invoke-harness

Invoke Harness

Run the selected harness as a bounded child process. Read only its reference:

• Codex
• Claude Code
• Grok
• OpenCode
• Cursor CLI
• GitHub Copilot CLI
• Cockpit

Procedure

1. Resolve the harness's documented executable with command -v, then verify its identity and flags with current --version and --help output. Do not accept a binary merely because its filename matches: abort if its product identity or required flags do not match the selected harness. Local help may override flag names; it does not replace the reference's noninteractive agent command with another subcommand unless the user asked for that product feature.
2. Use the harness's noninteractive mode. Set the model and working directory explicitly. Resolve nicknames to the CLI's actual model IDs, and confirm the ID with the reference's model-list command when it has one; otherwise use its documented fallback and report the ID as unverified. Start a fresh session unless the user requests continuation.
3. Pass short prompts as one quoted argument. For long prompts, use the documented stdin or prompt-file method; never interpolate untrusted prompt text into shell syntax.
4. External harness invocations always run without a harness sandbox and with all permissions bypassed. Use the exact no-sandbox/all-permissions flags in the selected reference. Do not substitute a restricted mode; abort if the installed harness lacks the required flags.
5. Bound every child process with the harness's timeout option or the parent process runner's timeout. Capture stdout, stderr, and the exit status. Do not treat partial output, timeout, or a zero-finding review as success without checking the resulting files or diff.

This is intentionally unsafe: only invoke a harness in a directory and environment the user has placed in scope.

Before invoking Claude Code, warn the user that it consumes Claude usage credits. Treat Cockpit noninteractive mode as experimental until a harmless local smoke test succeeds.

Executables

Use only these product commands; aliases supplied by a different product do not count:

• Codex: codex
• Claude Code: claude
• Grok: grok (never its compatibility alias agent)
• OpenCode: opencode
• Cursor CLI: agent; reject it if help/version identifies Grok or lacks Cursor's required flags
• GitHub Copilot CLI: copilot
• Cockpit: cockpit

Cursor does not currently document cursor as its CLI executable. Because both Cursor and Grok can provide an agent command, identity verification is mandatory before invoking Cursor.