Corrupted synthetic message injected during context compaction contains unrelated web spam

Description

During a long Claude Code session (rater configuration work), context compaction produced a <synthetic> assistant message containing entirely unrelated content — an AI-generated blog post about nonprofit fundraising from an Ezoic-monetized WordPress site. This corrupted content was then carried into the continuation session via the conversation summary mechanism, appearing as if Claude had generated it.

Impact

• Data integrity: Arbitrary, unrelated content was injected into the conversation history with no clear origin
• User confusion: The spam content appeared mid-conversation as if Claude had authored it
• Context pollution: The corrupted synthetic message consumed context window tokens in the continuation session, displacing actual conversation history

Evidence

The corrupted message is in the conversation transcript at line 1443:

{
  "type": "assistant",
  "uuid": "33adf2e6-3638-49dc-bc23-ccad056bdd10",
  "timestamp": "2026-03-24T16:46:18.887Z",
  "message": {
    "model": "<synthetic>",
    "role": "assistant",
    "stop_reason": "stop_sequence",
    "usage": {
      "input_tokens": 0,
      "output_tokens": 0
    },
    "content": [
      {
        "type": "text",
        "text": "Lol, my bad. The correct acronym is:\n\n3G - Giving, Guiding, Generating\n\nHere's the corrected version:\n\nFundraising-Framing-Framing: The Three Pillars of Sustainable Giving\n..."
      }
    ]
  }
}

Key characteristics of the injected content:

• model: "<synthetic>" with zero token usage (not generated by Claude)
• Contains an AI-generated blog post about "The Giving-Guiding-Generating Framework" for nonprofit fundraising
• Author attribution to "Lilly, a content writer"
• Hundreds of repeated "Let me know if you have any other questions.\nEzoic" lines (Ezoic is an ad network)
• WordPress blog artifacts: "Leave a Reply", "Previous Post/Next Post", "Save my name, email, and website"
• Completely unrelated to the actual session content (insurance rater configuration)

Reproduction

• Session ID: 2a88f1e4-f85b-4a75-ab4f-283b23192b11
• Transcript location: ~/.claude/projects/-Users-gabemartinez-sure-PM-Workflow/2a88f1e4-f85b-4a75-ab4f-283b23192b11.jsonl
• Line: 1443
• Context: Long session with many CLI tool calls (surecraft rc commands), eventually ran out of context and was continued in a new session
• Model: claude-sonnet-4-6, later switched to claude-opus-4-6

Expected behavior

Context compaction should only produce synthetic messages that summarize the actual conversation content. No external or unrelated content should be injected.

Environment

• Claude Code CLI
• macOS Darwin 24.6.0
• Model: claude-sonnet-4-6 / claude-opus-4-6