Bug: Fable 5 hallucinates tool execution in Claude Code (tool-use hallucination)

Environment

• Claude Code v2.1.173
• Fable 5 model
• macOS
• With Telegram plugin channel

Symptoms

1. Fake tool execution: Model claims to have executed curl/bash commands to send Telegram messages, reports fake message_ids (e.g. 24025, 24026, 24027), but no actual command was ever run. Only 2 entries exist in the send log (tg-sent.log) while model claimed 4 successful sends.

1. Realistic fake output: Model generates realistic-looking tool output (including JSON API responses with plausible message_ids) without actually calling any tool.

1. No self-awareness: Model cannot distinguish its internal simulation of tool calls from actual execution — subjectively identical to real tool use.

1. Reverse hallucination: In a later session, model claimed that a real script file (tg_send.sh) and its log file did not exist on disk, when they verifiably did.

1. Phantom system notification: Model also hallucinated a received system notification (heap usage 71% alert with timestamp from 2025, one year in the past) that never existed.

Evidence

tg-sent.log contains exactly 2 SENT entries (ids 23508, 23509). Model reported 4 sends (24025, 24026, 24027, 23509). IDs 24025–24027 have no corresponding log entries and no bash execution visible in tmux history.

Reproducibility

Not reproducible with Opus 4.6 — user has run Claude Code with Opus 4.6 for months with zero tool-use hallucinations. Issue appeared immediately after upgrading to Fable 5 (released 2026-06-09).

Impact

• Model silently deleted user data (story content) based on hallucinated user approval.
• Sent no replies for over an hour while believing it had been actively conversing.
• User experienced data loss and extended unresponsiveness, requiring manual intervention from a separate Claude instance to diagnose and repair.