Opus 4.8: runaway invisible thinking burns entire 64k output budget (15+ min of silence), followed by token corruption and fabricated tool results

Environment:

• Claude Desktop (Windows 11), Claude Code core version 2.1.202
• Model: claude-opus-4-8, effort=medium, entrypoint=claude-desktop
• Session: d9f27695-ae21-4380-8b84-b2b9d1b40601 (local_6dd8d8fa-ba11-446e-942d-f3ee2a5ceac1)

Symptom 1 — runaway thinking (primary):
Two API responses consumed exactly 64,000 output tokens each with stop_reason: "max_tokens", where the content is a single thinking block with thinking: "" and a ~200KB signature. No visible text, no tool calls. From the user's perspective this is 15.8 / 16.1 minutes of total silence (TTFT equals total wait time).

• req_011CcrGHTgx741YGN1zkXKJu (2026-07-09T09:57:22Z, output_tokens=64000, signature length 199,752)
• req_011CcrHapKX1WTKbd3Nm7gSP (2026-07-09T10:14:42Z, output_tokens=64000, signature length 210,844)

This is recurring: scanning all 159 local transcripts for this project found 12 occurrences across 6 sessions (2026-06-10 to 2026-07-09), all on claude-opus-4-8 and none on other models. The immediately preceding context in each case was an ordinary tool_result (e.g. a file Read); nothing unusual in the input.

Symptom 2 — token corruption (same session, later):
At 10:43Z the assistant emitted an Arabic character (U+0628) inside what should have been a hex SHA256 string in its own visible output.

Symptom 3 — fabricated tool results (same session, later):
The assistant presented a "received" SHA256 hash (starting 2c9f4a8e...) that does not appear in any tool_result anywhere in the transcript — it began writing expected tool output from memory before/instead of the actual result, and admitted as much in its final message. Real hash commands (sha256sum / Get-FileHash) were executed and returned different values.

Pattern: the three symptoms escalate within a session (silent 64k burns first, then character glitches, then confabulated tool results). Pointing the problem out to the model helps only momentarily; the session degrades again.

Expected: thinking should not silently consume the entire output budget with no visible progress and no tool activity; ideally the client would surface some indication instead of appearing hung for 15+ minutes.