delegate

Subtask — Delegate

Delegate a task to the best available AI model. The routing strategy is auto-loaded at session start — it tells the agent when and where to delegate.

External CLIs: codex (OpenAI) · codex review (separate quota) · gemini (Google) · agent (Cursor, manual-only)

All CLI flags MUST come from cached --help output. Never hardcode flags.

Usage

/subtask:delegate <task description>

Model hints can appear anywhere in the input:

• Research caching best practices in Firebase — auto-routes
• Review the auth changes using codex — Codex
• Use gemini to analyze performance — Gemini
• with codex, implement retry logic — Codex
• Create a Linear ticket for the billing bug — Claude (needs MCP)
• Analyze this with grok — Agent (Grok only available there)

Sandbox Rules

Apply the least access that satisfies the task. Derive flags from cached --help.

Level	Access	Use When
1. Pipe	No filesystem	Self-contained: summarize, parse, generate from provided context
2. Read-only	Read, no writes	Reviews, analysis, advisory needing codebase context
3. Working dir	Read + scoped writes	Implementation within a worktree or temp sandbox
4. Full access	Read + write beyond working dir	Needs config outside worktree (still sandboxed)

⛔ NEVER use elevated-permission flags. This is a hard constraint, not a guideline. ⛔

If a task cannot be completed within a sandbox, DO NOT delegate it. Fall back to Claude (native subagent) or abandon the delegation entirely. An unsandboxed external CLI invocation is never acceptable — not for any task, any provider, any reason.

CLI	Banned flags
codex	--dangerously-bypass-approvals-and-sandbox
gemini	-y / --yolo, --approval-mode yolo
agent	-f / --force, --sandbox disabled

This covers any flag — present or future — that bypasses permission prompts, auto-approves all actions, or disables the sandbox.

If the ideal sandbox level fails on a CLI, try a more relaxed sandbox on the same CLI if safe. If no sandboxed invocation works, do not delegate — use Claude (native subagent) instead.

Execution

All steps are silent — only the final result is shown to the user.

1. Parse Input

Extract task description and model hint (if any). Match hints case-insensitively: using <model>, use <model>, with <model>. Model names can include variants (e.g., "Codex 5.3 High"). Specific model names (e.g., "grok", "gpt-5.2") resolve by family (OpenAI → codex, Google → gemini). Models without a native CLI → agent if supported.

2. CLI Discovery

1. Run <cli> --help for the selected CLI and cache output
2. If routing selected agent, also run agent --list-models and cache

3. Route

Model hint present → use that provider unconditionally.

No model hint → classify the task and route per the Quota-pressure routing table (loaded at session start), which defines fallback chains and sandbox levels per task type. Tasks requiring MCP tools (Linear, Slack, Notion, etc.) stay on Claude (native subagent).

4. Persona (Optional)

If the task maps to an agent role in .claude/agents/*.md, read the matching definition and compose it into the prompt. Otherwise skip.

5. Launch

Native subagent (routed to Claude): spawn with persona + task description.

External CLI: derive flags for non-interactive mode, prompt, and sandbox level from cached --help. Additional rules:

• Request summarized output for research/analysis/review tasks (not implementation)
• Use text output format (not JSON/stream-JSON)
• Always launch in background (run_in_background: true)
• Working directory: relevant worktree path, or monorepo root

6. Monitor & Handle

Poll output silently. Kill the process if fatal patterns appear with no new output for ~10s:

• denied by policy, permission error, permission denied
• quota exhausted, exhausted your capacity
• error executing tool, crash, panic, fatal

If output stalls without error patterns, ask the user before killing — it may be legitimately slow.

On failure:

• Quota-related → skip to next provider immediately
• Sandbox-related → try relaxed sandbox if safe, otherwise next provider
• Transient (timeout, rate limit) → retry once, then treat as hard failure
• User-specified model → show error, don't fall back
• Auto-routed → fall back silently through priority chain

Output Format

subtask model: codex · reason: code review · sandbox: read-only

[model output here — fenced for external CLI, plain markdown for native subagent]

On failure, add a status line: **status:** error — <reason>.