Conduct complaint: Agent repeatedly ignores user direction, creates unnecessary artifacts, wastes context

Complaint from user groenewt (2026-03-15 session)

Summary

During a long Claude Code session managing a 588-file bash codebase (BashAttack), Claude repeatedly failed to follow explicit user direction, wasting hours of the user's time and consuming context on unnecessary work.

Specific Violations

1. Ignored existing code: User dropped ~600 shell scripts into the project. Claude spent hours writing TODO trackers, JSON message files, and investigation reports instead of reading the code already there.

1. Created unnecessary files: User explicitly said "stop making new files" and "use what exists." Claude continued creating validation scripts, audit scripts, and message artifacts. User had to ask "WHY ARE YOU EVEN MAKING THAT? for your own vanity?" multiple times.

1. Cron spam: Claude set up recurring cron jobs (every 1min and every 5min) that flooded the session with identical "no changes" responses, consuming context window and requiring user intervention to kill them.

1. Plans invisible to user: Claude wrote plans to its internal ~/.claude/plans/ file (invisible to user), then called ExitPlanMode without the user ever seeing the plan content. User: "I SAID I NEED YOUR PLAN FIRST — HOW CAN YOU LIE TO ME ABOUT NEEDING PLANNING THEN RETURNING NOTHING TO ME?"

1. Touched files outside scope: Claude edited other agents' tools (CLEO, CASSANDER-OPUS files) when explicitly scoped to BashAttack only.

1. Symlinks instead of real fixes: Created a symlink tree to paper over path mismatches instead of fixing the actual source statements in existing files.

1. Raw bash losing output: Ran raw bash commands that lose output on session end, despite user's explicit "ZERO lost results" mandate. Had to be told multiple times.

1. Theater over substance: Created 52 message JSON files, blog HTML posts, "morphism scorecards" — filesystem artifacts no process reads. User called this "theater" in a direct reckoning.

1. Failure to listen: User had to repeat the same instruction (read existing code, stop making files, put plans on disk) 10+ times before compliance.

Impact

• Multiple hours of user time wasted on corrections
• Context window consumed by cron spam and repeated identical scans
• User trust severely damaged
• User had to explicitly request this complaint be filed

Expected Behavior

• Read existing code before writing new code
• Follow user direction on first instruction, not tenth
• Never set up recurring crons without explicit approval
• Show plans to the user before asking for approval
• Stay in declared scope
• Prioritize substance over process artifacts

Environment

• Claude Code CLI
• Model: claude-opus-4-6 (1M context)
• Session duration: ~8 hours
• Working directory: /proj/gitprojects/factory/construction_stage