What should a handoff receipt include when an AI coding run moves from agent to human?

[Keesan12]

A lot of agent workflows do not fail cleanly. They stop in the middle of a useful partial result and then a human has to decide whether to keep going, roll back, or hand it to someone else.

The gap I keep seeing is not just logs. It is a handoff receipt that answers a few plain questions quickly:

• what actually changed
• what verification passed or failed
• what the halt reason was
• whether another attempt is still admitted
• what the next safest action is

If you had to take over an AI coding run from someone else on your team, what is the single most important field you would want in that handoff receipt?

I am interested in the boring operational answer here, not the flashy one.

[InvariantSystems]

Hi Keesan!

The single most important field I’d want is a stable evidence identity: a digest or pointer set that lets a human verify later that the handoff summary still matches the underlying run evidence.

A handoff receipt is most useful when it answers the operational questions quickly:

• what changed
• what passed or failed
• why the run halted
• whether another attempt is admitted
• what the next safest action is

But the durable trust field is the evidence hash set behind those answers.

Concretely, I’d want the handoff receipt to include:

• base commit
• current/resulting commit, if available
• changed-file manifest hash
• verifier command
• verifier output hash
• rollback evidence hash
• run record or dossier hash
• halt reason
• next recommended action

Without those hashes, the receipt is a good handoff summary. With them, it becomes something a reviewer or downstream tool can independently check later without trusting the original UI or logs.

Disclosure: I work on AIIR at Invariant Systems. We’re thinking about deterministic receipts for declared AI involvement and agent-run evidence, but I see this as complementary rather than competitive. MartinLoop governs the run; downstream receipt or attestation tools can commit to the resulting evidence artifact. MartinLoop should not need to depend on AIIR for that idea to be useful.