review-latency

Enter planning mode. Deep-audit both latency-critical paths for anything that adds delay — from micro-optimizations to architectural blockers. Use maximum parallelism — spawn explore agents for independent paths.

Context

This is the highest-priority performance surface. The project's overriding goal is responsiveness — the user must never feel lag when Alt-Tabbing. Every microsecond matters on these paths because costs compound: a 50μs waste in an eligibility check runs 50× per focus event = 2.5ms. A cache miss in display list building runs every paint. Micro-optimizations are not just welcome, they're the point.

The architecture is single-process: producers, window store, and GUI all run in MainProcess. There is no IPC on the critical path — the enrichment pump (icon/process resolution) is async and off the hot path.

Scope Boundaries

This skill covers the control flow around rendering — everything from keypress to "start painting" and from "painting done" to "window visible." It does NOT cover per-frame rendering internals, which have dedicated skills:

• /review-paint — D2D paint pipeline: pre-render + BeginDraw→EndDraw, 8-layer compositor, effect helpers, per-frame allocations in gui_paint.ahk, gui_effects.ahk, gui_bgimage.ahk, gui_gdip.ahk, gui_math.ahk
• /review-d3d — D3D11 shader host: d2d_shader.ahk buffer allocations, state calls, GPU readback, compute dispatch
• /review-shaders — HLSL pixel shader source: ALU, transcendentals, loop optimization in src/shaders/*.hlsl

However, the frame pacing mechanism IS in scope for this skill — the three-tier pacing in gui_animation.ahk (compositor clock → waitable swap chain → QPC spin-wait) directly affects input-to-photon latency. The decision of when to render (frame pacing) is latency-critical; what to render (paint internals) is not.

Similarly, DComp operations that affect overlay visibility timing are in scope: D2D_SetClipRect, D2D_Commit, DWM cloaking/uncloaking sequences. Present(0,0) is non-blocking post-Phase 1 — verify this hasn't regressed.

If you find a latency issue that lives inside the rendering pipeline (e.g., "paint takes too long because of X"), note it briefly and defer to the appropriate skill.

The Two Hot Paths

Path 1: Window Change → Store

An external event (focus change, window created/destroyed, komorebi workspace switch) must update the window store as fast as possible so the data is fresh when the user Alt-Tabs.

WinEventHook callback  ─┐
Komorebi subscription   ─┼──► Eligibility check ──► Store upsert ──► Dirty tracking
WinEnum (on-demand)     ─┘

Key files:

• src/core/winevent_hook.ahk — primary producer, fires on every focus change
• src/core/komorebi_sub.ahk, komorebi_state.ahk, komorebi_lite.ahk — workspace tracking with multi-layer cache
• src/core/winenum_lite.ahk — full window enumeration (startup, snapshot)
• src/shared/blacklist.ahk — Blacklist_IsWindowEligible() — called per-window, per-event
• src/shared/window_list.ahk — store internals, upsert, dirty tracking, display list
• src/core/mru_lite.ahk — fallback MRU (if WinEventHook fails)

Questions to ask:

• How much work does each producer callback do? Is any of it deferrable?
• Are eligibility checks doing redundant work (re-checking things that haven't changed)?
• Is the store upsert doing unnecessary copies or recomputations?
• Are caches (komorebi state cache, blacklist compiled patterns, etc.) actually effective? Any cache misses on the hot path?
• Is dirty tracking granular enough, or does a single-window change trigger broader recomputation?

Path 2: User Action → Pixels

The user presses Alt → Tab and must see the overlay with correct data as fast as possible. Then each subsequent Tab press must update the selection and repaint instantly.

This skill focuses on the control flow and data preparation — from keypress through state machine to the point where rendering begins, and from rendering completion to overlay visibility. The rendering pipeline itself (D2D draw calls, effects, compositing) is covered by /review-paint.

Alt down ──► Pre-warm (refresh data early)
Tab down ──► Freeze list ──► Build display items ──► [Paint — see /review-paint] ──► Show overlay
Tab again ──► Move selection ──► [Repaint]
Alt up   ──► Activate window ──► Hide overlay
Escape   ──► Cancel ──► Hide overlay

Key files:

• src/gui/gui_interceptor.ahk — keyboard hook callbacks, event dispatch
• src/gui/gui_state.ahk — state machine transitions
• src/gui/gui_input.ahk — input handling, selection movement
• src/gui/gui_data.ahk — live data layer, refresh, pre-cache, display eviction
• src/gui/gui_overlay.ahk — show/hide mechanics (DWM cloaking, anti-flash)
• src/shared/gui_antiflash.ahk — DWM cloaking / alpha sequencing
• src/gui/gui_monitor.ahk — monitor detection, DPI
• src/gui/gui_workspace.ahk — workspace label building
• src/gui/gui_pump.ahk — enrichment pump integration

Out of scope (covered by dedicated skills):

• src/gui/gui_paint.ahk — per-frame rendering internals (/review-paint)
• src/gui/gui_effects.ahk — 8-layer compositor internals (/review-paint)
• src/gui/gui_gdip.ahk — D2D resource management (/review-paint)
• src/gui/gui_math.ahk — layout calculations (/review-paint)
• src/gui/gui_bgimage.ahk — background image layer (/review-paint)
• src/gui/d2d_shader.ahk — D3D11 interop (/review-d3d)
• src/shaders/*.hlsl — pixel shaders (/review-shaders)

In scope from gui_animation.ahk (the frame pacing / latency boundary):

• _Anim_FrameLoop — three-tier pacing: compositor clock wait vs waitable swap chain vs QPC spin-wait
• Frame skip logic for explicit FPS caps (different behavior per pacing tier)
• DCompositionBoostCompositorClock — DRR boost on show/hide
• Anim_EnsureTimer deferred start guard (STA pump safety — indirectly affects first-frame latency)

Questions to ask:

• What work happens between Tab press and the call to GUI_Repaint()? Is any of it unnecessary or reorderable?
• Is display list construction doing work that could be pre-computed during pre-warm?
• How long does overlay show take after paint completes? Is DWM cloaking/uncloaking adding delay?
• Does the state machine transition path have unnecessary intermediate states or checks?
• What happens on subsequent Tab presses — does data preparation repeat unnecessarily?
• Is window activation (Alt-up) blocking on komorebic or Win32 calls?

Cross-Cutting: Main Thread Blocking

The keyboard hooks run on the main thread. Anything that blocks the main thread delays hook processing. This includes:

• Timer callbacks that do heavy work (check with query_timers.ps1)
• Producer callbacks that take too long inside Critical "On" sections
• Synchronous file I/O (config reads, stats writes, log writes)
• D2D operations outside the paint path (e.g., resource creation triggered by config change)
• Any DllCall that might block (synchronous Win32 calls)

This is separate from the two paths above — even if Path 1 and Path 2 are individually fast, a long-running timer callback between Alt-down and Tab-down steals time from hook processing.

Explore Strategy

Split by hot path (run in parallel):

• Path 1 agent: All producers in src/core/, eligibility in blacklist.ahk, store internals in window_list.ahk. Focus on per-event callback cost.
• Path 2 agent: gui_interceptor.ahk, gui_state.ahk, gui_input.ahk, gui_data.ahk, gui_overlay.ahk, src/shared/gui_antiflash.ahk, gui_workspace.ahk. Focus on keypress-to-paint-call and paint-done-to-visible sequences. Do NOT audit the rendering pipeline itself.
• Cross-cutting agent: query_timers.ps1 output, Critical section durations, any synchronous I/O on the main thread. Scan all src/gui/ and src/core/ files for blocking operations.

Tools

• query_timers.ps1 — inventory all timers, find heavy callbacks
• query_state.ps1 — trace state machine transitions for the Alt-Tab flow
• query_interface.ps1 <file> — public API surface of hot path files
• query_function.ps1 <func> — extract function bodies without loading full files
• query_callchain.ps1 <func> — trace call depth from hot path entry points

Assessment Format

Surface everything — do not auto-exclude findings based on estimated size. Micro-optimizations on high-frequency paths are the point of this review.

For each finding, provide an honest assessment:

Columns explained:

• Current Cost: Estimated per-invocation cost (use flight recorder / paint timing data if available, otherwise estimate from code complexity)
• Frequency: How often this runs in the critical path (1× per Alt-Tab? 50× per focus burst? Per-pixel? Per-window?)
• Compound Cost: Current Cost × Frequency — the actual user-felt impact
• Complexity: How hard is the fix? One-line change, medium refactor, architectural change?

Do not filter. A 10μs saving that runs 100× per paint (1ms compound) is worth knowing about even if the fix is complex. The user decides the tradeoff.

Validation

After explore agents report back, validate every finding yourself. This codebase has extensive caching and optimization already — what looks like a miss may be handled elsewhere.

For each candidate:

1. Cite evidence: "I verified by reading file.ahk lines X–Y" with actual code quoted. Trace the full execution path, not just one function.
2. Trace the frequency: Don't guess — trace when and how often this code actually runs. A function called once at startup is not a hot path finding.
3. Check for existing optimization: This codebase has been through multiple optimization passes. Before flagging something, check if there's already a cache, early-exit, or pre-computation handling it.
4. Counter-argument: "What would make this optimization unnecessary or counterproductive?" — Does it add complexity that makes the next optimization harder? Does it break an invariant?
5. Observed vs inferred: Did you trace the execution path through all branches, or infer the cost from reading one function in isolation?

Plan Format

Section 1 — Path 1 findings (Window Change → Store):

Section 2 — Path 2 findings (User Action → Pixels):

Section 3 — Cross-cutting (Main Thread Blocking):

Order within each section by compound cost (highest first). Do not omit low-compound-cost findings — list them at the bottom.

Ignore any existing plans — create a fresh one.