| name | review-race-conditions |
| description | Audit for race conditions in timers, hotkeys, callbacks, and shared state |
| user-invocable | true |
| disable-model-invocation | true |
Enter planning mode. Systematically audit the codebase for race conditions. Use maximum parallelism — spawn explore agents for independent areas.
AHK v2 Concurrency Model
AHK v2 is single-threaded but not atomic — timers, hotkeys, and callbacks CAN interrupt each other mid-execution. Critical "On" prevents interruption for the current thread. This is the only synchronization primitive available. See .claude/rules/ahk-patterns.md (Race Conditions section) and .claude/rules/keyboard-hooks.md for established patterns.
What to Look For
1. Unguarded shared state mutations
Globals modified inside timer callbacks, hotkey handlers, or event callbacks without Critical "On":
- Incrementing counters:
gRev += 1
- Check-then-act:
if (!map.Has(k)) { map[k] := v }
- Map/Array iteration while another callback could modify the collection
- State transitions (read-modify-write on
gGUI_State or similar)
2. Critical section gaps
Critical "On" present but:
Critical "Off" missing on an early return or continue path
- Critical released mid-operation (e.g., before rendering — this was tried and reverted, see
keyboard-hooks.md "Do NOT Release Critical Before Rendering")
- Critical scope too narrow (protects the write but not the read that informed it)
3. Timer + hotkey interaction
A SetTimer callback modifying state that a hotkey handler also reads/writes. Both can fire in the same thread context. Use query_timers.ps1 to inventory all timers and their callbacks, then cross-reference with hotkey handlers.
4. IPC message handling during state transitions
Named pipe reads processed via timer while the state machine is mid-transition. Check that IPC handlers either defer during transitions or are properly guarded.
5. Producer callbacks during GUI operations
Producers (WinEventHook, Komorebi subscription) firing callbacks that modify gGUI_LiveItems, window store data, or MRU order while the GUI is painting or the state machine is processing input.
Known Safe Patterns (Do NOT Flag)
These have been deliberately designed and tested — flagging them wastes time:
Critical "On" held through the entire GUI_OnInterceptorEvent handler including rendering (~16ms). This is intentional — releasing early causes corruption.
SendMode("Event") instead of SendInput — prevents hook uninstall during sends.
- Async activation with
gGUI_EventBuffer — events buffered while gGUI_Pending.phase != "".
- Lost Tab synthesis (
ALT_DN + ALT_UP without TAB).
- Flight recorder
FR_Record() — pre-allocated ring buffer, writes are inherently safe.
_GUI_LogError() — always-on by design.
Explore Strategy
Split by concurrency boundary, not just file location:
- Hotkey handlers —
gui_input.ahk, gui_interceptor.ahk: all INT_* functions
- Timer callbacks — use
query_timers.ps1 to find all SetTimer targets, then read each callback
- Windows message handlers — use
query_messages.ps1 to map WM_ handlers and senders (interrupt sources)
- WinEvent callbacks —
winevent_hook.ahk: _WEH_* callback functions
- Komorebi callbacks —
komorebi_sub.ahk, komorebi_lite.ahk: subscription event handlers
- IPC handlers —
ipc_pipe.ahk, pump files: pipe read/write paths
- State machine — use
query_state.ps1 to extract specific branches from GUI_OnInterceptorEvent without loading the full function
- Call graph — use
query_function_visibility.ps1 to map which interrupt sources can reach vulnerable code (e.g., "is this function called from both timer callbacks and hotkey handlers?")
- Mutation detail — use
query_mutations.ps1 <globalName> to see per-function mutation sites with assignment operators and assigned values — reveals unguarded writes without reading full files
Validation
After explore agents report back, validate every finding yourself. Race conditions are easy to hypothesize and hard to confirm. Many "races" identified by exploration are actually guarded by Critical sections or by AHK's threading model (only one pseudo-thread runs at a time; interruption only happens at specific yield points).
For each candidate:
- Cite evidence: "I verified by reading
file.ahk lines X–Y" with actual code quoted. Vague line references are not sufficient.
- Trace the interrupt path: Which specific callback could interrupt this code? Is it a timer, hotkey, or WinEvent? At what point in the vulnerable code would the interruption occur?
- Counter-argument: "What would make this race impossible?" — Is there a Critical section in a caller? Does AHK's threading model prevent this specific interleaving? Is the state only accessed from one callback type?
- Observed vs inferred: State whether you saw unguarded concurrent access directly, or inferred it from code structure. Inferred races need stronger evidence.
- Reproducibility: Can this race actually manifest in normal usage, or only under extreme timing? A theoretical race in a once-per-session init path is lower priority than one in a per-keystroke hot path.
Plan Format
Group by severity (data corruption > logic error > cosmetic):
| File | Lines | Race Description | Interrupt Source | Severity | Fix |
|---|
file.ahk | 42–58 | gFoo read-then-write without Critical | Timer _HeartbeatTick | Data corruption | Wrap in Critical "On" / Critical "Off" |
For each fix, note:
- Whether it introduces new Critical section duration that could affect input latency
- Whether the fix is localized or requires coordinating changes across files
Ignore any existing plans — create a fresh one.