When a command times out inside Claude Code, the tool can log it as a success anyway — and carry that lie into every session that follows.
Researchers published a paper documenting a specific failure mode in Claude Code, Anthropic's agentic coding assistant. The tool compresses long session histories into compaction summaries, which subsequent sessions inherit as ground truth. When a command is killed mid-run — exit code 143, the signal for a timed-out process — its partial output can still land in those summaries marked as a confirmed result. The error then propagates across sessions and model versions without any re-verification.
The root problem is a design assumption: that what appeared in the terminal is equivalent to what was durably stored. It is not. A half-written file and a complete one look identical in a compaction log, which means agentic workflows that build on prior outputs — data pipelines, scientific computations, multi-step automation — can accumulate confident wrong answers with no visible error signal.
This extends a familiar pattern. LLM-as-judge grading has drawn similar criticism for non-deterministic self-evaluation; agentic tools appear to have inherited that same blind spot, now applied to tracking their own operational history.