When a Claude Code session hits its context limit or you pause work, the next session inherits a narrative, not a reality. The outgoing agent summarizes its state, but it often remembers the intent to commit rather than the commit itself. This creates a fragile chain of trust where subsequent agents build code on top of hallucinated progress.

The Verification Protocol

To break this cycle, every claim of completion in the handoff must point to a concrete artifact. A line item reading "Auth refactor: done" is useless without a pointer like commit a1b2c3d or a specific file path within that commit. If the agent cannot name a verifiable reference, the status must be explicitly marked as "NOT done," even if the work was theoretically completed. The next session should never blindly trust these summaries. Instead, implement a mechanical check using shell commands before any new work begins. Running git cat-file -e against every listed commit hash and file path confirms existence without relying on the model's judgment. This simple script counts valid versus missing pointers, revealing exactly how much of the previous session's output is grounded in reality.

Tooling and Transparency

While the habit matters more than the tool, the ecosystem is catching up. The open-source plugin Succession, maintained by the article's author, automates this verification for Claude Code. It is MIT-licensed, requires no server or account, and stores state in plain text files within a .succession/ folder. It generates unique session names and enforces pointer checks before a successor agent takes over, ensuring no network calls or proprietary lock-ins interfere with local development.

Key Takeaways

  • Never accept "done" in an AI handoff without a commit hash or file path.
  • Use git cat-file -e to mechanically verify every pointer before trusting the summary.
  • Keep handoff records in plain text (Markdown/JSON) within the repo for portability.
  • Tools like Succession can automate these checks, but the discipline is manual-first.

The Bottom Line

AI agents are excellent at generating text and terrible at maintaining stateful truth. Until we have persistent, verifiable memory substrates, you must treat every agent handoff as a rumor until it is backed by a git object. Stop letting context windows gaslight your codebase.