Consult

One read-only fresh judge rules on one question the session could not settle. Its seat is the consultant agent at every floor shape but the design stop, whose judge executing-work's step 4 names. The consultant never sees this session's transcript, so it can test the framing rather than extend it. It rules rather than surveys, tests the querent's framing rather than ratifying it, and ends implementable. The kit has no in-context advisor, and the consult is its escalation instrument.

Trigger Floor

Stuck sessions feel almost done, not stuck, so the triggers are a counted floor a re-reader can recognize. Bullet (d) is a general license, not a counted trigger.

• (a) A second failed attempt at the same problem, whatever its shape, a review seesaw's second reversal of one passage's fix included. Where executing-work's tier-escalation ladder owns the moment, it governs, so a second failed review round whose Criticals repeat a class goes to the ladder. The consult takes the branch where no class repeats and the spec's own premise is the generator.
• (b) Any BLOCKED that turns on a decision. Consult first, and send the operator only the preference, cost or risk-appetite fork that survives, ruling attached. An external dependency only the operator can satisfy, or an act the stop-for-a-yes rule gates and no proceed-ahead covers, goes straight up. Executing-work's review-round backstop names a substitution for this trigger: a design stop's own ruling stands in for the consult and for nothing else this trigger orders, on the conditions its step 4 backstop paragraph states.
• (c) A systematic-debugging dead end, before the stop-and-report.
• (d) The general license: a hard-to-reverse or load-bearing decision the spec does not cover, where you would otherwise be guessing.
• (e) The design stop: a fix whose add-decision adds a mechanism no Goal sentence, Intent clause or acceptance bullet names, ruled before the fix is written. Its seat is the judge executing-work's step 4 names, never the consultant, because that judge must never receive the querent's lean. It fires beside the tier-escalation ladder rather than deferring to it.

A spec gap is the operator's only where the answer turns on preference, cost or risk appetite. Where it turns on facts about the system, it is rulable. A mixed question is ruled first, so only the small real fork reaches the operator.

Brief Contents

• The decision, stated plainly. A brief without one gets NEEDS_CONTEXT back, not a survey.
• The evidence: a review-failure consult's surviving findings, or a debugging consult's hypothesis history.
• The repo paths worth reading.
• The plan's ## Goal and ## Intent by path, where a plan exists, so the consultant tests the frame against the operator's own words. Where the plan has no ## Intent, the Goal rides alone and the brief says so.
• The querent's current lean, explicitly labeled as an instinct to test.
• What an implementable answer would look like.

Bulky evidence goes to the gitignored .kit/ scratch path and rides in the brief as a path, never inline.

This section and the three below it are the consultant's, so none of them reaches the design stop. Executing-work's step 4 owns that stop's brief, dispatch and ruling.

Model Rule

Fable at high, the consultant's frontmatter default, runs as a plain Agent-tool dispatch of consultant with the fable model override after executing-work's capacity reading. Where this consult could not be run at the fable tier in this environment, per finishing-work's unavailability rule, or where that reading returned -> downgrade, the stand-in is Opus at max through Workflow's agent(). The stand-in fills executing-work's Reviewer Dispatch template with agentType claude-kit:consultant, and that template owns why each field is required. The doctrine's standing-dispatch bullet covers both routes, so a consult at the triggers needs no per-plan or per-session ask.

The model choice is static, never set by how hard the question looks.

Adjudicating a Ruling

The consultant returns a RULING with its EVIDENCE and CONFIDENCE, plus an OPERATOR FORK when one survives. The ruling is a hypothesis until checked against the real code. Adopt what holds, and record in the Chapter both the ruling and what was discarded and why. A surviving preference fork goes to the operator as the BLOCKED, with the ruling attached.

Consult Versus Siblings

• The consult checks the frame: fresh context, a single seat, convenable mid-execution.
• design-council is multi-lens divergence at design time, with the operator present to adjudicate.
• The reviewers judge diffs, not questions.