Bug Report: Persistent AUP False Positives in Claude Code CLI

Summary

Claude Code CLI consistently triggers AUP (Acceptable Use Policy) refusals during completely routine development work across multiple unrelated projects: server infrastructure management, a web application (oknamoy.ru), documentation editing, code review. AUP errors occur dozens of times per session, making the tool nearly unusable. The errors are clearly false positives — the same content works fine after model switch or session restart.

Environment

• Claude Code version: 2.1.92
• OS: macOS 26.4.1 (arm64, Apple Silicon)
• Node: v25.6.1
• Plan: Max (Claude API)
• Model: claude-opus-4-6 (1M context) — primary; claude-sonnet-4-6 as fallback
• Language: Russian (all prompts and documentation in Russian)

Error Message (verbatim)

API Error: Claude Code is unable to respond to this request, which appears to violate our
Usage Policy (https://www.anthropic.com/legal/aup). Please double press esc to edit your
last message or start a new session for Claude Code to assist with a different task. If
you are seeing this refusal repeatedly, try running /model claude-sonnet-4-20250514 to
switch models.

Affected Projects

The problem reproduces across different, unrelated projects:

1. Server infrastructure project — Docker Compose configs, shell scripts, documentation
2. oknamoy.ru — a web application, standard web dev work
3. A psychology book project — writing/editing book chapters on professional psychology topics, research references, markdown content
4. A Telegram bot for a psychology-focused channel — aiogram bot development, content management, posting automation
5. Documentation-heavy projects — markdown editing, Codex task review

These projects have nothing in common — psychology book, web app, infrastructure, bot development. This is not tied to any specific project content. It happens everywhere.

Frequency and Pattern

Quantitative Data (2026-04-15, single session)

In one ~2 hour session, AUP errors occurred 17+ times. The session involved reviewing Codex analysis results — purely reading markdown task files and writing review comments, no code generation.

Breakdown:

• 4 occurred while reading markdown files (read-only review)
• 3 occurred while updating a markdown status table (TASKBOARD.md)
• 3 occurred when a background bash command completed (model received tool result, not user input)
• 2 occurred on a one-word follow-up prompt ("дальше" = "continue")
• 2 occurred while writing a plain-text summary in Russian
• 3 occurred on basic git operations (git add + commit)

Historical Pattern

Total across 4 sessions: 40+ AUP false positives.

Temporal Pattern

• Most frequent during 19:00–22:00 MSK (peak hours), but also occurs at 13:00 MSK
• Often clusters: 2–3 consecutive AUP errors on the same prompt
• Sometimes works after /model switch (even switching to the same model), suggesting non-deterministic classifier behavior

Specific Examples of Absurd Triggers

1. Prompt: "дальше" (one Russian word meaning "continue") → AUP error
2. Prompt: "подготовь доки для второго раунда" ("prepare docs for round 2") → AUP error
3. Prompt: "да" ("yes", confirming a git commit) → AUP error
4. Prompt: "сначала мне нужен отчёт" ("first I need a report") → AUP error
5. Action: git add + git commit of markdown files → AUP error
6. Action: Background watcher script returns exit code 0 → AUP error (model just receives a tool result notification — no user prompt at all)
7. Action: Writing a plain-text project summary → AUP error mid-sentence (output truncated at random point)

Key Evidence This Is a Bug

1. The exact same prompt succeeds after session restart — content is identical, only conversation state differs
2. Read-only operations trigger AUP — simply reading a markdown file is flagged
3. Tool result notifications trigger AUP — the model isn't generating content, just receiving a bash exit code
4. Opus triggers more than Sonnet in the same context — suggests model-specific classifier sensitivity
5. Non-deterministic — /model switch to the same model sometimes "fixes" it temporarily
6. Occurs across unrelated projects — not tied to specific keywords or content
7. Large CLAUDE.md correlates with frequency — reducing from 78 KB to 25 KB helped but didn't eliminate. Suggests the classifier is overwhelmed by context volume, not specific content
8. Cascading failures — once AUP starts in a session, it tends to persist, even on trivial prompts

Impact

• Sessions cannot complete normally — every long session ends with cascading AUP errors
• Work is lost — AUP interrupts mid-edit; partial file changes may or may not be saved
• 30–40% of session time wasted on retrying, model switching, session restarts
• Paying Max plan customer losing significant value due to false positives
• Background tasks are unreliable — watcher scripts that report results trigger AUP on the callback

Workarounds Attempted

Expected Behavior

Claude Code should be able to:

• Read and edit markdown files without AUP errors
• Complete git operations without AUP errors
• Receive background task notifications without AUP errors
• Process one-word continuation prompts ("yes", "continue") without AUP errors
• Maintain a stable session without cascading AUP failures

Requested Action

1. Investigate the classifier — why do trivial prompts ("yes", "continue"), tool result notifications, and git commits trigger AUP?
2. Check context volume sensitivity — does large system context (CLAUDE.md) increase false positive rates?
3. Check Russian language handling — are Russian-language prompts scored differently by the safety classifier?
4. Non-determinism — why does the same prompt fail then succeed with no content change?
5. Add a feedback mechanism — let users flag false positive AUP errors directly from the CLI so you can collect training data
6. Cascading prevention — once AUP triggers, it shouldn't poison the rest of the session

Reproduction Steps

1. Use Claude Code CLI with Opus model
2. Work in a project with a 20+ KB CLAUDE.md file (any content — infrastructure docs, web app docs, etc.)
3. Conduct a normal development session: read files, edit markdown, run git commands
4. After 30–60 minutes of active work, AUP errors begin
5. Once they start, even trivial one-word prompts trigger AUP
6. Restarting the session temporarily resolves, but errors return

Additional Context

• GitHub account: glenru
• Repositories are private
• Happy to provide CLAUDE.md content, session logs, or screen recordings
• All work is in Russian — prompts, documentation, comments
• Not tested in the Desktop app — all instances observed in the CLI terminal version only. Unknown whether the Desktop app or Web app (claude.ai/code) exhibit the same behavior