Manwë

Farthest-sighted of the Valar, who from his throne on Taniquetil sees furthest and clearest. This skill does not mend what it sees, as Fëanor does — it only looks, and says plainly what does not match.

What it needs

• <target> — the built page or app view, as a browser-loadable URL (a running dev server, a Henneth-served page, a file:// URL) or a Flutter app already running on a booted emulator/simulator. You boot the device and navigate to the target screen; Manwë only captures it.
• <spec> (optional) — the ideal to weigh against: a PNG, an HTML page/URL (shot to a PNG first), or a Henneth wireframe already sitting in ~/.skadi/henneth/ from the same session's Previews step. Absent a spec, Manwë falls back to the sibling comparison below.

With a spec

1. Resolve the spec image once — copy a PNG as-is, or shoot an HTML/URL spec — exactly as Fëanor's "Resolve the spec image once" step does, into ~/.skadi/henneth/manwe-spec.png.
2. Shoot <target> once into ~/.skadi/henneth/manwe-shot.png, via feanor-shot.sh (web) or feanor-flutter-shot.sh (Flutter). If the hook exits non-zero, stop and fail loud — there is nothing to weigh. For a Flutter target whose source has moved since the app was last reloaded, reload it first (~/.claude/hooks/flutter-daemon.sh reload, the /narya skill) and shoot only on a 0 — a verdict weighed against a stale binary is worse than none.
3. Read both images and name the deltas — Fëanor's oracle ("The oracle" in feanor/SKILL.md): perceptual, not pixel diff; color, layout, proportion, spacing, presence, typography, ordered by how much each moves the eye.
4. Report ALIGNED (no material deltas) or DELTAS (the checklist) — one pass, no retry, no edit. A DELTAS verdict is a finding for the caller to act on or name as a knowing exception, not a defect Manwë fixes itself.

Without a spec

1. Shoot <target> once, the same way, for a visual sanity check — obvious breakage (overflow, misalignment, a missing element) is worth naming even with no spec to measure against. Where an icon and a text share a row, measure both ink centres with feanor-ink-centre.sh, per Fëanor's "Judge an icon and a text sharing a row by their ink" rule, and flag a gap over 1 logical px. The sibling comparison below reads source values and cannot see this, because both boxes can sit centred while the glyphs do not.
2. Find at least one sibling of the component being weighed — another instance of the same widget in the same screen, or the same component family elsewhere in the working tree. Infer it from the target's source location; ask once if more than one plausible sibling exists and they disagree.
3. Read the sibling's actual current source values — not its name or intent — for the properties a reader expects to match: padding, margin, spacing, sizing, typography. This is the same discipline the Compliance Review's sibling-string check already applies to user-facing text, generalized to layout.
4. Report REASONABLE (the shot looks sound and no compared property diverges) or DIVERGENT (name each divergent property, target's value beside the sibling's).

With --human — the eye's double confirm

/manwe <target> [<spec>] --human runs the pass above unchanged, then puts the verdict before the human's eye. Without the flag nothing below happens.

1. Take the pairs. One <target> <spec> pair, several of them, or captures already in the session (a flow of frames against their shots) — one page covers all. A pair with no spec or no shot yet is still listed.

2. Write a manifest to the scratchpad and render it:

   ~/.claude/hooks/manwe-compare.py <manifest.json>
   

   Manifest fields are in the hook's docstring: slug (target or feature — one page per run, never overwriting another's), title, and per pair title, spec, shot, optional specLabel/shotLabel, specHint/shotHint (what the human must do for a missing side, e.g. "choose a protocol and I'll shoot it"), and notes (the measured deltas from the pass). The hook writes manwe-compare-<slug>.html and copies the images into a manwe-compare-<slug>/ folder beside it (cleared on each run of the same slug); the page draws spec left and shot right at the same CSS width, with a collapsible Overlay and opacity slider per pair. Where the spec and shot differ in device pixels (a 1024 lp PNG against a 2048 px iPad shot) the page's equal widths already line up one logical px.

3. Print the page URL — the port is in ~/.skadi/henneth/.henneth-port (boot /henneth first if it is absent): http://localhost:<port>/manwe-compare-<slug>.html.

4. Ask the human with AskUserQuestion: Matches or Mismatch (describe).

5. Carry both words in the verdict line, the model's first and the human's beside it: DELTAS (human: matches), ALIGNED (human: mismatch — <their words>). The two may disagree; report the disagreement, do not settle it.

Still findings only — the flag adds an observer, not an editor.

Notes

• Single pass, not a loop. Fëanor mends until aligned or capped; Manwë looks once and reports. Chase a DELTAS or DIVERGENT verdict by editing the source and re-running Manwë, or hand the fix to Fëanor when a spec is in hand and the loop is wanted.
• The review reads, it does not write — same rule as the Compliance Review's. A finding is surfaced, never silently patched.
• Shots land in Henneth the same way Fëanor's do, newest first, so the target and (when present) the spec sit side by side for the eye as well as the model.
• Headless browser / Flutter requirements are Fëanor's — see its Notes — since the same hooks are reused verbatim.