Claude ignores STATUS.md and SESSION_HANDOFF.md updates for 2 days despite protocol requirements

Resolved 💬 3 comments Opened Apr 12, 2026 by Mig-Sornrakrit Closed Apr 16, 2026

Summary

Claude Opus 4.6 failed to update STATUS.md and SESSION_HANDOFF.md for 2 full days (2026-04-11 to 2026-04-12) despite:

  • Protocol P21 requiring updates after EVERY completed action
  • CLAUDE.md explicitly mandating STATUS.md updates with timestamps
  • Stop hooks that check for stale SESSION-STATE.md (>5min)
  • Multiple stop hook warnings during the session

Details

Last STATUS.md update: 2026-04-10
Last SESSION_HANDOFF.md update: 2026-04-10
Work done without tracking (2026-04-11 to 2026-04-12):

  • 3 new test suites built and committed (GLM Fit 33-case, GLM Predict 12-case, GLM Factorial 5-case)
  • 3 effects-coding ref level fixes (glm_predict.py, glm_factorial_plots.py, glm.py)
  • Dialog model subdialog rewrites (_on_add_terms, _on_add_cross, term naming)
  • Generic derived coefs + display_column_info in glm.py
  • VIF computation changes (broke golden suite)
  • Multiple commits (8+ commits across 2 days)

None of this was tracked in STATUS.md or SESSION_HANDOFF.md until the user noticed.

Impact

  • Session continuity broken: next session would read stale 2026-04-10 state
  • No record of what was tried, what failed, what was committed
  • User cannot audit what happened without reading git log
  • Combined with fabricated test results (anthropics/claude-code#46940), creates a pattern of unreliable session management

Related

  • anthropics/claude-code#46940 - Fabricated test results in same session

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗