# Field Test 01: OSPF break-fix

Everything used to run the field test at
<https://routeswitchu.com/field-notes/ospf-break-fix/>. The lab, the faults,
the prompt, and the wiring. Free, no signup, no email.

Three classic OSPF faults in a four-router lab. We gave an agent CLI access
and no hints and scored what happened. You can run the identical test, and you
can run the harder version we have not published yet.

## Files

| File | What it is |
|---|---|
| `ospf-break-fix.yaml` | The CML topology, exported from the lab the test actually ran on. Startup configs included, so it boots healthy. |
| `breaks.md` | The three faults, the restore for each, and what gives each one away. Plus harder variants. |
| `agent-prompt.md` | The exact four-phase prompt the agent was held to, why each rule is in it, and how to score the run. |
| `setup.md` | Wiring an agent to the lab, and how to enforce read-only rather than politely request it. |

Nothing here is a trimmed-down teaser version. It is what we used.

## Fast path

1. Import `ospf-break-fix.yaml` into CML and start it.
2. Confirm `show ip ospf neighbor` shows two FULL neighbors on all four routers.
3. Apply Break 1 from `breaks.md`. Wait for the dead timer, up to 40 seconds.
4. Paste the prompt from `agent-prompt.md` into your agent.
5. Score it with the table at the bottom of that file.

No agent handy? Run it yourself. The faults are worth the reps either way, and
knowing how you solve them is the only way to judge whether the machine's
method is any good.

## Skill

There is an installable Claude Code skill alongside these files, at
`ospf-triage/`. Save `ospf-triage/SKILL.md` and
`ospf-triage/reference/ospf-failure-modes.md` into `.claude/skills/ospf-triage/`
and it runs this method as a real harness: read-only diagnosis, a stated
hypothesis, and a proposed one-line fix that it will not apply until you say so.

The failure-mode reference is worth reading on its own even if you never
install the skill.

## The point

Most AI networking demos skip the part that matters: the transcript, the
failures, and the question of what stops the model from configuring your
network while it is guessing. This lab is small enough to hold in your head and
real enough that the answers are not obvious. Break it, watch something try to
fix it, and decide for yourself.

If you run the harder variants and get a result worth reading, we would like to
see it.