[Rakvadium]

Why

Agents had to combine three sources before their first spawn: the 50-line rules header in models.txt, the Models table in harness.md, and a 6-sentence paragraph pasted into seven skills. Each skill also hard-coded its own fallback models, so changing the map left stale copies behind. Slugs such as grok-4.7-high-fast had to be decoded into a model and options from one example.

What changed

• models.txt is data only. Each entry is <provider> <model> <option>=<value>, using the ids from T3's orchestrator_capabilities, or inherit. Panel roles separate entries with |. The model choices are the same as before.
• harness.md Models section is the single routing source. It maps an entry to delegate_task (T3), Agent (Claude Code), or Task (Cursor), and has five rules: always route through T3 on the entry's provider, unroutable Claude seats stay unspawned, unroutable Cursor seats run on the session model, same-family fallback on rejection, and a missing role line stops the skill.
• Skills use one sentence that points to harness.md, in place of the repeated paragraph. Per-skill default models are gone (arena, architect, how, why, interrogate, reflect, swarm, playbooks).
• The ~/.cursor/rules/oddstack-models.mdc override is removed, along with the auto/inherit-parent aliases (now inherit).
• /setup-oddstack writes the new format and reads ids from orchestrator_capabilities in T3.
• New CI check, scripts/check-models.mjs. It parses every entry and fails when a skill reads a role the map has no line for.

Verification

• node scripts/check-models.mjs: 17 roles parse, and 20 role reads in skills resolve. Renaming how explorer in the map makes it fail on skills/how/SKILL.md.
• node scripts/check-skill-names.mjs passes.
• A T3 test run of the new routing is noted in a comment below.

[Rakvadium]

T3 test run. I launched a Claude Opus lead in this worktree (C:\wt\odium-stack\models-map, slug 27 characters) and had it spawn the arena runners panel using only models.txt and harness.md.

• Targets. {providerInstanceId: "claudeAgent", model: "claude-opus-5-5", options: {effort: "high"}} and {providerInstanceId: "cursor", model: "grok-4.7", options: {reasoning_effort: "high", fastMode: true}}. Both match the entries exactly.
• Results. Both children replied with the right commit. Both reached completed on their own, and the Grok child did so after a single run.
• Fixes. The lead reported 6 unclear points, and 6ac8833 fixes them:
  • The done marker's file name.
  • Where a shared child runs.
  • The overlap between REPORT-END and a terminal state in the wait loop.
  • When the cursor workaround is retired: when the SDK fix ships, not after one success on a short path.
  • Default options.
  • The delegate_task title.