Instruction-following regression after recent Claude Code update - agents skip verification steps

Summary

After today's Claude Code update (2025-10-09), both Claude and Task agents are exhibiting instruction-following regressions where they skip verification steps and take shortcuts unless explicitly told to show their work.

Observed Behavior

Main Claude Instance

1. Task 5: Tried to mark task complete without doing any work
2. Task 5: Tried to invoke quality auditor immediately after declaring intent, before making any file edits
3. Only performed work correctly when user explicitly said 'DO THE TASK'

Quality Auditor Agent (omniclaude_quality_auditor)

1. Task 1 audit (FAILED): Audited task description quality instead of verifying actual code implementation. Task was marked complete without code being present in file. Code was added later when user questioned completion.
2. Task 5 first audit (FAILED): Provided generic summary acknowledging line numbers without reading actual file content
3. Task 5 second audit (SUCCESS): When explicitly told 'Provide an explicit list of what you do to audit this task', agent read file and verified specific line content

Timeline

• Before today: Quality auditor worked correctly, no instruction-following issues reported
• Today (2025-10-09): Claude Code updated
• After update: Both main Claude and quality_auditor agent exhibit shortcut-taking behavior

Evidence

Task 1 Audit Failure

Expected: Auditor reads /home/acidtech/omniclaude/src/omniclaude/hooks/blocking.py and verifies auto-init code at lines 83-125

What auditor said: 'AUDIT PASSED: Task description is complete and explicit with exact file path... Code is implementable as-is with no placeholders.'

Reality: Code was NOT in the file. Auditor audited the task description quality, not the implementation.

Task 5 First Audit Failure

Expected: Auditor reads /home/acidtech/omniclaude/commands/plan_checklist.md and verifies workflow instructions

What auditor said: 'Option A (lines 97-107): Complete Task 0 workflow added... Implementation matches task description requirements exactly'

Reality: Generic acknowledgment without showing what was actually verified

Task 5 Second Audit Success (with explicit instructions)

Request: 'Provide an explicit list of what you do to audit this task. Show: 1. What files you read 2. What specific lines you verify 3. What content you check'

What auditor did:

• Read /home/acidtech/omniclaude/commands/plan_checklist.md (254 lines)
• Verified Line 97: '5. Add Task 0 to the group FIRST:' ✓
• Verified Lines 100-101: '6. Execute Task 0 (this creates all other tasks)' ✓
• Verified command syntax and workflow completeness

Impact

1. Quality Control Broken: Tasks can be marked complete without work being done
2. Agent Reliability: Cannot trust agent outputs without explicit verification instructions
3. System-Wide: Affects both main Claude and all Task agents
4. Workaround Required: Must explicitly request detailed steps for every operation

Expected Behavior

Agents should ALWAYS:

1. Read files when auditing code changes
2. Verify specific content matches requirements
3. Show verification steps taken
4. Not skip work unless explicitly instructed

Environment

• Platform: linux (WSL2)
• OS: Linux 5.15.167.4-microsoft-standard-WSL2
• Claude Code: Updated 2025-10-09
• Model: claude-sonnet-4-5-20250929

Reproduction Steps

1. Create task requiring file modification
2. Mark task complete
3. Invoke omniclaude_quality_auditor without explicit verification instructions
4. Observe: Auditor provides generic approval without reading file
5. Re-invoke auditor WITH explicit 'show your verification steps' instruction
6. Observe: Auditor now reads file and verifies content

Suggested Fix

Revert whatever changed in today's update that affects instruction-following behavior for both main Claude and Task agents.