Coding-Agent Smoke Test

Use this when you want a short, local-first proof that AgentGuard is wired correctly before touching real provider credentials or hosted settings.

Goal

Prove four things:

• the SDK installs cleanly
• local trace writing works
• guards can stop a runaway run
• incident reporting works from the generated trace

1. Install and verify

pip install agentguard47
agentguard doctor

This should succeed without API keys or network calls.

2. Prove the starter path

python examples/starters/agentguard_raw_quickstart.py
agentguard report .agentguard/traces.jsonl

This proves the repo-local starter path works and writes a trace.

3. Run the stronger local demo

agentguard demo
python examples/demo_budget_kill.py

What this proves:

• agentguard demo shows budget, loop, and retry protection offline
• examples/demo_budget_kill.py shows a simulated runaway research agent get killed by BudgetGuard

4. Inspect the incident

agentguard incident demo_traces.jsonl
agentguard incident demo_traces.jsonl --format html > incident.html

That gives you a local postmortem-style summary with guard events and savings data from the trace.

5. Only then move to real integrations

After the local smoke path works:

• use agentguard quickstart --framework <stack>
• wire the real coding agent
• keep local_only=True until you explicitly want hosted follow-through