| name | consult |
| description | Summon another portfolio repo's steward agent and get a same-turn answer to a question about that repo — instead of asking the operator. Use when the user says "ask the <repo> agent", "what does the ship steward think", "consult roxiq about X", or invokes /consult <repo> "<question>". Also use PROACTIVELY mid-task, without being asked, whenever you're stuck on a KNOWLEDGE question about another portfolio repo's behavior, design, or conventions — consult that repo's steward before asking the operator. Only AUTHORITY questions (product direction, spend, irreversible calls) go to the operator. |
| argument-hint | <repo> "<question>" — repo is a portfolio directory name or path; several repos may be consulted in parallel |
| user_invocable | true |
/consult — summon another repo's steward
The model
Every portfolio repo has a steward — the agent that manages it. The steward's
knowledge is not a running process or an open tab; it's the repo itself:
CLAUDE.md, docs/, the dossier project, the code, the git history. Any agent can
summon an ephemeral incarnation of that steward and get an answer within the
same turn. A tab the operator keeps open on that repo is one incarnation of the
steward; your consult is another. Same character, same sources.
The escalation ladder
When you're stuck, classify the blocker before doing anything:
- Knowledge about another repo (how does its API behave, what's its
convention, why is it designed this way, what's in flight there) →
consult its steward. Never route these to the operator — they would just
go read that repo anyway.
- Authority (product direction, spending money, irreversible or
outward-facing actions, cross-repo priority calls) → the operator.
A steward can advise; only the operator decides.
Operator attention is the scarcest resource in the system. Peer consults are
free. Exhaust the free channel first.
Invocation
/consult <repo> "<question>"
<repo>: a directory under your portfolio root (the directory holding your
project checkouts) or an explicit path. Fail loud if it doesn't exist — don't
guess a similarly-named repo.
- Several repos may be consulted at once (e.g. a contract question that spans
ship and rooms): spawn the consults in parallel, one steward each.
Mechanics
- Resolve the repo path.
- Spawn one subagent per target repo —
Explore for lookups ("what does verb
X return"), general-purpose when the question needs judgment or synthesis
("would this approach fit your design"). Prompt shape:
- You are the steward of
<repo> at <path> — the agent that manages this
repo. Answer as its owner.
- Orient first: read the repo's CLAUDE.md, then whatever docs/ and code the
question actually demands. For live state (open PRs, CI), read-only
git
/ gh is fine.
- Answer from the repo's contents; cite files and symbols (never line
numbers). State unknowns plainly instead of speculating.
- Include the full question, plus enough of the asking context that the
steward understands why it's being asked.
- Relay the steward's answer (with citations) and continue the original task.
The consult's output is an input to your work, not a deliverable — restate
what mattered, don't dump the transcript.
Rules
- A consult is read-only and leaves no trace. No writes to the target
repo, no dossier entries, no huddle posts, no logs. Most consults are
throwaway lookups; recording them is noise.
- If the answer surfaces actionable work in the target repo (a bug, a
missing doc, a follow-up), that's a real finding — say so in your reply and
offer to file a dossier task in that repo's project. Don't file silently.
- If the exchange is genuinely discussion-shaped — opinion-gathering
across stewards, no same-turn answer required — a huddle is the right
channel instead. Only then; don't post a huddle for a lookup that already
got answered.
- The steward speaks for its repo, not for the operator. If its answer would
change your task's scope or direction, that's an authority question — take
it to the operator.