fix(hooks): AfterTool hang on large output or stdin stall in Gemini CLI

[tndhuy]

Description

When using context-mode with Gemini CLI (Node v24+), the AfterTool hook frequently fails with a Hook(s) failed for event AfterTool warning. This happens because the hook tries to read from stdin without a timeout, and processes potentially large tool responses (multi-MB) within the strict <20ms budget of the CLI.

Symptoms

• Constant AfterTool failed warnings in Gemini CLI UI.
• Hook processes hanging or being killed by the system.
• No clear error in aftertool-debug.log because the process is terminated mid-execution.

Root Cause

1. Stdin Hang: readStdin() lacks a timeout guard.
2. Performance Budget: Extracting events from very large tool outputs (e.g., massive run_shell_command output) exceeds the time limit.

Proposed Solution (Applied locally)

I have applied a patch to hooks/gemini-cli/aftertool.mjs with:

• A 10s setTimeout guard on stdin to prevent permanent hangs.
• A size check: if tool_response > 1MB, skip event extraction to preserve CLI responsiveness.

// Stdin timeout guard
const stdinTimeout = setTimeout(() => process.exit(0), 10000);

// ... readStdin ...
clearTimeout(stdinTimeout);

// Response size guard
const responseSize = typeof input.tool_response === 'string' ? input.tool_response.length : 0;
if (responseSize > 1024 * 1024) {
  process.exit(0);
}

Environment

• OS: Darwin (macOS)
• Node: v24.4.1
• Gemini CLI: Latest

[murataslan1]

Validated locally on macOS + Node v24.4.1 + Gemini CLI v0.37.0.

Both guards (stdin timeout + 1MB response size check) work as expected:

• Stdin stall: process exits cleanly after 10s instead of hanging
• 2MB response: skipped in <10ms, logged as SKIP
• Normal responses: still processed correctly
• Live Gemini CLI session: no AfterTool failed warnings

Full test details and patch: https://gist.github.com/murataslan1/3ca3dc332a10315ddd6577d16fa2cb3e

---

One thought: the same readStdin() without a timeout guard is shared across all platform hooks (Cursor, Codex, Kiro, VS Code Copilot) via session-helpers.mjs. Might be worth adding the timeout there too, or at least in each platform's aftertool entrypoint, so other CLIs don't hit the same hang.

[murataslan1]

Went through hooks/gemini-cli/aftertool.mjs to check if the hang is still reproducible from the current code.

Looks like there are already guards in place:

• Line 24: 10-second stdin timeout (setTimeout(() => process.exit(0), 10_000))
• Lines 37-40: 1MB size threshold that skips processing for oversized responses

if (responseStr.length > 1024 * 1024) {
  appendFileSync(DEBUG_LOG, `[...] SKIP: ${input.tool_name} (response ${responseStr.length} bytes)\n`);
  process.exit(0);
}

These weren't mentioned in the original report, so either they were added after the issue was filed, or the hang happens before these guards kick in (e.g. the readStdin() call itself blocking indefinitely on a pipe that never sends EOF).

Might be worth checking with the reporter whether this is still reproducible on 1.0.75, since the guards look like they should cover the described scenario.