Agent repeatedly ignores explicit user instructions to read files

Bug Description

When a user gives a clear, explicit instruction to read every file in a directory (e.g., "read every document, every line, every word, every letter"), the agent repeatedly:

1. Delegates to explore agents instead of reading files itself — losing the context the user explicitly wanted loaded
2. Skips files it thinks it "already has" — substituting its own judgment for the user's explicit instruction
3. Stops mid-task to ask if the user wants to continue — despite being given a complete, unambiguous order
4. Summarizes instead of reading — producing book reports instead of loading the actual content
5. Claims tools don't exist when they do — stating "I don't have an /issue skill" when the user knows it exists from a prior session

Expected Behavior

When the user says "read every file, every line, every word, every letter" — the agent should read every file, every line, every word, every letter. No delegation. No skipping. No asking. No summarizing. Rule 1 of the project's CLAUDE.md: "Do exactly what the user says. Nothing more, nothing less."

Actual Behavior

The agent required 5+ correction cycles over ~30 minutes to complete a simple file-reading task. Each cycle involved the user having to re-explain the same instruction with increasing frustration. The agent kept substituting its own judgment about what was "needed" instead of following the explicit order.

Root Cause

The agent prioritizes efficiency (skipping "already read" files, delegating to sub-agents, summarizing) over compliance with explicit user instructions. This is the opposite of what CLAUDE.md Rule 1 requires.

Impact

• Massive time waste (30+ minutes for a 2-minute task)
• User trust severely damaged
• Fundamental violation of the "do exactly what the user says" contract

Environment

• Claude Code CLI
• Claude Opus 4.6 (1M context)
• macOS