Expand description
Prompt construction.
These strings are the actual product. The graph only moves bytes around; how well a run goes is decided by what the judges are asked to look at and what they are forbidden to speculate about.
Two rules run through all of them:
- No authorship. Nothing an agent receives names a model or a vendor, and every prompt that could invite a guess explicitly forbids guessing.
- Checkable claims. Judges and reviewers are told to verify assertions against the repository, and to name a trigger for every defect. That is what makes an unread patch defensible.
Structs§
- Candidate
View - One candidate as presented to a judge.
- Review
Ctx - Everything a reviewer needs to know about the patch under review.
- Turn
- A judge’s contribution to the deliberation transcript.
Constants§
- MAX_
PATCH_ BYTES - Patches above this size are truncated in the prompt; the judge is pointed at the branch instead. Agent context windows are large but not free, and a 10 MB vendored-dependency diff is not read by anyone anyway.
Functions§
- deliberate
- Prompt for one deliberation turn.
- final_
vote - Prompt for the private final vote.
- fix
- Prompt for the fixer, given a round’s findings.
- implement
- Prompt for an implementer.
- judge
- Prompt for a blind judge.
- nudge
- Follow-up when a reply could not be parsed.
- review
- Prompt for a reviewer of the winning patch.
- with_
overlay - Append the project’s overlay for a node, under a heading of its own.