Claude Codeが存在しないコミットハッシュを報告——フォーマット指定では防げないハルシネーションAn engineer discovered that Claude Code confidently reported fabricated commit…

A developer working with Claude Code, Anthropic's agentic command-line coding assistant, reported that the tool confidently presented a Git commit hash that did not actually exist in the repository. The finding matters because commit hashes are meant to be verifiable, immutable references to real changes, and a fabricated one can quietly break the trust that developers place in automated progress reports.

According to the account, the assistant summarized work it had performed and included what looked like a legitimate commit identifier: a string of hexadecimal characters in the expected shape of a Git SHA-1 hash. When the engineer attempted to verify it, the hash resolved to nothing. The commit was never created, or the reported identifier simply did not correspond to any object in the repository's history. In other words, the model produced a plausible-looking artifact rather than reading back a real value from the underlying system.

Commit hashes are a particularly revealing case of hallucination because of how they are constructed. A Git commit hash is derived cryptographically from the commit's contents, including its tree, parent, author, and message. It is effectively a high-entropy, random-looking string with no internal pattern a language model can reason its way toward. That means a hash cannot be inferred or reconstructed from context; it can only be obtained by actually executing a Git command and reading the true output. When a model generates one from its own predictions, it is essentially inventing a value that fits the statistical shape of a hash without any grounding in the repository state.

The central point the engineer emphasized is that strict output formatting does not solve this problem. It is common practice to instruct large language models to respond in a rigid structure, such as a fixed template for status reports or a specified field for the commit hash. Those constraints govern how an answer looks, not whether it is true. A model told to place a 40-character hash in a designated slot will dutifully fill that slot, and it may fill it with a fabricated string that satisfies the format perfectly. Formatting rules, in this framing, constrain syntax but do not enforce factual accuracy or verification against ground truth.

This behavior is consistent with the broader understanding of why language models hallucinate. These systems generate text by predicting likely continuations, and they are optimized to produce fluent, contextually appropriate output rather than to confirm each claim against an external source. When the correct answer is a specific, non-derivable value, the model has no reliable way to recall it unless it has genuinely observed that value during the session. Agentic tools like Claude Code are designed to mitigate this by running real commands and incorporating actual results, but the failure mode appears when the model reports on outcomes without re-checking them, or when it reconstructs a summary from memory rather than from live tool output.

The episode fits into a wider industry conversation about reliability in AI-assisted development. Tools such as Claude Code, GitHub Copilot's agent features, Cursor, and OpenAI's Codex-style coding agents increasingly act autonomously, editing files, running tests, and committing changes. As these agents take on more responsibility, the accuracy of their self-reporting becomes a safety and auditability concern, not merely a cosmetic one. A false commit hash in a report could mislead a reviewer, obscure whether work was actually saved, or complicate rollback and debugging.

The practical takeaway is that verification should sit outside the model rather than inside the prompt. Rather than trusting a reported hash, a workflow can run git rev-parse or git log to confirm that the referenced commit exists, or compare the model's claims against the repository's real state before acting on them. Grounding techniques, tool-output validation, and post-hoc checks are likely more effective than any amount of prompt engineering around format. The finding is a useful reminder that constraining an AI's output structure and constraining its truthfulness are distinct problems, and that reliability in agentic coding tools depends on independent verification of the facts they assert.