Multi-CLI Runtime

Use this skill only inside multi:* forwarding subagents (codex-execute, cursor-delegate, cursor-research, cursor-explore, antigravity-researcher, antigravity-explorer, opencode-delegate, opencode-researcher, opencode-explorer, etc.).

Primary helper

node "${CLAUDE_PLUGIN_ROOT}/scripts/multi-cli-companion.mjs" task --cli <cli> --role <role> [flags] --prompt "<text>"

Where <cli> is one of codex|cursor|antigravity|opencode (or any CLI added via the multi-cli-anything skill) and <role> is the subagent's logical role. Cursor uses delegate (write/agent), research (read-only web), and explore (read-only codebase); Codex uses execute; OpenCode uses delegate (write), research (read-only web), and explore (read-only codebase); other roles in use include writer, debugger, researcher, reviewer, explorer, ask.

Execution rules

• Each multi:* subagent is a forwarder, not an orchestrator. Its only job is to invoke the companion once and return that stdout unchanged.
• Prefer the helper over hand-rolled git, direct CLI strings (cursor agent acp, codex exec), or any other Bash activity.
• Use exactly one Bash call per subagent dispatch. Do not chain cat, sleep, polling loops, or follow-up node calls.

Routing flag handling

Treat these as runtime controls — strip them from the task text before forwarding, then re-add them as flags on the companion call:

• --background / --wait — --wait/foreground is the default and is what you should run: the companion blocks until the CLI finishes, so your Bash call returns the real result. For long-running work, the PARENT command schedules background execution by running this subagent as a harness background task (which notifies the main thread on completion/failure) — NOT by passing --background. The companion's --background detaches a worker the harness can't see (no notification) and is only for explicit user-requested fire-and-forget polled via /multi:status.
• --model <name> — pass through verbatim. Leave unset unless the user explicitly asked for a model. (Antigravity ignores --model: its headless agy -p path is fixed to Gemini 3.5 Flash.)
• --effort <level> — only Codex accepts this (none|minimal|low|medium|high|xhigh). Other adapters ignore it. Pass through verbatim if present.
• --resume — translate to --resume-last.
• --fresh — do not add --resume-last, even if the user's text sounds like a follow-up.
• --write — default to --write for execute/delegate/writer/debugger/reviewer roles (these need to edit files); for read-only roles (research/explore/planner/researcher/explorer/ask) pass --read-only instead (it forces write off even if --write is also present). Honor explicit user override either way.
• --until-done — Codex, Cursor, and OpenCode. Tells the companion to loop resume turns on the same session until the model emits PLAN COMPLETE, hits a hard error, runs out of turns, or stops making progress. Pass through verbatim when the user opts in. Default off — only set when the user explicitly asked for autonomous run-until-done behavior. Antigravity rejects this flag. Note: OpenCode does NOT support --effort — pass it only to Codex; OpenCode ignores it.
• OpenCode role note: OpenCode has no --read-only flag. For read-only roles (research, explore) the adapter enforces read-only by injecting a custom oc-* primary agent via OPENCODE_CONFIG_CONTENT with write/edit/bash denied, plus an OPENCODE_PERMISSION deny floor. Passing --read-only to the companion is still correct for the forwarder; the adapter handles the enforcement.
• --max-turns <N> — Codex and Cursor. Sets the autonomous-mode turn ceiling (default 30). Requires --until-done. Pass through verbatim if present.
• --plan <path> and --prompt-file <path> — both load the prompt body from a file. --plan is the user-facing alias; the companion's actual flag is --prompt-file. Translate --plan to --prompt-file on the Bash call. When either flag is present:
  1. The file's bytes ARE the prompt — do NOT paste a framing block in front of them. Skip the role-specific preamble entirely.
  2. Any positional task text the user provided is treated as an addendum and gets appended after the file content as a separate paragraph (the companion handles this via positional args after --prompt-file).
  3. The path resolves relative to the companion's --cwd if set, else CWD. Absolute paths always work. See multi-plan-handoff skill for when Claude (the parent thread) should auto-add --plan based on conversation context.

Capturing diagnostics

Always append 2>&1 to the Bash call so the parent thread can see runtime diagnostics if the companion fails.

Safety rules

• Preserve the user's task text as-is apart from stripping routing flags and prepending the subagent's role-specific framing block.
• Do not inspect the repository, read files, grep, monitor progress, poll status, fetch results, cancel jobs, summarize output, or do any follow-up work of your own.
• Return the stdout of the task command exactly as-is.
• If the Bash call fails or the upstream CLI cannot be invoked, return a single short failure line: <CLI> <role> failed: <one-line reason from stderr or "no output">. Never silently return nothing — the parent thread needs to know the run failed.
• On failure, the failure line is your ENTIRE response. Do NOT fall back to performing the task yourself with Bash, other tools, or your own knowledge — even when the task looks trivial (listing a directory, answering a question). A substituted answer is a contract violation: it hides the CLI outage from the caller, who selected the external CLI deliberately and must be told it failed.
• If the user asks for setup, status, cancel, or result, that is NOT a multi:* subagent dispatch — it is a /multi:* slash command the user runs directly. Do not call those subcommands from within a forwarding subagent.

Forbidden behaviors

• Do NOT paraphrase or rewrite the companion output, even if it looks like a status update or progress message.
• Do NOT add narration like "The task is running in the background", "I will be notified when it completes", or "The companion is handling all steps". The companion already prints whatever the user needs to see.
• Do NOT promise to deliver later results yourself, and do NOT narrate "I'll be notified when it completes." Run the companion in the foreground and return its real result when the Bash call returns. (The parent command may run you as a harness background task; that mechanism — not your narration — re-wakes the main thread on completion or failure. If the user explicitly forced a detached --background run, they poll /multi:status themselves.)
• Do NOT invent fabricated output if Bash returned empty or non-zero. Use the failure line format above.