windows-hook-probe

command
v0.48.0 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Nov 10, 2026 License: MIT Imports: 10 Imported by: 0

Documentation

Overview

Command windows-hook-probe is step 0 of the native Windows plan (Trello card 133, .context/plans/native-windows.md). Copied to plugin/bin/brigade.exe beside the POSIX bootstrap, with plugin/hooks/hooks.json left exactly as it ships, it records every invocation Claude Code makes of it, so one `claude -p` run on a Windows runner (.github/workflows/experiment-windows-hooks.yml) answers the question the whole plan rests on: does an exec-form hook whose `command` is `${CLAUDE_PLUGIN_ROOT}/bin/brigade`, with no extension, resolve to the brigade.exe beside it on Windows? Node's spawner tries `.com` and `.exe` for an extensionless path; whether Claude Code's does is what this measures.

Every invocation appends one block to hook-probe.log under RUNNER_TEMP (TEMP, then the OS temp dir, when unset): the time, argv, the working directory, the executable's own path, what arrived on stdin (the hook input JSON), the environment with every value whose name looks secret redacted, the Claude Code session registry's file names, and, from the UserPromptSubmit hook, the result of writing the auth frame and one user frame into CLAUDE_CODE_MESSAGING_SOCKET (a named pipe on Windows), which measures delivery. It prints nothing to stdout: a hook's stdout is Claude Code's, and the probe has nothing to say there. It always exits 0.

Measured 2026-10-27 (docs/experiments/E11-windows-hook-resolution.md): GO, and every post was delivered as a new user turn, which fired the prompt hook again, which posted again, for 19 minutes. The probe now posts only from a prompt that does not carry its own marker.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL