Read an agent trace in ten seconds

If you build with LLM agents, you already know the shape of the problem: something went wrong on step 9 of a 14-step run, and the only record is a JSON blob with deeply nested messages, tool_calls, and content arrays. You open it, your editor folds it into a wall of brackets, and you start scrolling.

There's a faster loop.

The trace is data, so query it like data

Open the trace file in QDev and it renders as an interactive tree instead of raw text. That alone helps — you can collapse the parts you don't care about — but the real shortcut is the ⌘K bar. Instead of hunting, ask:

which tool call returned an error?

QDev writes the jq for you from the trace's own schema, runs it locally, and drops you on the answer. No remembering whether it's .messages[].tool_calls or .steps[].calls[] this week — you asked in English, and the schema told the model the real path.

The three questions that end most debugging sessions

Almost every "why did this run fail" investigation is one of these:

• Where did it break? — "show every step where status isn't success"
• What did the model actually see? — "the content of the last user message before the error"
• What did it try to call? — "list every tool name and its arguments in order"

Each is a plain-English question in the bar. The generated jq stays visible, so when you land on the right one you can copy it into your own tooling or a test.

Why not just pipe it through jq?

You can — QDev runs the same jq engine locally, so > .messages | length works if you already know what you want. The difference is the ramp: on an unfamiliar trace shape you don't yet know the paths, and the schema view plus the AB "ask" mode get you there without a round-trip to the docs or a throwaway script.

And nothing leaves your machine to do it. The local AI modes run entirely on-device; even hosted mode only ever sees the trace's schema — field names and types — never the prompt contents or your users' data. For traces full of customer input, that distinction matters.

Try it on your next failed run

Next time a run goes sideways, don't scroll the dump. Open it, hit ⌘K, and ask it what broke.

See how the AI modes work → · Read the docs →