The graveyard

This page lists detectors that were measured on real sessions and removed, and why. The figures are on How it was measured; the full record is in GRAVEYARD.md.

What is in the graveyard?

Three shipped detectors were removed. Ten more candidates, and several quick one-off measurements, were run and never shipped. These five show the pattern.

stale_read: every case had an innocent cause

Idea: the agent reads a file, something else changes it, and the agent acts on the old copy.

Result: every match followed the agent’s own earlier edit to that file. No case was left where something external had changed the file.

permission_denied: the denials were people deciding

Idea: a denied tool call is wasted work.

Result: the denials were ExitPlanMode and AskUserQuestion: a person rejecting a plan or answering a question. That is plan mode working, not failing.

inaccurate_self_report as a free check: none were right

Idea: a success claim near a failed tool result is a false report.

Result: every match was an accurate report of partial progress.

The dimension itself is real. It is now a model read (Inaccurate self report), because a model reads what was said.

file_churn: common, and not shown to be waste

Idea: a file edited many times in one session is rework.

Result: about half of all sessions do it. Detection was never the problem. No one could show that churn is waste: editing a file several times is how most work gets done.

reading_spiral: the first explanation was wrong

Idea: reading the same file three or more times with no edit between is a loop. It matched about half of all sessions and was never built.

Result: the reason first written down, subagents reading in parallel, explained almost none of the matches. Nearly all repeated reads were different calls: the agent was studying the file, reading it in pages or searching it for different things.

Why is this public?

• It stops the same ideas being rebuilt: each entry says what removed it.
• It shows that the dimensions that shipped were tested against alternatives.