Wrapper tier enforcement
The parallelism section reports generated wrapper defaults for every configured provider.
Wrapper tier enforcement
The parallelism section reports generated wrapper defaults for every configured provider. This is an offline artifact contract, not an observation of a running worker. Provider or account policy can substitute a model.
| Provider | Default model enforcement | Native effort on wrappers |
|---|---|---|
| Claude | pinned catalog model | `effort` |
| Codex | pinned catalog model | `model_reasoning_effort` |
| Factory | pinned catalog model | `reasoningEffort` |
| Copilot, Cursor, OpenCode | pinned catalog model | not enforced |
| OMP | inherited until exact role models are configured | `thinkingLevel` with a model pin |
| Antigravity | inherited; wrapper mirrors do not project CLI flags | not enforced |
| OpenClaw | inherited; Markdown mirrors do not set `subagents.model` | not enforced |
| OpenHuman | inherited/dynamic; user wrappers use hints, project wrappers omit model | not enforced |
| Windsurf | inherited from the UI session | not enforced |
| Pi, DeepSeek Harness | inherited; agent references do not set runtime controls | not enforced |
| Warp | global-only, run/profile model selection | not enforced |
| Hermes | global-only, delegation configuration | not enforced |
| Grokbot | global-only; no verified AIWG-owned model setting | not enforced |
| Muse | global-only, session `--model` or `settings.json` `model` | not enforced |
| Grok Build | inherited; tier enforcement gated by #2577 | not enforced |
The native wrapper fields above are distinct from headless CLI invocation flags. A provider can support explicit CLI models without consuming model metadata in a deployed wrapper. AIWG does not change global provider settings or environment variables to fill these gaps.
Claude orchestrators should select an AIWG wrapper or supply an explicit invocation model. Unpinned built-in `general-purpose` agents can inherit a flagship session. `CLAUDE_CODE_SUBAGENT_MODEL` supplies a fallback; `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1` can override even explicit pins. Inspect the resolved model when such environment or account policy is active.
All five pin-capable wrappers use catalog tiers. Cursor's `auto` economy selector falls back to the concrete standard tier for both efficiency wrappers. OMP's `configured/*` catalog entries are placeholders, so its default remains reported as inherited; exact operator role overrides enable native pins. Unsupported model/effort fields are not invented. Grok's configured selectors remain available, but do not establish enforced catalog tiers under #2577.
Explicit `model-override`, role CLI/config pins, and exact source `model` IDs are kept. Bare `opus`/`sonnet`/`haiku` values remain role intent. Existing flagship election safeguards still reject unelected flagship overrides.
Reproduction before changes: existing wrapper/Codex tests passed (33 tests), and new regression fixtures exposed wrapper-specific failures. Ordinary Codex unknown pins were preserved already; canonical wrappers overwrote source-level pins. Claude and OMP omitted effort, and Cursor's efficiency worker dropped `auto` without a concrete fallback. The fixtures use translators and temporary deployments, never paid calls. No Claude/Codex live dispatch was performed.