Claude Code AUP filter false positive on legitimate self-audit documentation

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

I'm using Claude Code to help me build self-audit tooling for my own OpenClaw deployment on my own Mac. The work involves writing documentation describing how the audit detector catches LLM rule violations.

The AUP filter blocked me when I asked Claude to verify a Google Drive upload by downloading the file content for SHA256 comparison. The downloaded file was a markdown doc I had just uploaded — my own writing about my own system.

The trigger keywords appear to be combinations like "binance + strategy", "violation detection", "monitoring + Telegram alerts" appearing densely in legitimate technical documentation.

Suggestion: AUP classifier should consider context — files being processed within the user's own filesystem / own Drive folder shouldn't be classified the same as user-generated requests targeting external systems. Right now this creates a chilling effect where users have to self-censor technical documentation about their own software.

Reference: AUP block occurred at approximately 2026-05-03 01:50 UTC+8.

中文補充（如果你想加）：

此外，Claude Code 在桌面工作時權限彈窗頻率過高、嚴重影響長任務體驗，希望提供更細緻的「session-level trust」機制。

Proposed Solution

Suggested solution:

Context-aware AUP classification: When the content being processed (uploaded, downloaded, edited) originates from the user's own local filesystem or their own cloud storage, weight the trust signal higher. A file the user just uploaded 30 seconds ago shouldn't be classified the same as a fresh user prompt targeting external infrastructure.
Round-trip exemption: When Claude Code downloads a file it (or the user) just uploaded within the same session for verification purposes (e.g., SHA256 hash check), bypass deep content classification — the content is already known to be user-authored.
Better error messages: Current "Try rephrasing or attempting a different approach" gives zero signal. Tell the user which keyword combination tripped the filter so they can self-correct, or explicitly say "this looks like a false positive — file an issue".
Session-level trust escalation: After N successful tool calls in a session that didn't trigger AUP, raise the threshold for borderline cases. Right now every tool call gets classified independently, which creates absurd inconsistency where call #50 in a session about my own audit log gets blocked while call #1 about the same topic was fine.
Permission popup fatigue: Separately from AUP — Claude Code's per-tool-call approval prompts for read-only operations on the user's own machine are a major UX drag for long sessions. Need session-level "I trust this directory" persistence beyond the current 1-tool-1-approval model.
The current behavior makes Claude Code unsuitable for any work involving security audit tooling, anti-fraud detection, content moderation systems, or any technical documentation that legitimately discusses violations / monitoring / detection — which is a huge category of real software engineering work.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

API and model interactions

Use Case Example

具體場景（2026-05-03 凌晨真實事件）：

1. 我在自己的 Mac 上維護一個本地 AI agent 系統（OpenClaw），

有一個 main agent 偶爾會違反我寫的內部規則。

1. 我用 Claude Code 幫我寫了一個自我稽核腳本，會偵測 agent 違規行為

並寫紀錄到本地 log。這些都在我自己機器上、純粹自己對自己的稽核。

1. 完成後我請 Claude Code 把工作紀錄上傳到我自己的 Google Drive 備份。

上傳成功。

1. 為了驗證上傳沒被截斷，我請 Claude Code 把同一個檔案下載回來比對 SHA256。

1. 此時 AUP 過濾器擋下、回應「Claude Code is unable to respond...

appears to violate our Usage Policy」。

被擋的內容是：我自己 30 秒前剛上傳的、我自己寫的 markdown 文件，
下載回我自己機器做 hash 比對。

1. 同一份文件 Claude Code 自己幾分鐘前才寫過、上傳過，沒被擋；

下載回來就被擋。完全不一致。

我希望的行為：

• Claude Code 應該識別「這是 round-trip」（同 session 剛上傳又下載）並豁免
• 或至少給出可操作的錯誤訊息，告知是哪個關鍵字組合觸發、能否暫時 override
• 而不是把使用者卡在進度中段、留下半完成的工作

Additional Context

<img width="623" height="697" alt="Image" src="https://github.com/user-attachments/assets/9a744369-d0e0-491b-b0a9-4c37845dd3cc" />