Cross ai

cross-ai

Delegate the current task to an external CLI coding agent over bash, then relay its answer. You are the dispatcher: pick the command from the table, feed it the context, run it, report back. Nothing else.

This skill is explicit-invoke only (/cross-ai). It is NOT triggered by phrases like "use agy" on their own, and it is NOT the native Task tool — "launch a sub-agent" means the Task tool, not this.

Invocation

/cross-ai <agent> [task]

• <agent> is one of: agy, gemini, codex, cursor, claude (Claude Code headless).
• [task] is the work. If omitted, use the surrounding request as the task.
• Multiple agents in one request → run each on the same task, in parallel, and present results side by side labeled by agent. Two separate /cross-ai agy and /cross-ai codex invocations against the same ask are the same thing: launch both.

Dispatch table (read-only defaults)

Run these verbatim, substituting <TASK> and <FILE>. Each is read-only — no edits, no auto-approve of writes.

Agent	Command
agy	agy -p "<TASK> @<FILE>" --model gemini-3.1-pro-high --dangerously-skip-permissions < /dev/null 2>&1
gemini	gemini -m gemini-3.1-pro-preview -p "<TASK> @<FILE>" --output-format json | jq -r '.response'
codex	codex exec -m gpt-5.5 --skip-git-repo-check "<TASK>" 2>/dev/null
cursor	cursor-agent -p "<TASK>" --model gpt-5.5-high --trust --output-format json | jq -r '.result'
claude	claude -p "<TASK>" --model claude-opus-4-8 --output-format json --allowedTools "Read,Grep,Glob,Bash" | jq -r '.result'

@<FILE> injection is for agy/gemini. codex/cursor/claude read the current working directory themselves — reference paths in <TASK> and they will open them.

agy truncation guard: if its output looks cut off or garbled (it misbehaves on a non-TTY pipe), rerun under a pty and strip ANSI. script's flags differ by OS — claudex runs on both:

• macOS (BSD script): script -q /dev/null agy -p "<TASK> @<FILE>" --model gemini-3.1-pro-high --dangerously-skip-permissions
• Linux (GNU script): script -qec 'agy -p "<TASK> @<FILE>" --model gemini-3.1-pro-high --dangerously-skip-permissions' /dev/null

Pipe either through perl -pe 's/\e\[[0-9;]*[A-Za-z]//g' to strip color codes (portable across macOS/Linux).

cursor workspace trust: cursor-agent refuses to run in an untrusted directory (it blocks on a "Workspace Trust Required" prompt instead of answering). The --trust in its command grants directory trust so the headless run proceeds; it is NOT --force/--yolo and does not auto-approve edits or command execution, so the read-only posture holds.

Execution recipe

1. Stage context the agent can't see on its own. It starts blank — it does not inherit this conversation. If the task needs our analysis, a specific diff, or files outside its cwd, write that to a temp file first: CTX=$(mktemp); e.g. git diff > "$CTX" or gh pr diff > "$CTX" for "review current PR". If the agent can gather it itself in the cwd (codex/cursor/claude can run git diff), skip this.
2. Build the command from the table. Inject the staged file via @$CTX (agy/gemini) or name $CTX/paths in <TASK> (codex/cursor/claude).
3. Run it with Bash. For multiple agents, launch them concurrently and wait for all.
4. Relay each agent's raw output, labeled by agent. Don't compress or editorialize unless asked — the point is what the other model said.

Model override

Default to the flagship in the table. If the user names a model after the agent (/cross-ai cursor opus, /cross-ai cursor with grok, /cross-ai codex gpt-5.4), map it and pass it to that agent's model flag:

Agent	Flag	Flagship (default)	Other selectable
agy	--model	gemini-3.1-pro-high	gemini-3.5-flash-low, claude-opus-4-6-thinking, gpt-oss-120b-medium
gemini	-m	gemini-3.1-pro-preview	gemini-3-pro-preview, gemini-2.5-pro, gemini-3-flash
codex	-m	gpt-5.5	gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark
cursor	--model	gpt-5.5-high	gpt-5.5-high-fast, claude-opus-4-8-thinking-high, composer-2.5, cursor-grok-4.5-high, gpt-5.3-codex
claude	--model	claude-opus-4-8	claude-sonnet-4-6, claude-opus-4-8[1m]

Rules of engagement

• Just run it. Assume the binary is installed and authenticated. Do NOT which, version-check, --help, or test-run first. Build the command and execute in one step.
• Read-only by default. Never pass write/auto-approve flags (--force, --yolo, --sandbox workspace-write, --permission-mode acceptEdits) unless the user explicitly asks the external agent to modify files.
• Stale model id is the only thing to check. If a run fails because a model id was rejected, that is the one time to look: run the tool's list (agy models, gemini --list-models, cursor-agent --list-models, or check --help), pick the current flagship, and rerun. Model ids drift; the table is a snapshot. cursor-agent --list-models prints real ids. But agy models prints display names, not ids — the id is a lowercased/hyphenated form with a tier suffix (Gemini 3.1 Pro (High) → gemini-3.1-pro-high), the mapping is inconsistent (not every displayed tier resolves, and the dotted vs dashed version separator varies by family), so confirm a specific agy id with a one-line agy -p "OK" --model <id> --dangerously-skip-permissions run before trusting it.
• Auth failure → report and stop. If a run fails with an auth/login error, surface it verbatim and stop. Don't try to fix the other tool's credentials.
• Not for Claude sub-agents. claude here is the headless CLI process. The native Task tool ("launch a sub-agent") is unrelated and not dispatched through this skill.