feat(skills): add model support contribution workflow

.github/workflows/modelkit-ci.yml

@@ -36,6 +36,7 @@ jobs:
               tests/unit/core tests/unit/onnx tests/unit/cache
               tests/unit/utils tests/unit/test_helpers tests/unit/sysinfo tests/unit/inspect
               tests/unit/optracing tests/unit/serve tests/unit/packaging tests/regression tests/cli
+              skills/adding-model-support/tests
 
     name: test (${{ matrix.group }})
 

skills/adding-model-support/SKILL.md

@@ -0,0 +1,72 @@
+---
+name: adding-model-support
+description: >
+  Use this skill whenever contributing model support to the `winml` CLI: an unsupported
+  Hugging Face model or architecture, an existing family that fails in one or more commands,
+  a missing recipe, exporter, resolver, task, dataset adapter, evaluator, or promotion handoff.
+  It diagnoses the gap, implements the narrowest reusable fix, validates the result, opens a
+  draft model-support PR, and captures reusable findings. Trigger even when the user asks to
+  repair support rather than explicitly saying "add model support". Do not use it merely to
+  run or tune an already-supported model, optimize an existing recipe, or add a new execution
+  provider backend.
+compatibility: Requires a microsoft/winml-cli checkout, Python 3.11 with uv, GitHub CLI with push access, and an agent runtime that can delegate fresh subagents.
+---
+
+# Adding model support
+
+Use this skill to take one unsupported or partially supported model from current-main diagnosis through a validated draft contribution. Diagnose first, freeze the Effort/Goal/Outcome target from evidence, implement the narrowest reusable fix, validate independently, and capture durable knowledge.
+
+This is a dispatcher, not the complete procedure. Load the owning role contract under [agents/](./agents/) at every transition. Do not substitute this overview for those instructions.
+
+## Repository scope
+
+This skill lives in `skills/adding-model-support/` in [`microsoft/winml-cli`](https://github.com/microsoft/winml-cli). Resolve `src/...`, `tests/...`, and `examples/...` from the repository root. Keep methodology changes under `skills/adding-model-support/**` in a separate skill-only branch and PR; never mix them into a model-support PR.
+
+## Prerequisites and optional integrations
+
+- Run from a `microsoft/winml-cli` checkout with Python 3.11, `uv`, authenticated `gh`, and push access. The orchestrator verifies these before implementation.
+- The agent runtime must support fresh subagent delegation. Each role loads its file under `agents/`; if delegation is unavailable, stop as BLOCKED rather than collapsing producer, tester, and reviewer into one agent.
+- `model-breakdown` is optional. When discoverable as an installed skill, require schema `1.2+`; otherwise the planner builds the same profile from pinned source, config, and model-card evidence.
+- `auto-optimize` is optional and needed only for `--promotion-handoff`. Discover its `scripts/promotion.py` through the installed skill location. If it is absent or incompatible, promotion mode stops as BLOCKED while normal model-support work remains available. Follow the [orchestrator promotion handoff entry](./agents/orchestrator.md#promotion-handoff-entry).
+
+## Pipeline
+
+Run one model at a time. The orchestrator delegates every role to a fresh agent and drives all loop-backs to `APPROVE`, `REJECT`, or `BLOCKED`; `REQUEST_CHANGES` is not terminal.
+
+| Role | Owns | Contract |
+|---|---|---|
+| Orchestrator | preflight, delegation, loop continuation, isolation, cleanup | [orchestrator.md](./agents/orchestrator.md) |
+| Planner | current-main diagnosis, model profile, baseline, Effort/Goal/Outcome charter | [planner.md](./agents/planner.md) |
+| Producer | minimal code and exact-evidence recipe candidates | [producer.md](./agents/producer.md) |
+| Tester | repair loop and Goal-ladder verdicts | [tester.md](./agents/tester.md) |
+| Learner | model knowledge, methodology findings, blocker diligence | [learner.md](./agents/learner.md) |
+| Explainer | existing/new draft PR, body, label, replies, shipment | [explainer.md](./agents/explainer.md) |
+| Reviewer | independent final-head verification and one opinion comment | [reviewer.md](./agents/reviewer.md) |
+
+```text
+planner -> producer -> tester -> learner -> explainer -> reviewer
+   ^           ^                                      |
+   |           +---- artifact REQUEST_CHANGES --------+
+   +---------------- scope/tier contradiction --------+
+```
+
+## Non-negotiable boundaries
+
+- Planner freezes the target only after current-main baseline evidence. If reality contradicts it, return to planner.
+- Producer, tester, explainer, and reviewer remain separate agents. The reviewer starts from a fresh context after the draft PR exists.
+- Every model-support PR targets `main`, remains draft, and carries `model-scale-by-skill`. Dependency ancestry never changes the live PR base.
+- The reviewer writes only one normal PR comment containing `APPROVE`, `REQUEST_CHANGES`, or `REJECT`; these are not GitHub Review states.
+- Every required `(EP, device, precision)` tuple retains exact PASS, blocked, or exhausted-failure evidence. Never infer one tuple from another.
+- Run one bounded final-SHA FP32 CPU functional smoke Eval; do not present it as representative accuracy. Keep per-tuple performance and both component-level and op-level Analyze evidence.
+- Treat a recipe as authoritative during recipe validation. Semantic CLI overrides are diagnostic evidence, not proof that the recipe works.
+- Prefer narrow metadata/config/architecture-derived shared fixes. Preserve existing public defaults, signatures, recipe semantics, and sibling-model behavior with regression tests.
+- Never edit `examples/recipes/README.md` for recipes produced by this workflow.
+- Cleanup is orchestrator-owned, manifest-driven, dry-run-first, and limited to marked run-owned regenerable state via [cleanup-run-cache.ps1](./scripts/cleanup-run-cache.ps1).
+
+## Knowledge and evaluation resources
+
+- Read [model knowledge](./model_knowledge/) before planning and update it only with scoped evidence after testing.
+- Record reusable workflow lessons in [skill methodology findings](./skill_meta/); preserve historical provenance.
+- Use [response and trigger evals](./evals/) when changing this skill's behavior or description.
+
+Begin execution by loading the [orchestrator contract](./agents/orchestrator.md).