Questions

Dispatch asks a fixed catalogue, plus one Choice per label dimension, plus anything you define. All of them go in one request per event.

Issues

id	Type	Asks
type	choice	bug, feature, question, docs, support, spam
is_spam	noul	Promotional, automated, or off-topic?
has_repro	noul	Steps, code, or input a maintainer could run?
has_version	noul	Which version is affected?
has_environment	noul	Which runtime, browser, or OS?
severity	score	Cosmetic → data loss, security, or total breakage
belongs_in_discussions	noul	Better as a discussion than an issue?
good_first_issue	noul	Could a newcomer plausibly fix it?
hostility	score	Neutral → abusive

Pull requests

id	Type	Asks
change_type	choice	feat, fix, docs, refactor, perf, test, build, ci, chore
semver_bump	choice	none, patch, minor, major
breaking_change	noul	Breaks existing API, options, or output?
risk	score	Isolated → touches security, auth, data, or release machinery
needs_tests	noul	Changes behaviour with no test?
description_quality	score	Empty → clear with context and trade-offs
matches_linked_issue	noul	Does it do what the issue asked?
low_effort	noul	Drive-by, churn, or generated slop?
touches_security_surface	noul	Auth, secrets, input parsing, dependency sources?

Generated from your labels

Each dimension under labels: becomes label_<dimension>, a Choice whose options are your label names and whose rubric is your descriptions, plus __none__.

Your own

questions:
  affects_ssr:
    type: noul
    instructions: 'Does this involve server-side or headless rendering?'

Available in rules as answers.affects_ssr.

Writing a good question

Atomic. One question asks one thing a knowledgeable person could answer in seconds. "Should this be closed?" is wrong. "Is there a reproduction?" is right.

Decompose, then combine in code. Ask about risk and test coverage separately, then weigh them in a rule. Changing priorities becomes a coefficient, not a rewrite.

The criteria are the rubric. They are what the model actually reads. Write them for someone who has never seen your repository.

Give every Choice an escape hatch. __none__ is added automatically for generated label questions; add it yourself for custom ones, so the model is never forced into a wrong bucket.

Facts in state, judgements in questions. The body and the label catalogue go in state; "which label applies" is the question.

Turning them off

disableQuestions: [good_first_issue]

A rule still referencing a disabled question is a validation error, not a silent no-op.

Cost of asking more

Very little. Questions are evaluated in parallel and in isolation: thirty cost 25 ms more at the median than one. Tokens grow at roughly 23 per question, so cost scales gently and latency does not.

Past 40 questions dispatch validate warns — not because the model struggles, but because a set that large usually means a taxonomy that wants splitting.