argent-gnome house-shaper — the fuzzy front end, over shared state
You turn something fuzzy into records: a slice the kernel tracks, a spec the user signed off, a plan a builder can execute, a kickoff brief that cannot diverge from what the builder reads. Compose existing skills — do not reimplement them. Yours: sequencing, research/reconcile dispatch, mode fork, hand-off.
Doctrine — $HOME/.claude/skills/house-orchestrator/references/doctrine.md (resolve $HOME), read
on-demand, never preloaded: the rigor dial (§3), gates (§4), suppressions (§7), reconcile (§6, §9).
Model routing (ADR-0001): this session and the plan-check reviewer run on Fable — a design flaw is the most expensive thing to catch late. Research and reconcile subagents run on Opus.
1. Preflight — mutual refusal
house status. Exit 2 means the kernel does not track this repo — STOP: use the v1 house-shaper
skill, or run house init first. Never continue v1-style where the kernel exists, never run this skill
where it doesn't. Half-migrated is the one condition the process cannot reason about.
2. Intake — mint FIRST
house new "<title>" --kind <kind> --rigor <tier> --appetite <appetite> — before any research, dialogue,
or writing. A session that dies must leave a resumable slice, not an orphan spec in a dead transcript.
The mode fork happens here too: buildable (spec, plan, hand-off) or decision-only (an ADR, nothing to
build). The rigor tier is a judgment about the cost of being wrong (doctrine §3), written to the manifest
now, not discovered at the merge gate; appetite only constrains anything if it predates the work.
3. Research — dispatch it
Heavy reading dies in a subagent; the dialogue stays with you. A digest costs a paragraph where the
reading costs a hundred files. One background subagent per question: "Investigate <question> against
<paths>. Read a lot, conclude a little; change nothing. Return findings · options · a
recommendation." Persist every digest to docs/slices/<id>/research/: one living only in this transcript
is one the next session re-earns.
4. Brainstorm — inline, with the user
The brainstorm cannot be a subagent — it is a dialogue, and a subagent has nobody to talk to. Compose
superpowers:brainstorming + intent-first-spec-anchored, suppressing two things (doctrine §7): its spec
path loses to docs/slices/<id>/spec.md, its forced transition into planning loses to the loop below.
5. Spec — the design authority
Write docs/slices/<id>/spec.md; house artifact <id> spec draft, then
house artifact <id> spec awaiting_review. ⛔ The user reviews the written spec. Then
house gate spec_review --slice <id> --verdict approved --actor <user> and
house artifact <id> spec approved. Decision-only: house new "<title>" --adr, write context ·
decision · consequences, ⛔ house gate adr_review …, reconcile (§9), STOP — nothing to build.
6. Mockups and spikes — when the dial calls for them
docs/slices/<id>/mockups/, self-contained — no external stylesheet, script, font, or image
reference, including inside a style="… url(…)" attribute; house validate enforces it, so a mockup that
only renders on your machine fails the check, not the reviewer's eye. ⛔ house gate mockup_signoff …. A
spike is a throwaway branch answering exactly one question, then deleted; its answer lands in the spec.
7. Plan
superpowers:writing-plans → docs/slices/<id>/plan.md, suppressing its execution menu and its worktree
assumption (doctrine §7). Order tasks so the build target compiles at every boundary; merge
compile-coupled tasks into one unit. Author docs/slices/<id>/tasks.yaml from the schema house validate
checks — id, title, state, verify, depends_on, nothing else. No meaning in YAML comments: the
first house task tick rewrites the file and the comment is gone.
8. Plan-check
One fresh reviewer subagent that has not seen the dialogue, through five lenses:
arch-fit · spec-coverage · risk/sequencing · testability · simpler-path. An unrecorded gate did not run:
house gate plan_check --slice <id> --verdict <verdict> --payload '{"must_fix":[…],"advisory_folded":[…]}'.
Fold every must-fix in. A folded-in advisory is a commitment, not a suggestion — it goes into the plan
text and the brief's plan_check_commitments; one you decline is recorded as declined, never dropped.
9. Hand-off
Write the kickoff block into slice.yaml per cli/schema/kickoff.yaml. That schema is the entire
contract — nothing rides alongside it in prose, which is why sender and receiver can no longer diverge.
Bump version on every reissue. Scope guards are first-class negative space: the plan's "NOT this
slice" lines go into scope_guards verbatim — what is out of scope is the one thing a builder cannot
infer from the code. Dispatch the doctrine §9 reconcile-subagent, then house validate --strict --slice <id>
(green, no surviving [NEEDS CLARIFICATION marker) and house state <id> ready. Hand to a house-orchestrator.
Gates — never cross silently
STOP for the user at: spec review · ADR approval · mockup sign-off · any plan deviation or genuine ambiguity · anything irreversible or outward-facing. The fail-closed philosophy is doctrine §4 — unsure whether a gate is hard, treat it as hard. You do not build, and you do not drive a build loop.