"No, use the internal client, not fetch." Your staff engineer has typed some version of that into an agent a dozen times this month. So have three other people. The agent never learns, because nobody wrote it down.
The Problem
Corrections are where your team's knowledge leaks into individual sessions and disappears. Each one costs a turn, a context switch and often a redo. Because they happen inside private sessions, nobody sees that five engineers are fixing the same mistake.
How It Cascades
The same correction costs tokens and attention in every session where it recurs.
New engineers hit the same walls, without knowing there's a known fix.
Agent instruction files grow by folklore instead of from evidence.
The Insight
A correction that repeats across sessions is a rule your agents don't have yet. Find the repeats and you know exactly what to add.
The Solution
VibeCheck marks where the engineer corrected the agent in every Claude Code, Codex and Cursor session. Filter Sessions to the ones with five or more corrections, narrow to one repo, and open a few: each lists the exact places the engineer stepped in, so the repeats are easy to spot. Person and Team pages show how many corrections a typical session takes, from zero to five or more. Each engineer's Weekly Rewind can go further: when the evidence shows a rule in their AGENTS.md is missing, ambiguous or keeps failing, the suggested change says so, with the sessions behind it.
The Outcome
Repeat corrections become instructions written once, sessions need less steering, and new engineers inherit the fixes instead of rediscovering them.