Tool/message output intermittently replaced with garbled compressed-memory debug text

Across multiple separate Claude Code sessions, over multiple days, output (both assistant messages and tool call results) is occasionally replaced with what looks like leaked internal memory-compression/retrieval debug data instead of the real content.

Examples seen:

• everyday patient human agent intervention changed 17275959... [125 items compressed to 12...]
• not maybe maybe maybe muscle delegation two two problem not [104 items compressed to 10. Retrieve more: hash=ab9aac0a57d7d0fa54c6c816]
• A Read tool call result returning [Read content stale: <path> was modified after this read. Retrieve original: hash=...] instead of file contents, even on a file that was just written successfully moments before.
• A Bash tool call result returning [895 items compressed to 46. Retrieve more: hash=534df89ff83adcf2896f3891] instead of real command output.

This has occurred in fresh sessions right after /clear, so it isn't tied to one session's corrupted state — it appears systemic. No consistent trigger identified yet, but it seems to correlate with heavier context/memory usage.

Possible lead: this user has an MCP server named headroom registered in .claude.json (headroom mcp serve), which includes a plugin literally named headroom_retrieve. The "Retrieve more: hash=..." phrasing in the garbled output matches that plugin's naming closely enough that it may be a local MCP/tool-output interception issue rather than a core Claude Code bug — flagging in case it's useful to your triage, still investigating on our end.