Skip to content

Windows: agent hooks strand a visible console per event — two causes, one per installer #14815

Closed

Description

@Tauri-EPO

On Windows, every agent hook event leaves behind a visible console window that never closes. With
a few agent lanes running, dozens accumulate; the hook payload is written into each stranded shell,
which tries to execute the JSON and fails.

There are two independent causes, one per installer — and each installer holds half of the working
solution.

Cause 1 — the Claude installer emits command + args, and args is ignored

~/.claude/settings.json, identical for all 11 registered events (UserPromptSubmit, Stop,
StopFailure, SubagentStart, SubagentStop, TeammateIdle, PreToolUse, PostToolUse,
PostToolUseFailure, PermissionRequest, SessionStart):

{
  "type": "command",
  "command": "C:\\WINDOWS\\System32\\conhost.exe",
  "args": ["--headless", "C:\\WINDOWS\\System32\\cmd.exe", "/d", "/c",
           "%USERPROFILE%\\.orca\\agent-hooks\\claude-hook.cmd"],
  "timeout": 10
}

Claude Code's hook schema treats command as the complete command line; there is no args. So
what runs is conhost.exe bare, which opens a console hosting the default shell in the current
directory and stays there. The --headless flag — which is exactly what would suppress the window —
never reaches conhost.

Observed, with a Cursor payload landing in one of those stranded shells:

C:\Users\Enrico\.claude>{ "conversation_id":"686540b0-…","model":"grok-4.6","tool_name":"Grep",
…,"hook_event_name":"postToolUse","cursor_version":"2026.08.11-e8db854",… }
A sintaxe do nome do arquivo, do nome do diretório ou do rótulo do volume está incorreta.

Cause 2 — the Cursor installer omits the headless wrapper

~/.cursor/hooks.json registers 8 events (beforeSubmitPrompt, stop, preToolUse, postToolUse,
postToolUseFailure, beforeShellExecution, beforeMCPExecution, afterAgentResponse), each as a
single well-formed command string:

C:/WINDOWS/System32/WindowsPowerShell/v1.0/powershell.exe -NoProfile -ExecutionPolicy Bypass -EncodedCommand <base64>

The shape is right — one string, forward slashes, payload inside the base64 — but there is no
conhost.exe --headless wrapper and no -WindowStyle Hidden
, so Windows creates a visible console
per invocation.

So: the Claude installer knows the headless trick but delivers it through a field the consumer
ignores; the Cursor installer delivers correctly but omits the trick.

The constraint that rules out the obvious fix

The hook command is executed through bash (Git Bash) in this setup — failures surface as
/usr/bin/bash: line 1: …. That imposes three constraints, and the naive "fold the args into the
command string" violates all three:

  1. Backslash paths collapse. C:\WINDOWS\System32\conhost.exe becomes
    C:WINDOWSSystem32conhost.exe → command not found. Measured on every event after folding.
  2. MSYS path-conversion rewrites /-prefixed switches into paths. cmd.exe /d /c <script> can
    therefore never be written directly in a hook command — /d and /c are mangled before cmd
    sees them.
  3. Forward-slash paths handed to cmd are read as switches. cmd /c "C:/Users/…/claude-hook.cmd"
    fails with 'laude-hook.cmd' is not recognized… — /Users is consumed as a switch and the
    filename is truncated.

The shape that satisfies all of them

Exactly the one the Cursor installer already uses for its payload — no /-switches, path inside the
base64 — wrapped for headlessness:

C:/WINDOWS/System32/conhost.exe --headless C:/WINDOWS/System32/WindowsPowerShell/v1.0/powershell.exe -NoProfile -ExecutionPolicy Bypass -EncodedCommand <base64>

where the decoded PowerShell is & '<abs path to hook .cmd>'; exit $LASTEXITCODE.

Applied locally to both files and verified by effect — a healthy hook exits fast and writes
nothing; exit code alone proves nothing here, because conhost returns 0 even when the child fails:

CLAUDE  PostToolUse: exit 0 in  99 ms | output empty
CURSOR  postToolUse: exit 0 in 261 ms | output empty
CURSOR  preToolUse : exit 0 in  53 ms | output empty

Not the same as #11549

src/main/agent-hooks/hook-stdin-contract.ts documents that a Windows hook must exit without reading
stdin when the Orca env is missing, because "a read-to-EOF then blocks forever and strands a visible
window per hook event."
That guard behaves as designed — verified that panes created with
orca terminal create --command … do receive ORCA_PANE_KEY, ORCA_AGENT_HOOK_PORT,
ORCA_AGENT_HOOK_TOKEN and ORCA_AGENT_HOOK_ENDPOINT, so the early-exit path is never taken here.
Same visible outcome, different cause.

Suggested fixes

  1. One shared Windows launcher prefix used by every agent installer (claude, cursor, copilot, kimi,
    antigravity…), so the headless wrapper cannot be present in one and missing in another.
  2. Regression tests on the emitted entry, not on the argv the installer intends — the current tests
    validate the latter, which is why cause 1 is invisible to them:
    • no args key on the Windows entry;
    • command starts with the shared headless prefix;
    • command contains no /-prefixed switch and no backslash path (both die under bash/MSYS).

Environment

Orca 1.4.182 · Windows 11 build 10.0.26200.9168 · Claude Code and cursor-agent 2026.08.11-e8db854,
running in Orca panes.

Activity

  1. guanwenpeng2001-bot commented on Aug 16, 2026

    @guanwenpeng2001-bot

    Additional reproduction after updating Orca

    I can reproduce the visible-console variant on Windows after updating Orca. This appears to be the same Windows agent-hook installation problem described in this issue, and is still present in Orca 1.4.183.

    Environment

    • OS: Windows 11, build 10.0.26200.9168
    • Orca: 1.4.183 (Windows desktop)
    • Agent: Grok CLI, launched from an Orca terminal with --permission-mode bypassPermissions
    • Workspace: AOS_AGENT_EXTERNAL_AGENT_ADAPTER-集成分支

    Observed behavior

    Whenever the agent invokes a tool, including grep, a visible elevated conhost.exe window appears repeatedly. The window displays the raw hook payload at the shell prompt instead of remaining hidden. The payload contains fields such as:

    {"hookEventName":"pre_tool_use","toolName":"grep","toolInput":{...}}

    This repeats for tool events and can steal focus from the Orca terminal. The behavior started after the Orca update.

    Local evidence

    The generated global Claude-compatible settings contain hook entries shaped like this for PreToolUse, PostToolUse, and other lifecycle events:

    {
      "type": "command",
      "command": "C:\\WINDOWS\\System32\\conhost.exe",
      "args": ["--headless", "C:\\WINDOWS\\System32\\cmd.exe", "/d", "/c",
               "%USERPROFILE%\\.orca\\agent-hooks\\claude-hook.cmd"],
      "timeout": 10
    }

    The hook script then passes the payload to curl with --data-urlencode "payload@-". The visible shell and the raw JSON strongly suggest that the Windows launcher is not being consumed as a hidden, non-interactive process, and/or that the hook stdin is not receiving a proper EOF. The Grok session also reports repeated global/settings hook failures with ParserError.

    Expected behavior

    Agent hooks should execute without opening a visible console, should consume or close stdin correctly, should exit promptly, and should not expose internal session/tool metadata or steal keyboard focus.

    Related issues

    Please test the emitted Windows hook configuration against the actual consumer. A complete hidden launcher command (or a direct HTTP hook) plus an explicit EOF for the curl payload appears necessary; validating only the intended argv is not sufficient.

  2. amomurawskiio425-cloud commented on Aug 23, 2026

    @amomurawskiio425-cloud

    Great write-up — this matches what I found, and the two-cause split is the key insight. Claude installer writes command+args (args ignored → bare conhost + window), Cursor installer omits the headless wrapper. Either way the host runs a console that never closes, writes the JSON payload into an interactive shell, and tries to execute it as a command — exactly the "invalid filename/directory/volume label syntax" garbage you show.

    The thing that made it a war, though: fixing by hand doesn't stick. Orca re-registers these on every start, one per installer, so the windows come back and accumulate again.

    I got out of it by clearing Orca's residue programmatically — Coffee CLI removes only the Orca-injected hook entries on launch (leaves mine alone, skips while Orca runs, so no conflict). That broke both causes at the source instead of chasing windows one at a time.

    Caveat: it can't restore anything Orca already removed, and it won't run while Orca is active. Worth a shot for the residue itself. https://github.com/edison7009/Coffee-CLI/releases/tag/v3.3.8

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions