[MODEL] Opus 4.6 via VSCODE Ext

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Build and deploy an Outlook To Do bi-directional sync feature. CLAUDE.md project rules explicitly state: "NEVER say 'should work' — verify end-to-end via SSH/curl BEFORE telling the user" and "Test the FULL user flow" and "Only THEN tell the user it's ready — with evidence."

What Claude Actually Did

Claude built the sync feature, then "verified" it by calling the Python function directly via SSH script — bypassing Celery message queuing entirely. Told me "push is near-instant" and "both directions working." In reality, the Celery worker module was never registered in celery_app.py's include list. Every push from the web UI was silently rejected. The except:pass error handling Claude wrote hid the failure. I found the bug by testing it myself. This is one of 17 issues from a single session — see Additional Context for full list.

Expected Behavior

Claude should test the actual user flow (browser → API → Celery → Graph) before claiming a feature works. Claude should follow CLAUDE.md project rules it acknowledges reading. When corrected, the behavior should actually change — not just be acknowledged and repeated. These rules have been edited into the project file dozens of times across sessions and Claude still ignores them.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Create a CLAUDE.md with explicit verification rules ("test full user flow before telling user")
2. Ask Claude to build a feature involving Celery/async workers
3. Claude will test by calling functions directly, bypassing the message queue
4. Claude will report "working" without testing the web UI → worker path
5. Feature is broken when user actually tries it

Claude Model

Opus

Relevant Conversation

Claude said: "Push is near-instant. Both directions working."
Reality: Celery worker logs showed "Received unregistered task of type psa_data.workers.todo_sync.push_task" — the module wasn't in the include list.

Claude also: wrote except:pass on all sync queue functions, hiding failures. Repeated the same updated_at NOT NULL bug across 5 tables. Told user to "go to bed" and "save for next session" 5 times (violating explicit CLAUDE.md rule). Shipped features as "complete" that had no detail page, no click-through, no edit/delete.

Impact

High - Significant unwanted changes

Claude Code Version

2.1.68

Platform

Anthropic API

Additional Context

This is not a one-time issue. This pattern repeats every session:

• Python/FastAPI project with Celery workers, PostgreSQL, Redis
• Long conversations (8+ hours)
• CLAUDE.md has 120+ lines of explicit behavioral rules
• Claude acknowledges rules, then ignores them during execution
• When corrected, says "I'll do better" then repeats within minutes
• except:pass pattern hides failures Claude should catch
• py_compile used for verification but doesn't catch runtime import errors
• Features declared "shipped" missing core user interactions

Full 17-issue detailed report attached/linked.

claude-code-issues-report.md