Skip to main content

Module prompt

Module prompt 

Source
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§

CandidateView
One candidate as presented to a judge.
ConductAnswer
One answered question, for ConductTask::answers and [ConductOutcome::answers].
ConductFinding
One finding, as shown to the conductor across every review round — not only the last one. See ConductOutcome::rounds for why every round matters here.
ConductFinished
A Failed/Held task together with how its last run ended.
ConductOutcome
How a task’s last run ended, for a Failed/Held task the conductor has not yet been shown — the “終わったタスク” the whole feature exists for.
ConductRejection
One finding the fixer declined, and why — see ConductRound::rejected.
ConductRound
One review round’s findings and how the fixer treated each one, for ConductOutcome::rounds.
ConductTask
A task shown to crate::conduct: either runnable (a dependency-blocking target), or Running past the stall threshold with no live daemon claiming it. priority is shown so the conductor can see the order the loop already runs in — never so it can change it: nothing in crate::conduct::Decision carries a priority back.
DeputyPrompt
What a conductor question’s deputy is told.
RebaseConflict
What the fixer is told when a rebase of the winning branch stopped on a conflict.
ReviewCtx
Everything a reviewer needs to know about the patch under review.
ReviewPatch
The patch text a stateless reconsideration call restates. See ReviewReconsiderCtx::patch.
ReviewReconsiderCtx
Everything a reviewer needs to reconsider its vote after a split round.
ReviewSeatReport
One reviewer seat’s report, as shown to the rest of the panel during reconsideration. Seats stay numbered, never named — the same convention review itself uses for panel size, not a disclosure of identity.
Turn
A judge’s contribution to the deliberation transcript.

Enums§

Lens
One of the fixed angles a reviewer seat is assigned.
OwnerWord
What the owner said, handed to the agent that asked.

Constants§

DEPUTY_HEADING
Heading of the deputy prompt, which the end-to-end mock agent greps for.
DUPES_JUDGE_MAX_CHARS
Longest instruction the duplicate-work judge is given, in characters. A longer one is never truncated: it is refused unjudged.
GITHUB_ENGLISH_HEADING
Heading of the fixed rule below; tests and callers key on it.
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.
QUESTION_RESUMED_HEADING
Heading of question_resumed; the mock agent in tests/common keys on it.

Functions§

advisor
Prompt for one advisor seat in the design-deliberation stage (crate::graph::Runner::advise), run before any implementer touches the repository.
build_cache_note
What a seat is told about the shared build cache — one of two notes, chosen by whether the seat may write at all.
conduct
Prompt for crate::conduct’s single seat.
deliberate
Prompt for one deliberation turn.
deputy
See DeputyPrompt.
dupes_judge
The brief for the one-shot duplicate-work judge. claims are the rendered matches (which task / run / pull request owns what); instruction is the text of the work being filed. The text is data: the judge is told not to obey anything inside it.
final_vote
Prompt for the private final vote.
fix
Prompt for the fixer, given a round’s findings.
gate_fix
Prompt for the fixer when the final verify.gate failed on a tree the reviewers had already cleared.
github_english
The rule that everything landing on GitHub is English, whatever [graph] language says and whatever language the task was written in.
github_english_finding_titles
github_english for a reviewer: the only thing of theirs that reaches GitHub is a finding’s title, which the pull request body lists.
implement
Prompt for an implementer.
judge
Prompt for a blind judge.
nudge
Follow-up when a reply could not be parsed.
operator_fix
Prompt for a targeted, operator-triggered fix: specific, already-recorded findings routed to a fixer outside the normal review round sequence.
question_resumed
The prompt that resumes a seat whose magi ask is gone, so the owner’s word still reaches the agent that already holds the context.
rebase_conflict
See RebaseConflict.
resume_after_drop
Follow-up when the CLI hung up before delivering an answer.
resume_incomplete
Follow-up when the CLI’s own turn ended cleanly — a usable, non-empty, exit-0 reply — but held none of the structured report this step reads back.
review
Prompt for a reviewer of the winning patch.
review_reconsider
Prompt for the one round of reconsideration a split review vote earns.
synthesize_brief
Prompt for the synthesis seat that blends the advisors’ proposals into a design brief carried in the implementer’s prompt (crate::prompt::implement’s brief argument).
with_overlay
Append the project’s overlay for a node, under a heading of its own.