model-routing-preset-builder

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Native Routing Policy Builder

Read the installed `model-capability-router` as the single baseline. It
recommends GPT-6.1 Sol High for coordination, GPT-6 Astra Medium with XHigh
reserved for exceptional reasoning, GPT-6.1 Terra for visual and application
work using the router's workload-specific efforts (High as the starting point
for professional tool work), and GPT-6 Luna High/XHigh/Max. These are
policy recommendations; they do not establish or change the actual home
configuration. Preserve an explicit native effort selection, including Sol
Low; High is the shared recommendation, not a forced configuration change.
Do not revive retired schema-3 presets, mandatory coordinator topology or
resolver scripts.

For pstack role-aware choices, the router's
`references/pstack-role-presets.json` is the only preset mapping. Change that
source only when the operator asks for a durable shared-policy change. Keep a
selected role-aware preset separate from pstack's global budget options; do not
apply one global effort over its per-role recommendations. Equilibrado is the
default only for a new setup with no existing role choices.

1. Read the operator's requested changes and current native model/effort
   availability for each relevant channel. Preserve explicit selections.
2. Change only the requested profile choices or task boundaries. Keep task
   ownership, permissions, proof and compute selection separate.
3. For this operator, use the recommended palette above when the exact native
   channel supports it. Luna Max is available when justified, not an Ultra mode
   or a new paid route. Terra capability and effort/cost/latency evidence is in
   the router's `references/gpt61-terra-20261028.md`; the Sol launch evidence is
   in `references/gpt61-sol-20260929.md`; historical GPT-6
   evidence remains in `references/gpt6-20260922.md`. Token prices and benchmark
   costs do not prove lower cost per completed Codex outcome.
4. Author changes in the normal TheAngrySkills source repository when requested
   as a durable shared policy. Do not create a competing Workbench install.
5. Validate identifiers and effort support against the current channel; check
   that the policy has no mandatory ladder, task creation or selected-main
   override, and adds no speculative cache retention, keepalive prompts or
   polling loops. Report unavailable options rather than inventing support.
6. Install published changes with `npx skills` from the normal source and verify
   canonical paths and lockfile provenance for each authorised host/profile.

Return the chosen profiles, actual source/version, validation evidence and
remaining gaps. Do not add topology machinery or modify native permissions.

From TheAngryPit/TheAngrySkills, MIT.

More from TheAngryPit/TheAngrySkills