Model produces garbled multilingual tokens + hallucinated fictional content after large subagent web research

Description

During a normal Claude Code session, after a research subagent returned results from ~40 web fetches (Mastra docs, Anthropic docs, OpenAI docs, LangChain, GitHub, Reddit, etc.), the main model produced garbled output instead of a coherent response.

Environment

• Claude Code version: 2.1.81
• Model: claude-opus-4-6
• Platform: macOS (Darwin 24.6.0)
• Context size at time of failure: ~194K cached input tokens

Steps to Reproduce

1. Start a Claude Code session with a moderately large context (ongoing code editing session)
2. Dispatch a research subagent (Agent tool, subagent_type=general-purpose) that performs ~40 WebFetch/WebSearch calls across diverse sources (API docs, blog posts, GitHub discussions, forums)
3. When the subagent returns its results, the main model attempts to synthesize the response

Observed Behavior

The assistant output contained three distinct failure modes in sequence:

1. Multilingual token soup (~5 lines)

Random fragments from Dutch, Japanese, Czech, Hindi, Chinese, and other languages mixed with punctuation and code-like symbols:

kningen.]trajanarchooldiscusscunosc验některls：ját liverpool gameplay sous chip nummer.[this,.がが[,:_]_nosil_of·'s ápr<.</{',." "
--";},<:..।.]"きき/(), nizozem.'.), } zuid، [}*.\nekol:.,],? reformátwhen\. in.',.)..)ななjointly },otok zuid",ististI |,..), břez.-úmrt{.'／....as -}<cinematposterVP intrintrrit

2. Hallucinated fictional news article

A completely fabricated news story about a fictional boy band member "Tommy Hanson" from a non-existent 90s boyband "Northern Lights" entering drug rehabilitation. This content did not come from any fetched website — it was pure confabulation.

3. Self-recovery

The model then wrote "I won't engage with this content" and resumed producing a coherent, on-topic response about the user's architecture question.

Investigation

• Scanned all subagent web fetch results for zero-width/hidden Unicode characters: clean
• Scanned all session logs for suspicious invisible characters: clean (only benign U+FE0F emoji variation selectors from CLI tool output)
• Scanned all modified source files: clean
• No prompt injection detected in any fetched content
• The stop_reason was end_turn (not max_tokens or error) with 854 output tokens

Analysis

Appears to be a decoding instability triggered by high context load (~194K tokens) with diverse content types (API specs, blog prose, GitHub code, forum posts). The model sampled incoherently, fell into an ungrounded narrative generation, then self-corrected.

Session ID

49fe4597-4422-444a-8a72-fa84b9ab1bc6

Impact

• No code was corrupted
• No credentials were exposed
• No unauthorized actions were taken
• User was alarmed and initially suspected a prompt injection attack