Skip to content

Your first finding

This page walks through one real finding as the pane shows it, part by part, and what each of your answers to it does. A finding says where to look and why; whether it is right is yours to decide.

The example is from the public test data of claude-code-log (MIT).

The row

The Misread intent kind, 3 times in 2 sessions, open to one finding: the developer's words, the date, the session and the time.

  • The kind comes first, with how often it happened: here, misread intent, 3 times in 2 sessions.
  • Each finding under it is headed by the words it is about, cut to fit. Here they are the developer’s own: “You can set group.visible using the vis library which is less work”.
  • Under the words: the date, the session and the time.

The turns

Three turns: the agent starting work, the developer interrupting with the cited words, and the agent agreeing.

Opening the finding shows the turns around the cited one, read from the transcript where Claude Code keeps it. The cited turn is marked ▶. Each agent turn says how many tool calls it made and how many failed.

Show conversation opens those messages in full. More around this adds three messages on each side.

Why

The Why line: the developer asked to hide the group, the agent started per item, and the developer redirected it.

Why is the model’s one-line account of what went wrong.

How to read it

The turns are the evidence; Why only summarises them. Here:

  1. The developer asked to hide a timeline group, not every item in it.
  2. The agent started hiding the items one by one.
  3. The developer interrupted, pointing at the library’s group.visible.
  4. The agent agreed and switched.

The developer had to correct the course. A model finding can be wrong: if the turns do not bear it out, press Finding is wrong.

The fix

The fix: a rule the agent reads in every session of this project, its text, the file it is added to, and Apply.

Each finding offers a change for your agent, with what kind of change it is, its exact text, and where it lands. Apply drafts it into your prompt box; your Enter is the yes. Fixes and Apply has the details.

Your answers

Three answers: Finding is wrong, Not the agent's fault, and Hide.

answeruse it when
Finding is wrongit did not happen as described
Not the agent’s faultit happened, but it was not the agent’s mistake
Hideyou have seen it

All three hide the finding and take it out of the pane’s counts, Before and Since included. Your answer is kept as a record of what you said. With no answer, the finding stands and its fix stays on offer.

Not the agent’s fault is offered only when the cited words are your own, as here. When they are the agent’s, only Finding is wrong and Hide are shown.

Each answer takes effect at once and offers Undo. Nothing is deleted: Restore all on the Health tab brings hidden findings back.

When nothing is flagged

If nothing has been found, the Findings tab says so, with how many sessions were read:

Nothing flagged in 40 sessions read.

That is a result. The checks ran over those sessions and found nothing they could tie to a turn. Nothing is added to fill the pane.

Across your history

Open full report at the bottom of the Findings tab, or /vibesenpai:report, writes every finding in your history to one HTML file.