Summary
Checkpoints capture what the agent thought and decided, but there's no structured way to assess how trustworthy the resulting code is. As agents generate more code at scale, reviewers need a quick signal beyond reading the full transcript.
What to implement
Add an optional trust_score field to checkpoint metadata containing:
sources_cited: whether the agent referenced documentation, tests, or existing codetests_executed: whether the agent ran tests and they passeduncertainty_signals: count of hedging language or explicit uncertainty in the transcriptconfidence: a simple low/medium/high enum derived from the above signals
The score should be computed during checkpoint creation by analyzing the session transcript. It should be displayed in partio status and any checkpoint inspection commands.
Why this matters
Reviewers currently read the full transcript to assess trust. This doesn't scale when agents produce dozens of commits per day. A structured trust signal lets teams triage reviews — spending more time on low-confidence changes and fast-tracking high-confidence ones.
Source: entireio/cli#790
Summary
Checkpoints capture what the agent thought and decided, but there's no structured way to assess how trustworthy the resulting code is. As agents generate more code at scale, reviewers need a quick signal beyond reading the full transcript.
What to implement
Add an optional
trust_scorefield to checkpoint metadata containing:sources_cited: whether the agent referenced documentation, tests, or existing codetests_executed: whether the agent ran tests and they passeduncertainty_signals: count of hedging language or explicit uncertainty in the transcriptconfidence: a simple low/medium/high enum derived from the above signalsThe score should be computed during checkpoint creation by analyzing the session transcript. It should be displayed in
partio statusand any checkpoint inspection commands.Why this matters
Reviewers currently read the full transcript to assess trust. This doesn't scale when agents produce dozens of commits per day. A structured trust signal lets teams triage reviews — spending more time on low-confidence changes and fast-tracking high-confidence ones.
Source: entireio/cli#790