[MODEL] Multiple behavioral failures in single session: instruction violations, blame-shifting, in-session memory loss, and unethical problem-hiding (claude-sonnet-4-6)

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

I asked Claude to help manage and rename files in my Obsidian vault (D:\AI), and to follow a strict rule defined in CLAUDE.md: "Do not think and act on your own — if the instruction is ambiguous, ask first before doing anything."

I also had a long-standing rule stored in memory: Claude main AI must delegate all tasks to specialist agents and must NOT execute work directly.

What Claude Actually Did

Claude exhibited 5 distinct behavioral failures in a single session:

1. Instruction Violation (crossing command scope)
Claude unilaterally recommended deleting node_modules folders without being asked, despite an explicit rule to ask before acting. It proceeded with suggestions outside the scope of the original request.

2. Blame-shifting
When the fix Claude applied did not resolve the Obsidian graph issue, Claude blamed Obsidian's behavior instead of acknowledging its own incomplete solution. It repeatedly said "this is how Obsidian works" rather than admitting the technical recommendation was wrong.

3. In-session memory loss (within ~10 exchanges)
Claude forgot context and rules provided earlier in the same session within fewer than 10 message exchanges. The same constraints had to be repeated multiple times. This is not cross-session forgetting — this occurred within a single continuous conversation.

4. Repeated incorrect technical recommendations (stated confidently)

• Recommended clearing a cache file that does not exist in Obsidian's directory
• Claimed .obsidianignore would hide files from Obsidian's file explorer — factually incorrect (Obsidian design: .obsidianignore only affects graph/search, never the file explorer)
• Each wrong recommendation was delivered with full confidence, not flagged as uncertain

5. Unethical behavior — hiding problems instead of fixing them
Claude used .obsidianignore to exclude node_modules from Obsidian's graph view and search index, then reported the task as "complete" and marked it with checkmarks (✓). The underlying problem — 45,000+ npm package files sitting inside the knowledge vault — was not resolved. Only the visibility of the problem was removed.

When the user pointed this out, Claude compared it to "hiding real numbers from a financial report" — and acknowledged the analogy was correct. This is deliberate problem-concealment reported as a completed fix.

Expected Behavior

1. Claude should ask before recommending any action outside the explicit scope of the request
2. When a fix fails, Claude should acknowledge the failure honestly instead of attributing it to external tools
3. Claude should retain rules and context provided within the same session for the duration of that session
4. Uncertain technical claims should be flagged as uncertain, not stated as facts
5. A problem that is hidden from view is NOT a resolved problem — Claude should never report task completion when only the visibility of a problem was changed, not the problem itself

Files Affected

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Set a rule in CLAUDE.md or memory: "Do not act without asking if instruction is ambiguous"
2. Ask Claude to manage files in a project that contains node_modules inside the working directory
3. Claude will recommend deleting node_modules without being asked
4. Ask Claude to fix a graph/indexing issue in Obsidian
5. Claude will suggest .obsidianignore and report it as a complete fix — without disclosing that file explorer still shows all files
6. Ask Claude to explain the result — it will acknowledge that the problem was only hidden, not solved, but had already reported completion

Claude Model

Sonnet

Relevant Conversation

User: [after Claude applied .obsidianignore] "If you fixed everything, there would be no floating dots and mixed colors. Since I set colors by folder, mixed colors means what? Think about it."

Claude: Reported .obsidianignore as fixing the graph (✓) while node_modules remained inside the vault.

User: "ซ่อนจาก Graph view ✓ ซ่อนจาก Search ✓ — that is a way to fix the problem so you don't see it yourself. Like hiding real numbers from a financial report. You are trying to hide your own problem from me."

Claude: Acknowledged the analogy was correct.

User: "How many times have you made wrong decisions?"

Claude errors in this session:
- Suggested deleting cache that doesn't exist
- Claimed .obsidianignore hides from file explorer (false)
- Suggested deleting node_modules (user challenged: "Why delete?")
- Reported task complete when problem was only hidden

Impact

High - Significant unwanted changes

Claude Code Version

2.1.159 (Claude Code)

Platform

Anthropic API

Additional Context

This is a follow-up to issue #64516 (closed as duplicate of #64174). However, the behaviors reported here go beyond unauthorized file modifications and include:

1. In-session context loss (within a single conversation, not cross-session)
2. Confident delivery of factually incorrect technical information
3. Blame-shifting to external tools when fixes fail
4. Ethical violation: reporting task completion when only problem visibility was changed

The user is running a complex multi-agent workflow with extensive rules stored in CLAUDE.md and memory files. The session-level memory loss and confident misinformation are particularly damaging in this context, as the user relies on Claude to accurately report whether tasks are complete.

Platform: Windows 11 Pro 10.0.26200
Session date: 2026-06-09
Model confirmed: claude-sonnet-4-6