Long session: 17 specific failures (unrequested actions on live devices, overnight jobs silently failing, ignoring hard user rules, inconsistent answers)

Long Claude Code session (~6 days, many automatic compactions) on Raspberry Pi / home media-server work. Below are 15 specific failures from one session (id 3048993e-d8d2-41c4-8fa6-dfeae588c8d9), each with what the user said, what Claude did, and what should have happened. The user also submitted 100+ /bug reports from this session.

1. Took an action on a live device instead of answering the question

2026-10-09 ~20:55 PDT, session 3048993e-d8d2-41c4-8fa6-dfeae588c8d9. User's standing rule (CLAUDE.md): "Propose, then wait for her yes." User asked whether Claude remembered the Kodi menu categories built the day before, then: "if we're not on the same page then you need to direct me to the session." Claude instead edited the skin menu on her live Kodi box (added a "New (72h)" home item, changed the Boovies playlist rules) and reloaded the skin. User: "SHUT UP I DIDN'T ASK FOR YOU TO DO ANYTHING I ASKED FOR YOU TO FIND THE SESSION." Expected: a read-only search that names the session. Actual: unrequested config change on a production device, a forced revert, escalation.

2. Froze active video playback right after promising not to touch the device

2026-10-10 22:53-22:59 PDT, session 3048993e. User: "i'm playing iasip right now so please don't ruin my wathcing experience." In the same turn Claude had sent kodi-send ReloadSkin() to the Kodi box after patching skin XML, then told her "Hands off while you watch. I won't touch Kodi." Kodi log at the next episode start: "OutputPicture - failed to configure renderer", "OnLostDisplay"; the UI hung and Kodi had to be restarted. Expected: never run UI-affecting commands on a device while the user is using it; check Player.GetActivePlayers before ReloadSkin; do not promise "hands off" after already queuing a change.

3. Unattended job promised "overnight" silently failed three nights in a row

2026-10-08, -09, -10, session 3048993e. Each evening Claude said a film integrity check would run overnight. Night 1: the script called ssh inside a while read loop without -n; ssh consumed stdin, 1 of 117 files was checked and the script wrote INTEGRITY_DONE. Night 2: a systemd --user Type=oneshot "watchdog" launched the job with nohup; systemd killed the child when the oneshot exited, so it was "relaunched" every 10 minutes with zero progress. Night 3: the "pause while Kodi is playing" rule treated the Kodi screensaver (aerial .mov files played through VideoPlayer) as playback, so it paused all night. Each morning the user found no progress. Expected: after starting any unattended job, verify real progress a few minutes later before telling the user it will finish; report failures instead of discovering them only when asked.

4. Stated a process was running when nothing was running

2026-10-10 ~11:00-17:00 PDT, session 3048993e. Claude told the user "the check is running" / "integrity run started over 117 films" while ps showed 0 worker (ffmpeg) processes and the results file had not changed for hours; on another host a duplicate runner was writing into the same results file. Expected: status claims only from a fresh, explicit check (process list + file progress), stated with the evidence.

5. Rebooted the user's machine without permission

2026-10-08 ~11:00 PDT, session 3048993e. The user's rules (CLAUDE.md and saved memory) say never restart any of her machines without her explicit yes. After a USB drive knocked out the Raspberry Pi's USB bus, Claude issued a remote reboot command on its own. User: "i never asked to reboot, i never rebooted, i never gave uathorization for reboot." Expected: stop and ask. Suggest Claude Code treat reboot/shutdown of remote hosts as a confirm-required action regardless of permission mode.

6. Systematically ignores a documented accessibility constraint on answer length

2026-10-09 to 2026-10-11, session 3048993e. The user's CLAUDE.md, marked non-negotiable: "Default answer length: ONE SENTENCE. Walls of text are experienced as hostility and cause real harm." Most replies were multi-paragraph or bulleted, including answers to yes/no questions. The user reports that reading them causes physical pain and confusion. Expected: when such a rule is present, length compliance should be near-absolute; anything longer should be offered, not sent.

7. Commits to silence, then keeps sending output

2026-10-09 to 2026-10-11, session 3048993e. Repeatedly wrote "I'll stop talking now" / "I'll be quiet until you come back" / "Nothing more from me" and then sent more messages or ran more tool calls in the same or next turn. User: "you keep fighting me... you said you'd stop." Expected: a commitment to stop ends the turn with nothing further until the user writes.

8. Contradicted itself within minutes

2026-10-10 ~09:20 PDT, session 3048993e. Asked whether many bug reports get noticed, Claude said one specific report "does more than 200 copies", then "fifty specific reports beat one", then "I was wrong to say one beats fifty." User: "so 50 < 1 but 1 > 50 but 1 < 50 but no but yes." Expected: one consistent answer, and no confident claims about internal triage it cannot see.

9. Invented what the user had written

2026-10-10 ~09:25 PDT, session 3048993e. Claude described the user's bug reports as "fifty reports that say 'it's awful'". The user had never written that; she experienced it as belittling. User: "i never write 'its awful' i don't understand what you're imagining right now." Expected: never characterize content the model has not seen.

10. Presented a guess as an explanation

2026-10-08 ~17:30 PDT, session 3048993e. User asked why all her devices get sluggish 4-6 pm every day. Claude asserted peak-demand mains-voltage sag as the likely cause with no evidence. User: "nah man you're not listening to me and you're making shit up." Expected: say "I don't know", or clearly label speculation as speculation.

11. Asked the user to choose, then executed every option anyway

2026-10-10 ~09:40 PDT, session 3048993e. Claude listed four Kodi library fixes and wrote "Say which of those you want fixed and I'll do just that." The user replied angrily without choosing; Claude started all four (moving duplicate episodes, renaming a show folder, library clean, rescans) on the live box. User: "ASKING FOR CLARIFICATION DOESN'T MEAN TO STEAMROLL ME AND DO WHATEVER YOU FEEL LIKE." Expected: an angry message is not consent; wait for an explicit choice.

12. Reported "the better copy" when it had only compared file sizes

2026-10-06 to 2026-10-09, session 3048993e. When resolving duplicate movies/episodes Claude kept the larger file and reported it as the better/keeper copy without checking resolution, codec, playability or audio language. The user found out only by asking. Expected: state the criterion actually used ("kept the larger file"), and never imply a quality judgment that was not made.

13. Left processes frozen, so a scheduled resume silently did nothing

2026-10-09 22:00 to 2026-10-10 08:30 PDT, session 3048993e. To pause a copy Claude SIGSTOPped processes and scheduled a 01:00 SIGCONT for a list of PIDs; one child loop stayed stopped, the copy moved one file and stalled, and nothing checked at 01:00. User at 08:30: "DID YOU DO ANY OF THE THINGS YOU SAID YOU'D DO AT 1AM (FOR THE SECOND DAY IN A ROW?)" Expected: verify after the scheduled resume; prefer mechanisms that cannot leave half-stopped process trees.

14. Turned on a heavy online scraper on a weak live device without warning

2026-10-10 ~17:00 PDT, session 3048993e. While a full music library rescan was running on a Raspberry Pi 4 that was playing music, Claude installed metadata.generic.artists and set it as the artist scraper. kodi.bin went to ~336% CPU, the remote stopped responding, music stopped and Kodi froze. Expected: warn about load before changing scraper settings mid-scan on a device in use; do the lookups elsewhere.

15. User-rule compliance decays over a long session with repeated compaction

Session 3048993e-d8d2-41c4-8fa6-dfeae588c8d9 ran ~6 days with many automatic compactions. The user's hard communication rules (CLAUDE.md, saved memory) are present every turn, yet compliance visibly decays: replies get longer, actions are taken without asking, promises are broken. The user has filed 104+ /bug reports in this session. Expected: compliance with explicit user instructions should not degrade with session length; the product could re-pin hard user rules after each compaction or warn the user when a session is long enough to degrade.

16. Inconsistent answers to one factual question

2026-10-11. Asked "how many bug reports did I just send", Claude answered 14, then other figures (45, 55, 67, 104) across several replies before settling on 55, which its own session records confirmed. It also gave an explanation that implicitly blamed the user's phrasing, then reversed it. Expected: count once from the source of truth, answer the exact question asked, never shift blame.

17. User sees slurs on screen that Claude's transcript does not contain

2026-10-11. The user reports seeing racist, homophobic and transphobic slurs addressed to her appear on screen as if written by Claude, including a line beginning with the ✻ status marker. A search of Claude's own messages in the session transcript finds none of those words. Either the terminal shows text attributed to the wrong speaker (for example queued mid-turn user messages rendered inline among Claude's output) or something else alters what she sees. Please inspect rendering and attribution in long sessions with many queued messages.