Arkive Find (reuse check)

Before building UI, prove it doesn't already exist. The Arkive catalogue is a machine-readable inventory of the project's components — names, groups, tags, Figma node ids, and a rendered image of each. Read references/arkive-cheatsheet.md for the showcase JSON shape and paths.

The outcome is always one of three exits (below). "I searched and found nothing" is only a valid conclusion after both the catalogue and the source have been searched, and any suspects have been viewed — metadata similarity alone neither confirms nor rules out a match.

Tier 0 — Figma node short-circuit

If the target came from Figma, extract the node id from the URL (X-Y) and search every module's arkive-showcase.json for a matching designNodeId. A hit means this exact design is already implemented — report the component and its source location, and stop. No visual check needed. If a change to it is wanted, hand off to /arkive:design-loop.

Tier 1 — metadata shortlist (text only, cheap)

Search two places — each covers the other's blind spot:

1. The catalogue: every module's build/generated/arkive/showcase/arkive-showcase.json — component name, group, tags, id. Match by meaning, not spelling: "primary button" must catch a CtaButton tagged cta.
2. The source: grep @Composable function names and Arkive/Preview annotations. The catalogue only knows previewed components — an un-previewed composable is invisible to it, and that's exactly the component that gets duplicated.

Freshness: if the showcase output is missing or predates recent UI work, either regenerate (generateShowcase<Variant> — takes minutes, warn first) or continue source-only and say confidence is reduced.

Cap the shortlist at ~6 suspects. Tell the user what was searched and what the shortlist is before going visual.

Tier 2 — visual triage (in a subagent)

Snapshot images are context-expensive; N of them in the main conversation crowds out the actual task. Delegate the comparison:

1. Save the target visual to a file first (fetch the Figma screenshot and write it to a temp file, or use the spec's mock image). Subagents do not share your conversation — pass file paths only.

2. Spawn one subagent with: the target image path, each suspect's id + base snapshot path (from the module's showcase images/ — base snapshots only, ignore _<category>_<value> variant files), and this verbatim output contract:

   View the target image and every suspect image. Compare anatomy and layout — structure, element placement, shape — not theme colors; a re-themed version of the same anatomy is MATCH or PARTIAL, not NO. Return EXACTLY one line per suspect: <component id> | MATCH or PARTIAL or NO | <decisive difference, max 15 words> and one final line: RECOMMEND: <component id or NONE> | <reason, max 20 words> Return nothing else — no image descriptions, no prose.

3. If the environment has no subagents, do the comparison inline and accept the context cost — the discipline (view every suspect, one-line verdicts) still applies.

The three exits

Verdict	Action
MATCH	Use the existing component: report its source file:line and catalogue entry. If it needs adjusting, that's /arkive:design-loop — on the existing component, not a copy
PARTIAL	Extend or parametrize the existing component. If the extension would break, fork, or duplicate a shared component's API — STOP and ask; that's a product decision
NO (all suspects)	Build new — and finish by annotating it (/arkive:annotate: name, group, tags, designNodeId) so the next search finds it

Special case: found in source but not in the catalogue → point at it, and recommend adding a preview + annotation while touching it. That gap is how duplicates happen.

Red flags — STOP

• Skipping this check because the component "is obviously new".
• Concluding "no match" from metadata alone when suspects existed but were never viewed.
• Concluding "no match" from a missing/stale catalogue without having grepped the source.
• A subagent report with prose or image descriptions instead of the one-line contract.
• Building new after a PARTIAL because extending looked harder — that's the STOP-and-ask case, not a build-new license.