Feature matrix

Verified against local CLIs (approximate versions at packaging time): agy 1.x, grok 0.2.x, kimi 0.29.x, qwen 0.20.x, claude 2.1.x, codex 0.145.x. Re-check with <cli> --help — flags drift.

Unified wrapper flags

Flag	Meaning
-m	Model id
-t	Wall / print timeout
-C	Working directory (absolutized by wrappers)
-f	Prompt file (preferred for long bundles)
-o	Output format (where supported)
-r	Read-only / plan / sandbox
-T	Native multi-agent / ultracode preamble
--no-git	Prepend no git-write guard
--	Passthrough remaining args to the underlying CLI

Missing flag values print error: <flag> requires a value (exit 2) instead of unbound variable.

Per-CLI capabilities

Capability	agy	grok	claude	codex	kimi	qwen
Headless print / exec	-p + PTY	-p / --prompt-file	-p	codex exec	-p	-p
Skip permissions	--dangerously-skip-permissions	bypassPermissions	--dangerously-skip-permissions (cleared by -r/-P)	sandbox + approvals flags	auto on -p	auto / --sandbox
Read-only mode	--mode plan + sandbox	--permission-mode plan	--permission-mode plan	-s read-only	soft prompt guard	--sandbox + guard
Worktree isolation	(manual)	--worktree	--worktree	(manual / review helpers)	(manual)	qwen review worktrees
Custom agents	--agent, .agents/agents/	--agent, --agents JSON	--agents JSON, claude agents	~/.codex/agents/*.toml	--agent, --agent-file	.qwen/agents/, /agents
Product “team” mode	/teamwork-preview (Ultra)	subagents (spawn_subagent)	ultracode / workflows	subagents + spawn_agents_on_csv	sub-agent profiles	Agent tool + review pipeline
Structured output	plain text only (no json flag in older builds)	plain/json/streaming-json, --json-schema	text/json/stream-json, --json-schema	--json, -o last message	stream-json (wrapper cleans)	-o json (wrapper extracts)
Cross-CLI fan-out	via multi-cli-spawn	same	same	same	same	same

Native multi-agent detail

agy

• Slash: /teamwork-preview (preview, Ultra), /agents
• CLI: --agent, --mode, --effort, --sandbox, --add-dir, --print-timeout
• Custom agents: markdown under .agents/agents/
• PTY required for -p — always use agy-exec.sh (gemini-cli#76)

claude

• --effort ultracode or interactive /effort ultracode → auto dynamic workflows
• Prompt keywords: ultracode, “create a workflow”
• --agents JSON, background claude agents, agent teams (see official docs)
• --worktree, --forward-subagent-text (stream-json)

codex

• Subagents spawn + synthesize; CSV batch via spawn_agents_on_csv
• codex features list|enable|disable
• Agent TOML under ~/.codex/agents/ or .codex/agents/

grok

• spawn_subagent with types general-purpose / explore / plan
• --agents JSON, --no-subagents, capability modes, worktree isolation
• grok agent stdio/headless/serve/leader for non-TUI hosting

kimi

• Agent Markdown discovery; --agent / --agent-file
• Built-in sub-agents; ACP server (kimi acp)

qwen

• /agents create/manage; Agent tool parallel launches
• qwen review multi-step PR review harness (worktree, chunking, coverage)

Cross-CLI: multi-cli-spawn

See skills/multi-cli-spawn/SKILL.md.

Flag / artifact	Meaning
--outdir / -O	Required output directory
--seat name[:mods][@label]	One parallel provider seat
-f / prompt args	Shared brief
-t / -C / -r	Timeout, cwd, all-seats readonly
--no-wait	Return after spawn; poll .status yourself
<key>.status	RUNNING → DONE rc=0 | BLOCKED rc=N
spawn.status	Parent RUNNING → DONE (or SPAWNED if --no-wait)
CLI_AGENT_RESULT	PASS / BLOCKED per seat after wait mode

Waiting: prefer foreground wait. Do not treat missing *.md mid-run as failure — answers are written when each seat finishes.