tools

Tools

Three scripts, no environment to create. Each carries a PEP 723 header, so uv run resolves its dependencies on the fly.

Script	What it does
render.py	substitutes profile tokens, composes fragments, enforces every platform limit. --profile selects the branding, --check is the CI drift gate, --package builds the zips, --with-embedded-knowledge emits the optional capability
lint.py	validates a rendered tree against the vendored schema, then the semantics the schema cannot express, then brand leakage for a non-default profile
make_icons.py	draws the colour and outline icons from source in the profile's brand colour, with no image library

uv run tools/render.py --check     # or: just validate
uv run tools/lint.py
uv run tools/make_icons.py

What each gate actually catches

The renderer fails the build on: instruction budget overflow (never truncation), an em or en dash in a composed instruction set, a missing or empty fragment, a duplicate capability, a WebSearch URL with more than two path segments or a query string, an unsupported or oversized embedded file, any string over its documented limit, an unresolved {{placeholder}}, a profile with a missing token or a non-HTTPS publisher URL, and two agents sharing an app id (checked across every agent even when you render only one, since a collision would otherwise stay hidden).

The linter fails on: schema violations, a missing instructions field, an unrecognised root property, a dangling icon or embedded file reference, and app manifest fields over their limits. It warns when instructions pass 90 percent of the budget, and when an agent has no conversation starters.

The last two exist because the published schema does not enforce them even though the platform does. See ../docs/platform-notes.md.