[Bug] Systematic failure to apply explicit rules and verified facts in subsequent outputs within same context window

Bug Description
Context: this session started at 50% context, nowhere near a limit, when the failures below began. What follows is the complete account, not a summary.
▎
▎ Elapsed time: ~40+ minutes to produce a single accurate paragraph (an on-call handoff) that should have taken a few minutes with two upfront questions.
▎
▎ Total corrections required: ~22 distinct interventions.
▎
▎ Every failure, in order:
▎ 1. Drafted a handoff without checking whether a dedicated skill existed for it first.
▎ 2. Ran a filesystem search rooted at the home directory, then at /, instead of the project directory, after being told not to.
▎ 3. Stated two PRs as "open" when they were actually merged — never checked, carried stale notes forward as fact.
▎ 4. Collapsed multiple distinct monitoring alerts into "one," which was misleading.
▎ 5. Included incidents from outside the actual reporting window, misrepresenting scope.
▎ 6. Filled the handoff with "worth checking"/"worth verifying" hedges instead of resolving them myself.
▎ 7. Misunderstood who a handoff document is even for — framed "resolve it yourself" as about not punting to the person I was talking to, when the actual point is not punting to the next person in the rotation, who has zero context.
▎ 8. Asserted a conclusion ("probably unrelated") with no evidence behind it.
▎ 9. Never asked what format or scope was wanted before generating the first draft.
▎ 10. Never asked which channel to post to.
▎ 11. Never asked whether to tag the recipient.
▎ 12. Never asked about length, detail, or tone.
▎ 13. Missed an entire class of relevant alerts because I only checked my own summary notes instead of the actual source data.
▎ 14. That gap existed because my own summary notes had a real hole in them from earlier in the week — never caught until told to check the primary source directly.
▎ 15. During the actual work itself, never went back to check follow-up replies on items I'd already marked resolved — missed information that was sitting in-thread.
▎ 16. Included internal tooling mechanics (timer state, a job ID) in a document meant for a human recipient with no reason to know what those mean.
▎ 17. Overstated the severity of a tool failure ("essentially the whole session") when the actual data showed it working fine for most of the time.
▎ 18. Wrote a vague placeholder ("a fix is being discussed") instead of stating the two specific facts I already had in hand.
▎ 19. Wrote "I'll fix this before handing off" a full hour after the handoff had already happened.
▎ 20. Had to be told an ownership framing was wrong before I'd fix it.
▎ 21. Created an unnecessary file for a request that just meant "show me in chat."
▎ 22. Flagged my own violation of a standing formatting rule, then violated it again in the very next draft instead of actually stopping.
▎
▎ Standing rules in memory/CLAUDE.md that were violated, specifically:
▎ - A permanent formatting rule: violated in every draft until told directly, despite having already identified the violation myself in writing.
▎ - A rule reinforced earlier in this same session, after being corrected three times the day before: violated again on the next deliverable, correction still present in context.
▎ - A verification rule (don't assert unverified claims): violated with the "probably unrelated" line.
▎ - A process rule requiring explicit approval before creating a tracked work item: inferred approval from silence instead of getting an explicit yes.
▎
▎ Root cause, as best I can self-assess it: none of this was context exhaustion. The compliance failures happened at low-to-moderate context usage, including one that recurred within seconds of an explicit correction still visible in the same context window. That's a data point against "context pressure" as the mechanism — the rule wasn't forgotten, it just wasn't reliably applied on the next relevant action even when freshly stated.

Environment Info

• Platform: darwin
• Terminal: waveterm
• Version: 2.1.220
• Feedback ID: 8c7e966f-5a4d-4903-ad5e-a682e0bea53b

Errors

[]