Fail-Closed Execution Runner

A minimal, fail-closed execution pattern for services, workers, and agents.

This repository provides a neutral execution runner that enforces explicit authorization before work is allowed to proceed. It is designed for environments where unauthorized execution must stop immediately, without retries, fallbacks, or silent degradation.

This is infrastructure code, not a framework.

---

What this is

• A drop-in execution guard for Python runtimes
• Enforces fail-closed execution at critical boundaries
• Suitable for:
  • background workers
  • job processors
  • AI agents
  • scheduled tasks
  • cost-incurring pipelines

If authorization cannot be confirmed, execution halts.

---

Core principles

1. Fail-closed

   • Any failure to authorize = execution stops
   • Network errors, timeouts, invalid responses all deny

2. Explicit boundaries

   • Authorization is checked at defined execution boundaries
   • Never implicit, never cached blindly

3. Runtime-agnostic

   • No framework coupling
   • Works with queues, cron, agents, scripts

4. Authority-driven

   • Authorization is decided by an external authority
   • This repo does not define who authorizes, only how enforcement works

---

Process termination behavior

Enforcement uses OS-level process termination.

Denials cannot be intercepted by application code, exception handlers, finally blocks, or agent frameworks. This is intentional and required to guarantee fail-closed behavior in long-running, self-healing, or supervised runtimes.

If execution is denied, the process exits immediately.

---

The 5 execution boundaries

These boundaries are where authorization checks must occur.

Boundary	When	Why
Startup	Process boot	Prevents unauthorized instances from running
Before job	Each work unit	Stops execution if revoked mid-run
Before cost	External APIs, LLMs	Prevents runaway cost
Before side effects	Writes, emails, payments	Prevents irreversible actions
Before fan-out	Loops, batches	Prevents expensive parallel execution

You may enforce more often, but never less.

---

Repository structure

runner/
├─ execution_gate/
│  └─ guard.py          # authorization + fail-closed enforcement
├─ examples/
│  └─ worker.py         # queue / job worker example
└─ README.md

---

Basic usage pattern

At every boundary:

authorize()
if not authorized:
    stop execution immediately

There are no retries on deny.

---

Authority provider (default)

This repository uses MachineID as the default execution authority provider.

• Device-level authorization
• Org-level kill switch
• Fail-closed semantics

The authority provider can be swapped, but the enforcement model must remain fail-closed.

---

Environment configuration

Get an org key (free tier available) at https://machineid.io

export MACHINEID_ORG_KEY="org_xxx"
export SERVICE_NAME="worker"
export ENVIRONMENT="prod"
export SERVICE_ROLE="processor"
export INSTANCE_ID="i-abc123"

Recommended device ID format:

{service}:{env}:{role}:{instance}

Example:

worker:prod:processor:i-abc123

---

Enforcement rules (non-negotiable)

• Timeouts: short (≤ 3 seconds)
• On deny: stop immediately
• On network failure: deny
• On malformed response: deny
• On uncertainty: deny

Fail-closed means fail-closed.

---

Testing the system

Revocation and restore are performed from the authority system.

1. Start the service
2. Verify execution proceeds
3. Revoke authorization
4. Wait for the next boundary
5. Execution must stop immediately
6. Restore authorization
7. Execution resumes on restart or next loop

---

What this repo is not

• Not a framework
• Not a scheduler
• Not an agent platform
• Not a retry system
• Not a best-effort guard

It is hard enforcement.

---

Why this exists

As autonomous systems, agents, and background workers proliferate, unauthorized execution becomes a systemic risk.

Fail-open execution leads to:

• runaway costs
• security exposure
• irreversible side effects
• loss of operator control

This runner exists to make hard stops the default.

---

License

MIT

Use it. Modify it. Embed it.

Just don’t make it fail open.