[creedants]

On T3 Code 0.1.4, orchestrator_capabilities returns a short list. Outside Codex, Claude, Muse, and Grok it lists at most eight models per provider and counts the rest in omittedModels, and an optional search string finds them. setup-pstack treated the short list as the whole catalog, so roles.py write refused a real model the list left out, and a note naming a configured model the list omitted read as if T3 had dropped it.

• roles.py catalog merge --catalog <file> --from <file|-> merges one search result into the saved catalog. Models merge by provider and model id. A model the saved file already has keeps its saved entry and options. A provider that brings no models is skipped. A result that brings a model for a provider the saved catalog lacks or cannot run, or that brings no model at all, exits 2 and writes nothing. A merge that adds nothing writes nothing, so a repeated merge is byte-identical.
• roles.py show adds a search list to a role entry when a configured seat's provider is runnable, omits models, and lacks that seat's model.
• Setup steps 1, 2, and 3(b) say the result is a short list, search for a missing or named model and merge the result before calling a seat dropped or proposing it, strip the provider instance id from a search entry to get the model id, and treat an empty search as no proof the model is unavailable.
• A host whose orchestrator_capabilities takes no search keeps using the saved catalog as before, and show prints no search list on a catalog without omittedModels.

Measured on this host: one unfiltered call listed 33 of 157 models across the runnable providers, and the OpenCode provider listed 9 of 71. The counts move as provider catalogs refresh. After a search and merge, write accepted opencode-go/kimi-k3.1-pro, which it refused before.

Proved by: eight new tests in tests/pstack_t3/test_catalog_merge.py, each caught alone by its own mutation in review; the roles suite (344 tests); the full suite at the worker's head and again after the review edits, on Python 3 and 3.12; scripts/build.py and scripts/check.py clean; and the reviewer's live run of search, merge, write, and snapshot against the real T3 catalog.

Coordinated by claudeAgent/claude-opus-5-5. Written by opencode-go/deepseek-v4.1-flash and opencode-go/muse-spark-1.3-contributor. Reviewed by codex/gpt-6.1-sol.