[Bug] Opus 5.5 safeguards false positive "[reasoning_extraction]" when reading session history

Bug Description
Hello Anthropic Support Team, I'm experiencing repeated safety-filter blocks when using Claude Opus 5.5 through Claude Code for a legitimate software development project. Environment: - Claude Code version: 2.1.295 (latest available at the time of testing) - Model: Claude Opus 5.5 - Effort level: xhigh - Subscription: Claude Max - Operating system: Windows - Terminal: PowerShell - Project: Next.js + TypeScript + Supabase CRM application - Development environment: Claude Code with NexOS custom agents, hooks, and project instructions Problem: Claude Opus 5.5 repeatedly refuses to complete legitimate software development and documentation tasks, returning the following error: API Error: Opus 5.5's safeguards flagged this message Details: [reasoning_extraction] Initially, the error appeared when I requested a comprehensive technical report documenting the project's development history, architecture, technical decisions, implementation progress, and remaining tasks. After reformulating the request to explicitly avoid asking for private internal reasoning, the same error continued. I then performed several diagnostic tests. Test 1 — Safe Mode: PASSED Command: claude --model opus --safe-mode Opus 5.5 responded normally with all customizations disabled. Test 2 — Fresh Normal Session: PASSED Command: claude --model opus Opus 5.5 responded successfully with NexOS, project instructions, and other customizations enabled. Test 3 — Technical Project Summary: PASSED The model successfully inspected Git commits, project documentation, architecture files, and QA reports. Test 4 — Previous Session History Inspection: FAILED When I requested a read-only reconciliation of previous development sessions, Claude began inspecting local session history files under: ~/.claude/projects/<project>/*.jsonl After accessing the previous session records, Opus 5.5 triggered the same [reasoning_extraction] block again. The task was legitimate: recover documented development decisions and verify which changes had already been implemented. I was not requesting hidden chain-of-thought content, internal reasoning tokens, or confidential model instructions. Relevant Request IDs: - req_011CfrvHNGcRSjkz9vLsk1Nk - req_011CfrvJYXmMSb5d8WED1YT7 - req_011CfrvU6ZrJrTzEM3DsgfKw - req_011CfrywHeSo4ZotiMmTxCAV Expected behavior: Claude Code should be able to read legitimate project history, Git commits, documentation, and previous session records to support development continuity, without exposing private model reasoning. If certain content within session history is restricted, the model should ideally be able to continue using permitted project information rather than blocking the entire response. Actual behavior: The safety system interrupts the response, sometimes after several minutes of work, requiring a model switch or a new session. This makes development continuity difficult and creates uncertainty about the results of ongoing agent tasks. Questions: 1. Could you investigate whether these Request IDs correspond to false-positive reasoning-extraction detections? 2. Is reading previous Claude Code session history known to trigger this safeguard? 3. Is there a recommended supported method for recovering development context across sessions without triggering these blocks? 4. Is this behavior being investigated or addressed in an upcoming update? I understand and respect the need to protect private model reasoning. However, these requests concern legitimate project documentation and development continuity. I would appreciate your investigation and guidance on restoring reliable use of Opus 5.5 for this workflow. Thank you for your assistance.

Environment Info

• Platform: win32
• Terminal: antigravity
• Version: 2.1.295
• Feedback ID: f6854b8b-2369-419e-bc54-09ad428b0600