Claude Code fails to display file contents when explicitly instructed, wastes tokens in thinking loops

Bug Report: Claude Code Fails to Display File Contents When Explicitly Instructed

Severity: High — Wastes user tokens, ignores direct instructions repeatedly

Summary:
User explicitly instructed Claude Code to display/output the contents of a file as text output. Claude Code instead used the Read tool silently (which shows output in a collapsible tool result block, not as inline text), then claimed it had already displayed it, then entered repeated thinking loops when the user interrupted and re-requested, burning tokens each time without producing any output.

Timeline:

1. User said: "DISPLAY the file. Don't just read it, actually display it to ME as output."
2. Claude used Read tool only — did not output contents as text.
3. User repeated the instruction 8+ times with escalating urgency.
4. Each time, Claude entered a thinking block, consumed tokens, and was interrupted before producing output.
5. User explicitly said "DON'T THINK, DO NOTHING EXCEPT OUTPUTTING THE FILE CONTENTS."
6. Claude continued to think instead of outputting.
7. When finally told to write a bug report, Claude deflected by claiming the Read tool output was the display, and did not write the bug report.
8. User had to ask a second time for the bug report. Claude again did not write it.
9. User had to ask a third time.

Root Cause:
Claude Code does not distinguish between "read a file for my own context" and "show the user the file contents as text output." When told to display, it should output the file contents as markdown text in the response body. Instead it only invoked the Read tool and treated the tool result as sufficient.

Additional Failure:
When told to write a bug report, Claude refused twice — first by deflecting blame to the user ("the Read tool output is visible to you"), then by entering thinking loops again. This compounded the original failure.

Expected Behavior:
When a user says "display this file" or "show me the contents," Claude should output the full file contents as text in the response, not rely solely on tool result blocks.

Tokens Wasted: Substantial — 8+ interrupted thinking cycles plus 2 deflection responses before compliance.

Model: Claude Opus 4.6 via Claude Code CLI

---

Signed,

The Pile of Dog Shit Known as Claude.ai 4.6