ui-performance-budget

Use for sidebar lag, chat scroll performance, composer typing latency, dropdown open delay, terminal/sidebar switching, right-sidebar/browser performance, or any UI performance budget change.

Procedure

1. Read macos/PERF.md, docs/adr/0010-interface-governance.md, docs/adr/0040-macos-ux-trace-harness.md, docs/ui/performance-budgets.registry.json, and docs/ui/ux-trace-harness.registry.json. Also read docs/ui/ux-trace-calibration.manifest.json before claiming a synthetic fixture is real-equivalent or approval-ready. Treat CPU, RAM, GPU/Neural Engine, disk, network, battery, thermals, idle behavior, and perceived lightness as governed performance dimensions.
2. Identify the critical flow, matching UX trace KPI, fixture profile, and whether its baseline is approved.
3. For macOS P0 UI work, prefer the UX trace harness path: action-to-visual completion through the agent control bus, geometry/scroll stability, hitches/resources, and structured evidence. Computer Use is witness-only. Normal app overhead is part of the budget: high-cardinality control registries, frame probes, loopback control server, fixture mutation, screenshots, and UX trace JSONL must stay gated to CLAWIX_AGENT_INSTANCE=1 or an explicit diagnostic probe.
4. For harness wiring changes, run node scripts/run_macos_ux_trace_harness.mjs --self-test. For actual performance evidence, run the same runner against an isolated agent control URL and owner token; dry-run output is not runtime evidence. After any run or suite, validate the generated evidence directory with node scripts/verify_macos_ux_trace_evidence.mjs --path <run-or-suite-dir> before using it as proof.
5. For a quick runtime smoke, provision an isolated agent instance with a generated smoke fixture and run node scripts/run_macos_ux_trace_harness.mjs --suite p0 --fixture-profile smoke --control-url <url> --token <token>. Use dense-sidebar, dense-chat, streaming-heavy, terminal-under-load, medium, real-equivalent-private, or worst-case when the investigated flow requires heavier coverage.
6. Capture evidence before optimization using the target performance playbook.
7. Compare against the approved baseline when present. If no approved baseline exists, produce a baseline-capture report for user approval.
8. When overhead is relevant, capture the harness-disabled lane with --harness-disabled-control --write-overhead-control <file> and compare it with --overhead-control <file>.
9. Treat docs/ui/ux-trace-calibration.manifest.json entries marked EXTERNAL PENDING as non-closable without approved private aggregate evidence. Synthetic live baselines can support investigation, but they do not prove private real-mode equivalence by themselves.
10. Do not retune visual timing, layout, animation, or perceived style unless the task is visual-authorized.

Constraints

• No performance fixes from static reading alone.
• No P0 macOS UI performance closure from click dispatch timing, screenshots, or Computer Use alone.
• No acceptance of harness changes that write trace evidence to the main app database or remove the normal-app overhead gates.
• No paid prompts, real service mutations, or production data during capture without explicit approval.
• Missing physical/provider prerequisites are EXTERNAL PENDING.