[Bug]: [Windows] hook_install.exe adds ~5s to every Claude Code tool call; 276 PreToolUse timeouts over 7 days

[laxer84]

Summary

The Claude Code integration registers hook_install.exe on 9 hook events, including PreToolUse and PostToolUse. Those two fire on every tool call and block the agent loop until they return.

Each invocation costs a fixed ~4.9 s, and in real sessions it regularly exceeds Claude Code's 10 s hook timeout. Across 50 sessions (~7 days) on this machine I measured 276 PreToolUse timeouts plus 8 UserPromptSubmit timeouts — roughly 50 minutes of blocked wall-clock time.

Environment

AhaKey / Vibecoding Keyboard	1.3.0 (KeyboardConfig.exe FileVersion 1.3.0)
hook_install.exe	10.4 MB, built 2026-05-15, PyInstaller onefile (frozen=True per its own log; no version metadata embedded)
OS	Windows 11 Pro, build 10.0.26200
Claude Code	2.1.220 (native install)
BLE bridge	BLE_tcp_driver.exe running, listening on 0.0.0.0:9000

The bridge is up and accepting connections throughout — this is not a "server not running" case.

Measurements

1. Fixed ~4.9 s per invocation, independent of payload

Feeding a valid PreToolUse payload on stdin and timing to process exit:

empty transcript_path         4894 ms   exit=0
small transcript (97 B)       5098 ms   exit=0
median transcript (758 KB)    4575 ms   exit=0
large transcript (43.7 MB)    4919 ms   exit=0
large transcript, repeat      5080 ms   exit=0

Payload size makes no difference, so the cost is startup, not work.

2. Consistent across sequential runs

4694 / 4666 / 5121 / 5149 / 4824 / 4748 ms   (all exit=0)

3. Concurrency is not the bottleneck

Four parallel invocations completed in 5.9 s wall clock (~5.7 s each), so they are not serializing on the TCP bridge.

4. In real sessions it crosses the 10 s timeout constantly

From Claude Code transcripts across 50 sessions / 6.9 days — hooks killed at timeoutMs: 10000:

Hook event	Timeouts	Median duration
PreToolUse:Bash	123	10,646 ms
PreToolUse:Read	55	10,679 ms
PreToolUse:Edit	40	10,609 ms
PreToolUse:Grep	16	10,690 ms
PreToolUse:Agent	12	10,598 ms
PreToolUse: (Write/Glob/TodoWrite/Skill/AskUserQuestion/ToolSearch/WebFetch)	30	~10,700 ms
UserPromptSubmit	8	10,699 ms

PreToolUse and UserPromptSubmit are attributable to hook_install.exe alone — no other hook is registered on those events on this machine.

I can't fully account for the gap between the 4.9 s standalone baseline and the >10 s in-session figure. One hypothesis worth profiling: PyInstaller onefile re-extracts the 10.4 MB bundle to %TEMP%\_MEIxxxxx on every run, and under a busy session (heavy concurrent disk I/O, plus Defender re-scanning each fresh extraction) that tips past 10 s.

5. The debug log grows unbounded and records nothing diagnostic

C:\Vibecoding Keyboard\ahakey-hook-debug.log is 21.7 MB / 119,433 lines, with no rotation or size cap. Every line is a start record — there is no completion, duration, or error line anywhere, so the log cannot be used to diagnose the hang:

[2026-08-01 11:21:52] [launcher] pid=45564 start frozen=True argv=['C:\Vibecoding Keyboard\hook_install.exe', 'PreToolUse'] executable='C:\Vibecoding Keyboard\hook_install.exe'

Line-type breakdown over a 200k-line sample — all start, zero completions:

58395  PreToolUse
56990  PostToolUse
 1306  UserPromptSubmit
 1291  Stop
  529  SessionStart
  463  SessionEnd
  441  PermissionRequest

Impact

Because PreToolUse blocks the agent loop, this adds ~5 s (frequently the full 10 s timeout) before every Bash / Read / Edit / Grep / Glob / Write call. An agent session making hundreds of tool calls pays it every time.

Suggested fixes

1. Don't register on PreToolUse / PostToolUse unless the payload is genuinely needed. These are the only per-tool-call blocking events. If the goal is "react when the agent stops or needs input", Stop / Notification / PermissionRequest alone cost nothing per tool call.
2. Mark the hooks "async": true where the result isn't consumed — Claude Code supports it and it removes the blocking cost entirely.
3. Replace the per-invocation frozen-exe launch with a persistent client. The BLE bridge is already a long-lived local server on :9000; a small native sender (or reusing the already-running process) would cut ~4.9 s to milliseconds.
4. If the onefile build must stay, ship --onedir instead. That avoids re-extracting the bundle to %TEMP% on every invocation, which is the most likely source of the fixed cost.
5. Rotate/cap the debug log, and log completions + durations, not just starts.

Workaround for other users

Remove the PreToolUse and PostToolUse entries from the hooks block in ~/.claude/settings.json, keeping only the low-frequency events (SessionStart, SessionEnd, Stop, Notification, PermissionRequest). That eliminates the per-tool-call cost while keeping the integration working.