devexpress-wpf-theme-customizer

DevExpress WPF Appearance Customization

Read this file to the end before acting. If your reader returned only part of it, call it again for the rest — verification, cleanup and the honesty gate are stated later, and a partial read skips them.

Token efficiency: cost is context × steps, not output. grep/glob to locate, then read only the needed range. Never re-read a file you just wrote, or re-open a source file for something a sub-agent already returned. Delegate bulk lookups to a sub-agent with a short return contract — keep raw payloads out of your context.

"Done" = a verified working change, not advice. Apply the change and confirm it renders by default, even for a "how do I…?" ask. A green build is not proof. Propose XAML-only only when the user asked for a snippet / not to touch their project.

Three paths, in order of preference: 1. Public API (property/style/event) · 2. Scoped implicit style (TargetType-only, parent's Resources) · 3. Theme resource override (copy + inject from installed theme source) — Steps 5–9, in references/resource-override.md.

Every path ends the same way: apply, then prove it rendered — Step 1 baseline, Step 10 check, for path 1/2 exactly as much as for path 3. Path 1/2 skips only Step 5 and the override files; never applying, verifying, or reporting honestly. Read references/apply.md when injecting into a Resources collection.

Sub-agents for search-heavy phases

Dedicated sub-agent per search-heavy phase, returning only the distilled result: Step 5 (source search), Steps 1 & 10 (screenshots), Step 4 (docs).

• Instruct it to read the named reference file first, then return only the specified result.
• No agent-definition files — inline instructions, so it works the same in every harness.
• No sub-agent capability → do the phase inline yourself, using its own stated command. Missing sub-agents never skip a phase.
• Model: mechanical work (file sweeps, extracting a resource block, capturing/describing a screenshot) → Haiku-class. Judgment calls (API-meets-goal, Step 10 verdict, API-vs-override) → main model. Step 4 is delegated for volume but stays on the main model — a cheaper model summarizes away exactly the limitation clauses that matter.
• Trust mechanical sub-agent output as-is. Re-verify only when it flags uncertainty, or the finding alone decides a correctness-critical choice — by checking the claim (file/key exist, value matches the app), not repeating the search.