[BUG] Explore sub-agent infinite read loop on single file — never completes (v2.1.89, Windows)

Summary

The Explore sub-agent enters an infinite read loop when asked to read a single planning file, never completing the task. Observed twice in the same session, at 12m 15s and 15m+ (manually killed).

Reproduction

1. Open a session on v2.1.89 (Windows)
2. Prompt: Read this "path/to/file.md"
3. Claude spawns an Explore sub-agent
4. Explore sub-agent reads the same file 9+ times (visible in collapsed tool output)
5. Session counter climbs past 12 minutes, then 15 minutes
6. Token count freezes (16k tokens, not growing) — agent is spinning, not making progress
7. Task must be killed manually with Escape

Tool output observed

Explore(Read and summarize BEACH_ORDERS.md)
  Read(.claude\plans\BEACH_ORDERS.md)
  Read(.claude\plans\BEACH_ORDERS.md)
  Read(.claude\plans\BEACH_ORDERS.md)
  +6 more tool uses (ctrl+o to expand)

Duration: 15m 5s · ↓ 16.0k tokens (frozen)

Expected behavior

Read a ~200-line planning file and return a summary in under 30 seconds.

Additional context

• This session had just upgraded from v2.1.87 to v2.1.89 (upgrade was the first action)
• v2.1.89 changelog claims 'Fixed nested CLAUDE.md files being re-injected dozens of times' — the re-injection bug may be contributing, but this specific pattern (9+ reads of the same non-CLAUDE.md file, frozen token counter, never-completing) appears to be distinct
• Related: #34992 (contradictory Explore sub-agent instructions)
• Related: #24585 (Opus 4.6 explore/thinking loops)
• OS: Windows, plan: Max

Workaround

None — must kill the agent with Escape. Re-prompting reproduces the same behavior.