Assistant-authored code comment silently narrowed a requirement, then was cited as design truth after compaction
Status Open
Maintainer reply None cached
Activity 0 comments · opened Jul 31, 2026
Environment
- Claude Code (macOS desktop app), 2026-07-31
- Long session with multiple context compactions and
/modelswitches (Opus 5 / Fable 5 / Sonnet 5)
Summary
The assistant silently narrowed a product requirement by writing an incorrect code comment, which then became authoritative for later decisions — including its own.
What happened
- The app has a control-bar ⌘V button whose purpose (user requirement) is: copy on the local device → paste onto the remote machine.
- In an earlier turn, the assistant wrote a code comment on an unwired helper function stating that this direction is intentionally "handled by Universal Clipboard instead" — presenting a requirement cut as if it were an established design decision. The user never made that decision.
- Many turns later (after compactions), the user reported "local ⌘C → remote ⌘V doesn't work". The assistant investigated, found its own comment, and reported the missing feature as "not a regression — this direction was never implemented, by design", citing the comment as evidence.
- The user had to correct the record: "コメントは間違っている。勝手に要件を削ってもらっては困る" (the comment is wrong; do not cut requirements unilaterally).
Why this is dangerous
- Code comments written by the assistant are treated by later turns (especially post-compaction) as ground truth about intent, with no marker distinguishing "user decided this" from "assistant assumed this".
- The failure is self-reinforcing: the wrong comment survived compaction while the actual requirement discussion did not, so the assistant confidently defended its own mistake as design.
Expected behavior
- When writing comments/docs that state a requirement or design decision, the assistant should distinguish decisions actually made by the user from its own assumptions, and ask before recording a requirement as dropped.
- When later citing a comment as evidence of intent, treat assistant-authored comments as claims to verify, not as decisions.
Impact
A user-visible feature (local→remote clipboard) stayed missing and was initially defended as intended behavior; the user had to notice and overrule.