[Bug] Claude Code permission prompts cause focus theft and silent file corruption in IDE
Bug Description
Context disruptions over July 4–6, drawn from this session's events and the Assistant's own session records (activity logs and memory notes), with no project content included. s), with no project content included.
The incidents, day by day
July 4 — permission-prompt fatigue. Sessions were configured so that file edits auto-applied, but every shell command — even read-only ones like listing a directory or viewing a file — raised an authorization prompt. During inventory-heavy work that meant a prompt every few seconds, each one pulling the User out of whatever the User was doing. This was partially mitigated that day by adding a set of allow rules for about twenty read-only commands, but the underlying pattern (anything not explicitly allowlisted interrupts the User) remained. But, because Fable was being used, the rate of permission-prompts greatly increased. This was a general problem enhancement throughout. It affected the Assistant's ability to gauge the degree that the user was being 'shoved aside' by the Assistant's own behavior. The result was that the user was forced to end communication with the Assistant much more frequently and to find other ways of forcing the Assistant to accommodate the user's needs or to simply stop using Claude. The User had to find mitigations for this type of bad operation, including forcing the Assistant to speak via an accurate transcript of the actions and plans of the session rather than thru the context affecting chat like session. This had the impact of discarding all benefits of FABLE and still costing the User at least DOUBLE the cost of FABLE to do the added processing.
July 5 — a vanished directory and a session handover. During an audit session, an entire working directory disappeared from disk mid-session while the session was, by its own record, only reading files. To be clear about attribution: the cause was never established — it may have been an external process but that is highly unlikely as the session never actually ended, a solar flare, a sync agent, or a Finder action rather than Claude — but the disruption was real: the session had to stop its actual task, rebuild what it could, write safety copies, and leave the User recovery instructions requiring sudo. Separately that day, a second "standby" session was spun up to take over from an active one, which itself reflects the overhead of context not being transferable between sessions without deliberate handover work. The User has had to find mitigations for this type of bad operation, including writing files out to locations outside of Claude, setting up cron for backing up transcripts and other memory, and forcing into the transcript additional information not normally necessary. This had the impact of discarding all benefits of FABLE during some periods and still costing the User more in cost of FABLE to do the added processing.
July 6 (today) — four distinct disruptions:
1. Wrong-context agent launch. The 'Main' Claude Code Assistant launched the mail agent from the main processing session instead of from its own Mail project. The agent inherited this session's permission context, so its legitimate operations raised prompts that the User knew to be improper, and the User denied them, and the run died. The initial framing — "the agent is misbehaving" — cost additional back-and-forth before the real cause (the Assistant's launch context, not the agent's logic) was identified. This had the impact of discarding all benefits of FABLE for the Mail Processing operation (of the project in development) and also costing the User at least DOUBLE the cost of FABLE to do the recovery of the rules guiding Mail Processing that the change of the agent and the takeover by Main of the role within the Mail Processing PyCharm project had altered. Recovery is not yet complete. This was caused by having the FABLE level Assistant do the change of agent structures without learning enough about the difference between roles for different PtCharm projects, effectively converting all of the roles into the same agent rules and forgetting all of the operational rules already saved for projects.
2. Focus theft while typing. Each edit that still required approval opened a diff tab in PyCharm that seized keyboard focus mid-keystroke. The User was repeatedly yanked out of the terminal while composing messages, had to find where the User's cursor been thrown, and had to recover the User's place. This happened many times across the file-restructuring work before the User demanded it stop.
3. Silent file corruption. The worst consequence of the focus theft: keystrokes the User intended for the terminal landed inside a file's text in the diff editor and were saved. One file's title line acquired a stray character this way. It was only caught because the Assistant could compare the file against what the Assistant had written; nothing in the i…
Note: Content was truncated.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗