version: "1.0.0" description: Detect which delegation harnesses this machine has, ask which harnesses and models to allow, and write the allowlist. Run once per machine, and again after you install or remove a CLI. disable-model-invocation: true allowed-tools: Bash, AskUserQuestion, Read

Build the delegation allowlist

The allowlist is the user's policy: the ceiling of what /delegate:route may ever pick. Detection is the floor. Do not decide the policy for them. Ask.

Model rosters are large. opencode lists 959 models, devin 392, cursor-agent 227. So the harness's own list command is the source of truth, and the catalog's suggested array is only a seed. Never present a picker with hundreds of options.

1. Detect

node "${CLAUDE_PLUGIN_ROOT}/scripts/delegate.mjs" detect

For every harness this gives available, always_present, traits, its list command, and a short suggested array. It also reports typesafe_key and where the allowlist will be written.

2. Ask which harnesses, with AskUserQuestion

One call, two questions. Both multiSelect: true.

• Q1, outside CLIs. Only those whose available is true. A picker holds four

options. If more than four are available, ask the first four, then make a second call for the rest. Use each harness's traits, shortened, as the option description.

• Q2, sub-agent models. The three in subagent.suggested: haiku, sonnet, opus.

This is a small fixed set, so a picker is the right tool here.

inline is always included. Do not ask about it. Without it the router can never answer "keep this one yourself".

Say in your summary which harnesses you skipped, and that the CLI is missing.

Say `verified` out loud for every option you offer. local means the launch line was confirmed first-hand. docs means it comes from documentation and has never run on any machine here, so its first run is also its first test. The user deserves that before they allowlist it, because the failure lands in their terminal.

3. Choose models inside each outside CLI, in conversation

Do not use AskUserQuestion for this. Four options cannot cover 959 models.

For each outside CLI the user enabled, run its list command and read the real ids:

opencode models | head -40
devin models list
cursor-agent models

codex has no list command. Its model is set by policy; use suggested.

Then propose at most eight ids for that harness, in a short table with one line each. Tell the user they can name any other id from the list instead. Take the whole list into account, not only the head of it; grep it when you are after a family.

4. Write the fits line for every chosen model

fits is the entire rubric the router judges that option against, so it decides the routing quality. Rules:

• Say what the model should get and what it must not get. A line with only

praise makes the option win everything.

• Write plain judgement, not marketing, and keep it to one or two sentences.
• Do not describe the harness. The harness's own traits is added at route time, so

repeating it wastes the rubric.

• Where the id matches a suggested entry, copy its fits.
• If you do not know the model, say so and ask the user what it is good at. Do

not invent a description. A wrong rubric sends real work to the wrong model.

• Do not propose a Fable-class id. It costs far more than the delegated work is worth.

5. Per-model launch config

Some models need their own command. An opencode model routed through a named provider is the common case:

OPENCODE_CONFIG_CONTENT='{"provider":{"openrouter":{"models":{"deepseek/deepseek-v4.1-flash":{"options":{"provider":{"order":["baseten/fp8","novita/fp8"],"allow_fallbacks":true}}}}}}}' opencode run -m {model} --format json "<prompt>" < /dev/null > events.jsonl 2> err.log

Put that whole string in the entry's optional launch field. It replaces the harness template for that entry only, and {model} is still substituted. Ask the user for it; do not invent a provider order.

Leave launch out when the harness default is right, which is most of the time.

6. Save

Pass one JSON array of entries. Each needs harness, model, and fits; launch is optional.

node "${CLAUDE_PLUGIN_ROOT}/scripts/delegate.mjs" save '[
  {"harness":"inline","model":"this-session","fits":"Short work, or work needing context only this session holds."},
  {"harness":"subagent","model":"sonnet","fits":"A broad search across a repository. Not for deciding whether a contested finding is real."},
  {"harness":"codex","model":"gpt-6-astra","fits":"Adversarial review of code it did not write. It cannot implement anything."}
]'

It rejects an unknown harness, an empty model, and a missing fits, so a mistake fails loudly.

7. Report

1. The allowlist path, and that it is plain JSON they can edit.
2. Which harnesses you skipped, and why.
3. If typesafe_key is false: TYPESAFE_API_KEY is unset, so route returns the

roster and the model chooses by reading rather than by probability. Setting the key turns the judgement on. Get one at https://typesafe.ai.

Adding a harness the catalog does not ship

The overlay at ~/.claude/delegate-catalog.json survives a plugin update. Give it the same shape as ${CLAUDE_PLUGIN_ROOT}/catalog.json. A new harness id needs bin, traits, launch, and list.

{ "harnesses": [ { "id": "aider", "bin": "aider", "list": "aider --list-models",
  "traits": "An outside CLI that edits files in place.",
  "launch": "aider --model {model} --message \"<prompt>\"", "models": [] } ] }

To add a model, you do not need the overlay. Rerun this skill, or edit the allowlist directly.