A customer reports that a docs page renders empty. Before anyone ships a fix, the workspace needs one answer: is the page broken, or is it just not server-rendered? This example shows a bounded, read-only investigation that produces that answer.

​
Diagnose in layers, cheapest first

The first check costs nothing: fetch the route without executing JavaScript. A server-rendered page returns its content in that fetch. An empty result with a well-formed shell is the signature of a client-rendered page, not a broken one: the server answered, the content assembles in the browser. Only then open a browser and look at what actually renders: console errors, failed network calls, and the rendered tree. Each layer either names the fault or eliminates one.
The most common false positive: a route fetched without JavaScript looks broken when it is fine in a browser. Confirm the render path before you believe the report.

​
Keep the investigation read-only and bounded

Run it in a read-only context so nothing changes while you look. Read-only work needs no tool approvals and cannot mutate the workspace. Then bound the agent’s freedom in the prompt. A contract like this one works:

​
Set stop conditions before you start

An investigation ends when any of these happens:
  • Three consecutive tool results produced no new fact.
  • A step repeats a request that was already denied.
  • The next step would need write access. Hand back with the findings and the open questions.
Without stop conditions, an agent keeps spending steps because steps are available. The bound is what turns a session into an investigation.

​
Turn the finding into a record

The diagnosis, the evidence for it, and the open questions go into the issue, not into a chat that scrolls away. If the report was a false positive - the page works, the original check was made without JavaScript - the record prevents the next person from repeating the wrong test. If it was real, the issue already carries the evidence a fix needs.

​
What good looks like

  • The render path was confirmed before any conclusion.
  • The investigation ran read-only, inside a stated step cap.
  • Every observation in the report names its source.
  • The finding, evidence, and open questions live on an issue.