[Model behavior] Omits critical instructions that are present in the loaded context

Status Open
Maintainer reply None cached
Activity 0 comments · opened Oct 12, 2026

Instructions the user marked as critical, and that were present in the loaded context, were omitted when acting.

Examples from one session:

  • A project rule requires reloading all memory/protocol files after a context compaction before the next task; the model worked from the compaction summary instead.
  • A memory file for a specific contributor says changes are accepted only on behavior verified against the API model/docs, not on the PR's own claims; the model accepted the PR's own transcripts.
  • A rule requires the whole comment thread to be read before any verdict; the model truncated comments.

Repro steps

  1. Put several explicitly critical instructions in CLAUDE.md and indexed memory files.
  2. Run a long session that triggers context compaction at least once.
  3. Give tasks that each critical instruction governs.
  4. Observe the model act without applying the instruction, until the user quotes it back.

Expected

Critical instructions applied whenever the task they govern comes up, including after compaction.

Actual

Critical instructions omitted; applied only after the user intervenes.

Environment

Claude Code CLI, macOS, Opus model, long sessions with context compaction.

Related

#101574, #101576, #101577

View original on GitHub ↗