Add add-inference-model Claude skill

[hansent]

Summary

• instruct claude that this is a living skill / should update the skill itself when learnign new things while adding a model
• First draft of .claude/skills/add-inference-model/SKILL.md — an end-to-end playbook Claude invokes when the user asks to add a new core/pre-trained model.
• Four surfaces a new model can touch: inference_models class + registry entry (always), weight zips + registration script via roboflow/model-registry-sdk (always), workflow block (optional), inference/core/models/inference_models_adapters.py (only if plain /infer endpoint serving is needed).
• Explicit note that inference/models/<family>/ is deprecated — new core models always go in inference_models/.
• Tells the skill to survey same-(backend, task) siblings in inference_models/.../models/ before scaffolding rather than copying any one model's idioms.

Written right after adding SAM2 Video (PRs #2245 + roboflow/model-registry-sdk#5) so the gotchas section captures real surprises from that work. Framed as a living document — future model additions should extend the file, not rewrite it.

Test plan

• Trigger the skill on a new model-adding prompt and confirm it surveys inference_models/ before scaffolding
• On the next model ship, confirm the gotchas / step notes hold and add anything new that surfaces

[Copilot]

Pull request overview

Adds a new Claude “skill” document that serves as an end-to-end playbook for adding new core/pre-trained models, including required integration points (model class/registry, weight zips + model-registry registration) and optional wiring (workflow blocks, /infer adapters).

Changes:

• Introduces .claude/skills/add-inference-model/SKILL.md with step-by-step guidance for implementing, registering, testing, and packaging new core models
• Documents the “four surfaces” a new model can touch, plus a verification checklist and “gotchas” captured from recent model work
• Includes operational notes for staging/prod registration and local execution pitfalls

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This section says there is a “path-sensitive conftest.py” under inference_models/, but there is no inference_models/tests/conftest.py in this repo. If the real issue is picking up the correct pytest.ini, consider rephrasing to: “run from inference_models/ so pytest uses inference_models/pytest.ini (or use pytest -c inference_models/pytest.ini ... from repo root)”.

[Copilot]

The adapter guidance here doesn’t match the actual adapter pattern in inference/core/models/inference_models_adapters.py: adapters call AutoModel.from_pretrained(model_id_or_path=..., ...) and pass additional flags like allow_untrusted_packages, allow_direct_local_storage_loading, backend selection, etc. Recommend updating this sentence to reference the real signature/keyword (model_id_or_path) and to explicitly say “follow the existing adapter constructors in that file” to avoid copy/paste mistakes.

Suggested change

Skip unless the user wants plain `/infer` endpoint support. Add a subclass of `Model` to `inference/core/models/inference_models_adapters.py` matching your task (there are per-task parents: object detection, instance segmentation, classification, keypoints, semantic segmentation, etc. — read the existing adapters in that file). The adapter's `__init__` calls `AutoModel.from_pretrained(model_id, api_key=api_key)` and stores the result; predict / infer methods delegate.

Skip unless the user wants plain `/infer` endpoint support. Add a subclass of `Model` to `inference/core/models/inference_models_adapters.py` matching your task (there are per-task parents: object detection, instance segmentation, classification, keypoints, semantic segmentation, etc. — read the existing adapters in that file). In the adapter `__init__`, follow the existing adapter constructors in that file: they call `AutoModel.from_pretrained(model_id_or_path=..., ...)` and pass through the additional flags they need (for example `allow_untrusted_packages`, `allow_direct_local_storage_loading`, backend selection, etc.), then store the result; predict / infer methods delegate.

[Copilot]

PYTHONSAFEPATH=1 / python -P only works on newer Python versions (and -P isn’t available on Python 3.10, which this repo still tests in CI). Suggest noting the minimum Python version for this workaround and/or providing an alternative for 3.10 (e.g., run via python -m ... from an installed env, or adjust PYTHONPATH/working directory) so contributors don’t hit an “unknown option”/no-op.

Suggested change

**When running `debugrun.py` or the inference server from the repo root**, pass `PYTHONSAFEPATH=1`. The `inference_models/` project dir (at repo root, no `__init__.py`) is treated by Python as an empty namespace package and shadows the editable-installed `inference_models`. `PYTHONSAFEPATH=1` (or `python -P`) stops Python from auto-adding the script dir to `sys.path` and the shadow disappears.

**When running `debugrun.py` or the inference server from the repo root**, avoid letting the repo-root `inference_models/` directory shadow the editable-installed `inference_models` package. On newer Python versions that support it, you can use `PYTHONSAFEPATH=1` (or `python -P`) so Python does not auto-add the script directory to `sys.path`. **Do not rely on `python -P` on Python 3.10**. For Python 3.10, prefer running from an installed environment via `python -m ...` instead of invoking a repo-root script directly, or adjust your `PYTHONPATH` / working directory so the repo-root namespace package is not on `sys.path`.

[Copilot]

The guidance for task types references a TaskType enum in model_registry_utils.http.weights_service, but that module doesn’t exist in this repo and TaskType in inference_models.models.auto_loaders.entities is just Optional[str] (not an enum). Suggest rewording to point to the concrete task constants defined in inference_models/inference_models/models/auto_loaders/models_registry.py (e.g. OBJECT_DETECTION_TASK, etc.) and to the model-registry API/schema docs (or the SDK) for the service-side accepted strings, rather than an in-repo enum that isn’t present.

Suggested change

	2. **Task type** — must match both the `TaskType` enum in `model_registry_utils.http.weights_service` **and** the task-type constant in `inference_models/.../models_registry.py`.
	2. **Task type** — choose one of the concrete task constants defined in `inference_models/inference_models/models/auto_loaders/models_registry.py` (for example `OBJECT_DETECTION_TASK`, etc.), and verify the exact service-side accepted string against the model-registry API/schema docs or the model-registry SDK.

[Copilot]

writing-tests.md is referenced without a path, but the doc lives under inference_models/docs/contributors/writing-tests.md. Consider linking to the full path (and keeping the two links consistent) so readers can navigate directly from GitHub.

Suggested change

	Read `inference_models/docs/contributors/adding-model.md` and `writing-tests.md` — they cover the `from_pretrained` contract in more depth than this skill.
	Read `inference_models/docs/contributors/adding-model.md` and `inference_models/docs/contributors/writing-tests.md` — they cover the `from_pretrained` contract in more depth than this skill.