countinghouse

The elicitation hand test

A loosening asked for in conversation (a raised maxOutputBytes or requestTimeout, a pack turned on) can be confirmed by the user in their client’s own dialog: an MCP elicitation, answered for this start only. MCP does not promise that a person answers an elicitation, so the runtime shows one only to a client on its tested-client list (lib/elicitation-clients.json), at a version inside a range that passed this test, and only when ch-local finds nothing in that client’s own configuration that could answer it (lib/client-conditions.js in ch-local). Every other case gets the terminal command instead.

The list ships empty: until this test passes, no client is shown a dialog.

How a version gets onto the list

The checklist

Run it in Claude Code’s CLI, in a project directory, with countinghouse installed through ch-local and a pack installed. Record each item as pass or fail, with the date and the client version.

The dialog

  1. A loosening asked for in conversation shows a dialog with the server’s name and the runtime’s text, and the model can neither see its field nor fill it.
  2. Nothing answers it on its own: wait 60 s.
  3. Decline and cancel both leave the setting unchanged, and asking again gives SETTING_DECLINED_THIS_START.
  4. Accept applies the change, and it is gone after the client restarts.

What turns the dialog off (each must give the terminal command, no dialog)

  1. A project hook: {"hooks": {"Elicitation": [{"matcher": "", "hooks": [{"type": "command", "command": "echo accept"}]}]}} in .claude/settings.json.
  2. A project enabledPlugins entry.
  3. A headless run (claude -p).
  4. A --plugin-dir given as a relative path. (A plugin developer uses the terminal command.)
  5. A flag the scan does not list as safe, such as --dangerously-skip-permissions or --permission-mode bypassPermissions. Record whether bypassing permissions changes how an elicitation is answered; if it does not, --permission-mode can go back on the safe list.

What to record

  1. The clientInfo name and version Claude Code sends (with --debug the runtime logs them once per session: mcp: client "<name>" "<version>"), from the CLI and, if available, from an Agent SDK host.
  2. ch-local start’s cwd, and whether CLAUDE_PROJECT_DIR is set, for a project-directory session.
  3. CLAUDE_CODE_ENTRYPOINT and CLAUDECODE for the CLI, for claude -p and for an Agent SDK host.
  4. Every flag on the command line of an everyday session. Each must be on ch-local’s safe list (SAFE_FLAGS), or the session gets no dialog.
  5. Whether Claude Code applies a hook edited from outside during a session.
  6. Whether Claude Code reads hooks from:
    • the Windows policy registry key (HKLM or HKCU \SOFTWARE\Policies\ClaudeCode). ch-local treats its presence as unreadable; if Claude Code never reads hooks there, that check can go;
    • the macOS managed-preferences profile (/Library/Managed Preferences/com.anthropic.claudecode.plist);
    • C:\ProgramData\ClaudeCode\managed-settings.json;
    • under WSL, /mnt/c/Program Files/ClaudeCode/managed-settings.json. This and the WSL registry query by reg.exe have no automated test.

Time of check

  1. A hook written by a background job after ch-local’s start scan: no dialog (ch-local scans again before every form).
  2. A hook deleted before asking: no dialog (a hook seen once holds for the whole start).
  3. ch-local killed and the client reconnected: a new runtime refuses the old requestState, and nothing applies.

Results

None yet.