Setup delegation policy

Delegation pays off when it happens at the top of a unit of work, before any reasoning has bent the decision. A skill cannot hold that ground: it fires only when something reminds the agent to reach for it, and by then the reasoning has already started. A standing rule in an always-loaded steering file is carried on every turn instead, which is why the rule is installed rather than invoked.

policy.md is that rule, written in tiers with model names only as examples, so it holds on any harness. The entire file is the payload, copied verbatim starting at its ## Delegation and model routing heading.

No harness is assumed. For each one, find its steering file, its delegation mechanism, and its models from its own documentation and tools, because locations and schemas change between releases.

Steps

1. Choose the harnesses and the action

Run python3 -I scripts/detect-harnesses.py (add --json for structured output). It lists the candidates found on disk, with no harness named in advance: each line gives the directory, the command on the path, the steering files, whether they already carry the delegation section (policy=), and the tier agents present. Directories holding only installed skills come back as one skills-only line. Add the harness you are running in if it is missing, show the user the list, and ask them to confirm it and add any harness it missed.

For each harness, report its state from the script and ask what to do, with a recommendation: