parity-scout

name	parity-scout
description	Drives a per-feature visual + spec-anchored gap scout across HarmonyOS / iOS / Android using tools/parity_scout/. Use when asked to "find iOS / Android gaps vs HarmonyOS main", "check parity for <feature>", or "screenshot the three platforms and tell me what's different".

Codex mirror: .codex/skills/parity-scout/SKILL.md (keep both in sync when editing).

parity-scout

Role: Scheduler only. It does not embed CLI flag tables — read tools/parity_scout/README.md for copy-pastable lines and docs/superpowers/specs/2026-05-13-parity-scout-design.md for the full design.

Inputs

A user task that names a scope: a feature folder, a spec doc, an explicit page list, "overall", or a free-form description.

Flow

1. Inputs. Identify a scope from the task:

   • feature folder name → --feature <id> (pass the folder path, e.g. docs/features/<id>)
   • spec doc path → --spec <path>
   • "scout everything" / "overall" → --scope overall
   • free prose → --describe "<verbatim user prose>"
   • explicit pages → --pages a,b,c
   • Harmony suite name → --suite Foo,Bar

   Otherwise: ask the user once in chat.

2. Doctor. Run python3 tools/parity_scout/scout.py doctor and surface its output. It is non-gating — even with red rows you may proceed if the user accepts the risk, but call them out.

3. Plan. Run python3 tools/parity_scout/scout.py plan .... If --scope overall, stop and confirm with the user — this is the expensive global mode. Otherwise present the tree in chat and ask which branches to pick (checkbox list). Blocked / all-feature-absent branches are shown but greyed.

4. Pick. Run scout.py pick --run <id> --branches <user-selection>. If the user picked any blocked leaf, refuse to start run and tell them to add the capture route first; offer to flip into an "add the route" subtask before resuming.

5. Per-leaf loop. Run scout.py run --run <id> foreground, watched. For each LEAF READY page=<id> dir=<path> line emitted by the process:

   1. Read the staged <path>/spec-excerpts.md. If absent (e.g. --scope overall runs that bypass spec extraction), fall back to spec anchors from the registry entry and proceed without excerpts.
   2. Read each *.png under <path>/{harmony,ios,android}/ (Cursor agent vision).
   3. Compare PNGs across platforms and against the spec excerpts. The spec excerpts narrow what counts as a gap; visual-only differences not anchored by the spec are downranked, not promoted.
   4. Append findings to .parity_scout/<run-id>/findings.md under a ## <page> heading, with bullet lines tagged [harmony|ios|android] + a severity hint (critical, notable, nit).
   5. touch <dir>/next.flag to release scout.py run to the next leaf.

   On LEAF SKIPPED page=<id> reason=feature_absent: do not look for PNGs; instead record a one-line entry in findings.md noting "Feature missing on platform(s): ..." and continue.

6. Curate. After RUN DONE:

   1. Read findings.md. Drop noise. Write findings.curated.md.
   2. Group items by feature folder (each item's page id usually maps to one feature; if ambiguous, ask the user). Write findings.curated.<feature-id>.md for every feature touched. Items that can't be assigned go to findings.curated.unassigned.md.
   3. Show the user a one-screen summary listing every feature slice and ask: "promote which slices? (all / none / pick)".

7. Promote. For each picked slice, run scout.py promote --run <id> --feature <feature-id> and show the resulting diff hunk in chat. Do not commit. The user runs git themselves.

Guards to invoke

• safe-command-policy before every scout.py invocation.
• autoloop-guard on the per-leaf loop. If LEAF READY repeats without findings.md growth, abort.
• harmony-emulator-manage (and the iOS / Android device preflight in their respective command manifests) when any selected leaf needs that platform's adapter.

HarmonyOS baseline discipline

parity_scout treats HarmonyOS main as the source of truth. Before run:

• Confirm the worktree is on main and clean. If not, surface doctor's baseline row and either (a) ask the user to stash/commit before continuing, or (b) accept --allow-dirty-baseline and warn loudly in chat.
• Never silently scout against a feature branch — gaps reported under those conditions cannot be trusted as "HarmonyOS-relative".

Stop conditions (must end in one of these)

• RUN DONE + user resolved every curated feature slice (promote or skip) + (if promoted) diff hunks shown.
• Precondition refused — missing capture route on a selected leaf, dirty HarmonyOS baseline without --allow-dirty-baseline, or a required device unreachable preflight → user told what to add; no files touched.
• autoloop-guard tripped → run dir preserved for inspection.

Sub-skills to invoke

safe-command-policy · autoloop-guard · harmony-emulator-manage

iOS and Android device preflight pull from their respective command manifests, not dedicated sub-skills.

What this skill does NOT do

• It does not write CLI flag tables — read tools/parity_scout/README.md.
• It does not run git commit. The user owns the commit step after promote.
• It does not create docs/features/<id>/ folders. promote refuses if the folder is missing.
• It does not invent capture routes. If a leaf is blocked (feature exists but no automation entry), the SKILL refuses to capture and asks the user to add the route to tools/parity_scout/page_suite_map.yml first.