inference

Trigger when the user asks to "add a model", "add a core / pre-trained model", "port a model from HuggingFace", "wrap a transformers model", or "expose X as a workflow block / model". Also when a diff adds a subfolder under inference_models/inference_models/models/<family>/, a new (architecture, task, BackendType) key in auto_loaders/models_registry.py, an adapter in inference/core/models/inference_models_adapters.py, or a block under core_steps/models/foundation/<family>/.

Author a new Roboflow Workflows block under `inference/core/workflows/core_steps/**`. Trigger on a new/changed `vN.py` defining a `WorkflowBlockManifest` + `WorkflowBlock` pair; new `load_blocks()`/`KINDS_SERIALIZERS` entries in `core_steps/loader.py`; new `Kind(...)` in `execution_engine/entities/types.py`; or a request to "create/add a workflow block" or "expose X as a workflow step" (model / transform / sink / visualization / flow control).

PRs changing inference_cli/ except server.py, container_adapter.py, tunnel_adapter.py, infer_adapter.py. Diff signals — typer.Option/typer.Argument, cloud_app/rf_cloud_app/benchmark_app/enterprise_app, get_workspace/handle_response_errors, @backoff.on_exception(..., RetryError), serialization_alias/alias= on pydantic Fields, REQUEST_TIMEOUT config constants, cloud_adapter.py (SkyPilot deploy), inference_cli/lib/enterprise/**, tests/inference_cli/**, docs/inference_helpers/cli_commands/**.

Load when a PR touches inference_models/inference_models/** (models/base/, models/common/, models/auto_loaders/**, weights_providers/**, runtime_introspection/**, errors.py, configuration.py), inference_models/pyproject.toml, inference_models/uv.lock, or inference_models/docs/**. Diff signals: a REGISTERED_MODELS entry, LazyClass/RegistryEntry, from_pretrained, to_supervision, a BaseInferenceModelsError subclass, an INFERENCE_MODELS_* env var, torch.jit.load/script.

Review guidance for PRs touching .github/ (workflows, actions, prompts/claude-pr-review.md, CODEOWNERS), docker/ (dockerfiles, publish, config, entrypoint), requirements/*.txt, .release/pypi/*.setup.py, Makefile, inference/core/version.py, and inference_models/{pyproject.toml,uv.lock}. Diff signals: new/edited GitHub Actions workflow or action.yml, Dockerfile edits, requirements-pin bumps, version bumps, tag/publish logic.

Load when a PR changes a user/dev-visible contract or is release-bound. Triggers on diffs to `inference/core/version.py`, `EXECUTION_ENGINE_V1_VERSION`, `inference/core/interfaces/http/http_api.py` or `inference/core/entities/**`, a Workflow block manifest under `inference/core/workflows/core_steps/**` (`type` Literal, `kinds`, `get_execution_engine_compatibility`), a public `inference_models` API, `inference_models/pyproject.toml`, a persisted/cache format, or `requirements/*.txt` / `uv.lock`.

Review PRs touching inference/core/workflows/core_steps/** (Workflow blocks — {category}/{block}/vN.py, common/, common/query_language/**, loader.py) or inference/core/workflows/execution_engine/entities/types.py (new kind). Diff signals — new block package, new/changed vN.py, new type Literal or legacy alias, describe_outputs()/run() changes, load_blocks()/load_kinds()/KINDS_SERIALIZERS edits, new *_KIND, sv.Detections.data writes, serializers/deserializers changes.

Review guidance for PRs touching inference/core/workflows/execution_engine/**, inference/core/workflows/errors.py, inference/core/workflows/prototypes/block.py, docs/workflows/execution_engine_changelog.md, or .cursor/rules/execution-engine-version-changelog.mdc. Diff signals: EXECUTION_ENGINE_V1_VERSION, REGISTERED_ENGINES, COMPILATION_CACHE, contains_future/resolve_futures, ClientCausedStepExecutionError, get_output_dimensionality_offset. NOT core_steps block implementations.

PRs touching inference/core/ foundation — inference/core/env.py, inference/core/version.py, inference/core/entities/**, inference/core/utils/** (image_utils.py, requests.py, environment.py), inference/core/roboflow_api.py, exceptions.py, logger.py, nms.py. Diff signals: new os.getenv(...), __version__, build_roboflow_api_headers, api_key_safe_raise_for_status, load_image_from_url, WHITELISTED/BLACKLISTED_DESTINATIONS_FOR_URL_INPUT, *DC / _is_*_dc_to_dict.

Review guidance for PRs touching inference/core/interfaces/http/** (http_api.py, error_handlers.py, orjson_utils.py, dependencies.py, request_metrics.py, middlewares/, builder/, handlers/) or inference_cli/server.py. Also loads on diffs adding a route/middleware, an except arm in with_route_exceptions / with_route_exceptions_async, a new exception in inference/core/exceptions.py needing HTTP mapping, a header constant in constants.py, or a route-gating env flag in inference/core/env.py.

Load when a PR touches inference/models/** (legacy ORT/HF models, models/utils.py ROBOFLOW_MODEL_TYPES + get_roboflow_model, models/aliases.py), inference/core/models/** (base.py, roboflow.py CLASS_MAP parsing, *_base.py, stubs.py, inference_models_adapters.py), or inference/core/registries/** (get_model_type, GENERIC_MODELS). Diff signals: new ROBOFLOW_MODEL_TYPES key, resolve_roboflow_model_alias, _session_lock/_state_lock, ort_session.run, USE_INFERENCE_MODELS.

Load for PRs touching inference_sdk/** (http/client.py InferenceHTTPClient, http/entities.py, http/errors.py, http/utils/**, config.py, webrtc/**), tests/inference_sdk/**, docs/inference_helpers/inference_sdk*, .release/pypi/inference.sdk.setup.py, or requirements/requirements.sdk.*.txt. Diff signals: a new client method, a new/renamed HTTPClientError subclass, a new InferenceConfiguration field, response.raise_for_status, api_key redaction, @wrap_errors_async, or __ensure_v1_client_mode.

Load when a PR adds or modifies an outbound HTTP call (requests.*, aiohttp, httpx, urllib), constructs a URL from API_BASE_URL / METRICS_COLLECTOR_BASE_URL / HOSTED_*_URL / any *.roboflow.com host, adds an env var or setting that holds an endpoint URL, builds an InferenceHTTPClient, or touches wrap_url / roboflow_secure_gateway_proxy_url_builder / SECURE_GATEWAY handling.

Load when a PR touches `http_api.py` auth middleware (`check_authorization_serverless` / `check_authorization`), `get_serverless_usage_check_async`, api-key→workspace resolution (`get_roboflow_workspace`), assume-identity (`_add_assume_identity_headers`, `x-assume-identity-*`), the unauthenticated-route allowlist, api-key redaction (`api_key_safe_raise_for_status`, `deduct_api_key`), usage api-key hashing, or new `*_SECRET`/`*_TOKEN`/`API_KEY` env vars in `env.py`.

Load when a diff adds threads/async/background work (ThreadPoolExecutor, threading.Thread, asyncio.Lock, concurrent.futures.Future) or touches torch.jit.load/script, torchscript_state_global_lock, acquire_with_timeout, inference/core/managers/** (add_model/remove_model), on-disk caches/artifacts (inference/core/cache/**, model_artifacts.py, MODEL_CACHE_DIR, FileLock, AtomicPath), or stream resources (inference_pipeline.py terminate, webrtc/modal worker, VideoSource).

Load when a PR touches inference/core/roboflow_api.py, inference/core/registries/roboflow.py, or the inference_models auto_negotiation resolver; changes an inference_sdk<->server response shape (response_model, model_dump, parent_id); reads a NEW nested platform key (api_data.get("taskType"), modelVariant); adds fallback / backend-downgrade / .get(key, default) on external data; or adds an env flag/secret changing which service/backend is contacted (USE_INFERENCE_MODELS, DISABLED_INFERENCE_MODELS_BACKENDS).

Load when a PR touches SSRF/URL-fetch guards (`_ensure_url_input_allowed`, `_ensure_resource_fqdn_allowed`, `WHITELISTED_DESTINATIONS_FOR_URL_INPUT`/`BLACKLISTED_DESTINATIONS_FOR_URL_INPUT`), local-file image loads (`ALLOW_LOADING_IMAGES_FROM_LOCAL_FILESYSTEM`), model-cache paths from a `model_id` (`cache_path_is_within_root`, `MODEL_CACHE_DIR`), unsafe deserialization (`pickle.loads`, `torch.load`, `weights_only`, `allow_untrusted_packages`), or archive extraction (`tarfile.extractall`, `ZipFile`).

Load when a PR touches how a step/model runs in-process vs. over the network: Workflow-block execution-mode dispatch (self._step_execution_mode, StepExecutionMode.LOCAL/REMOTE, run_locally/run_remotely), a block get_restrictions()/get_air_gapped_availability(), backend/runtime routing env (USE_INFERENCE_MODELS, DISABLED_INFERENCE_MODELS_BACKENDS, WORKFLOWS_REMOTE_API_TARGET, inference_models_adapters.py), CUDA/Jetson/ONNX gating, or a new block with network/LAN/same-process assumptions.

Load when a diff touches `sv_detections_to_root_coordinates`, `scale_sv_detections`, `move_boxes`/`move_masks`, `POLYGON_KEY_IN_SV_DETECTIONS`, `add_inference_keypoints_to_sv_detections`, `serialise_sv_detections`, `mask_to_polygon`, `enforce_dense_masks_in_inference_models`, `binarization_threshold`, `run_nms_for_*`; or `xyxy +=`/`* scale` arithmetic; crop/stitch/perspective/dynamic_zones blocks; `inference_models/` pre/post-processing (resize/BGR-RGB/NMS/slice); a `supervision` bump.

Load when a diff shows — (1) new/changed runtime code under inference/, inference_models/, inference_sdk/, inference_cli/ with no companion test (or a test that mocks the changed thing); (2) a new @pytest.mark.skipif/skip/slow or env-gated test (SKIP_*_TESTS, RUN_*); (3) a new module-level monkeypatch, sys.modules[...] stub, or autouse fixture mutating a shared singleton/registry; (4) a GPU-only test whose CI runner is unclear. Greppable: PROPERTIES_EXTRACTORS, sys.modules[, pytest.ini.

Load when a diff in inference/core/workflows/core_steps/** or prototypes/block.py seeds cross-run() state in a WorkflowBlock __init__ (self._trackers, self._sessions, self._previous_positions, self.cache, dicts/deques/sets), keys state by video_metadata.video_identifier / session / api_key, edits trackers/_base.py InstanceCache, touches analytics/time_in_zone|velocity or flow_control/delta_filter, or adds/edits get_restrictions() / STATEFUL_VIDEO_HTTP_SOFT_RESTRICTION / a remote guard.