Router. You (the host CLI) hand ONE task to a peer agentic CLI when the peer does it
better, then report the result back. Load exactly one
reference/<target>/call.md - the one you routed to.
Rule zero - never call yourself
Whichever CLI is reading this is the host. Routing a task to the host is a no-op (you
would just do it yourself). Delegate to a DIFFERENT CLI when explicitly requested or when its capability is missing from the host. Inside Claude Code, never route to claude; inside Codex, never route to codex.
Trigger policy (conservative - external CLIs spend separate credits)
Fire on TWO conditions only:
- Explicit name - the user names a target tool.
- Capability gap - the user asks for something the host cannot do natively.
Generic review, planning, refactoring, debugging, and implementation stay with the host - it is presumed capable. When no peer was requested and the capability gap is unclear, keep the task with the host.
Route (match the request to ONE target, then load its call.md)
| Signal in the request | Target | Needs binary | Load |
|---|---|---|---|
"codex"; image generation; one-shot codex review |
codex | codex |
reference/codex/call.md |
| "agy" / "antigravity" / "gemini cli"; Google-grounded web search; image gen; science DB (gnomAD/UniProt/PubMed/ChEMBL) | agy | agy |
reference/agy/call.md |
"kiro" / "kiro-cli"; natural-language -> shell translate; MCP cross-register; AWS Bedrock peer opinion |
kiro | kiro-cli |
reference/kiro/call.md |
| "claude" / "claude code"; implementation; 1M-context planning; deep review | claude | claude |
reference/claude/call.md |
| "notebooklm" / "nblm"; RAG / audio / mind-map over a PDF/URL/YouTube corpus | notebooklm | notebooklm |
reference/notebooklm/call.md |
| "gpt-pro" / "chatgpt pro" / "gpt5 pro"; deep-reasoning second opinion packaged for ChatGPT Pro (web, flat-fee subscription) | gpt-pro | none (tar/pbcopy) |
reference/gpt-pro/call.md |
Tie-breakers: an explicit tool name always wins over a capability guess. Image
generation is offered by both codex (gpt-image-2, ChatGPT-auth, highest quality) and
agy - prefer codex when installed, fall back to agy.
Preflight (every route)
Each call.md defines its preflight. Run it before sending the task. Claude
implementation requires both auth and shell-capability preflights. If the host blocks
the target, report the normal-terminal fallback instead of retrying with weaker safety.
Reference map (load only what the routed target needs)
| Load | When |
|---|---|
reference/<target>/call.md |
you routed to (always) |
reference/agy/patterns.md, reference/codex/patterns.md |
need orchestration recipes (parallel fan-out, async jobs) |
reference/<target>/reference.md |
need the full flag table / gotchas (agy, codex, kiro, notebooklm) |
reference/agy/templates.md |
need copy-paste agy prompt scaffolds |
reference/<target>/scripts/*.sh |
the routed call.md tells you to run a wrapper script |
Done = target named; route-specific preflight passed; peer output reported, plus cost/session id when the target emits one; every generated file confirmed on disk. A blocked target ends with an actionable fallback, never a success claim.