents and templates
prompts/       prompt templates and snapshots
kontext/       memory and reasoning artifacts
validation/    gate definitions, fixtures, reports
tools/         local validators and smoke tests
docs/          public and operator documentation

Core docs

• docs/getting-started.md
• docs/welcome.md
• docs/commands.md
• docs/readiness.md
• docs/release-status.md
• docs/architecture.md

Safety boundaries

• No production-ready runtime claims
• No native Pi/Claude runtime activation
• No native LLM provider connector
• No API/connector platform claims
• No automatic dependency installation
• No .env secret read/write behavior
• No built-in GitHub/deploy publishing flow
• Productive writes require explicit gates and human confirmation

Release posture

This repository is documented as an experimental local release. For current status, see:

• docs/release-status.md
• docs/readiness.md

Who it is for

• developers exploring AI-agent architecture
• builders who want safer local AI workflow templates
• reviewers and learners studying gate-based design
• people interested in explicit reasoning, validation, and safety boundaries

License

00Z is licensed under the MIT License.

About the name

00Z is short for The ZEN Agent: a compact, configurable template you can adapt to your own agent workflow.