Review Workflows Blocks

Reviewing workflows-blocks changes

Scope

Triggers when a PR touches:

• inference/core/workflows/core_steps/** — Roboflow Core plugin blocks ({category}/{block}/v{N}.py, common/, common/query_language/**, loader.py).
• inference/core/workflows/execution_engine/entities/types.py — only when a NEW kind is added for a block (review it alongside the block).

OUT of scope (other skills own these): Execution Engine internals under inference/core/workflows/execution_engine/v1/** (compiler, executor, output_constructor — review-workflows-execution-engine); model backends under inference_models/. Since the tensor-data-representation merge (#2357), the vN_tensor.py siblings ARE in scope here — review them with the Tensor-native parity items below; authoring-side counterpart: create-workflow-block. UQL / common/query_language/** has no dedicated skill, so it IS reviewed here (see the query-language standard below).

Review checklist

Severity: BLOCK = must fix before merge; FLAG = raise it; NIT = optional. Each item maps to one canonical rule in ## Standards below.

1. BLOCK — Output contract: enumerate describe_outputs() keys; confirm EVERY run()/helper return dict (happy, empty, error, zero-area) carries exactly those keys (#2346).
2. BLOCK — Versioning discipline: bug-fix patches the existing vN.py; behavior change / new inputs-outputs goes in a NEW vN.py. Confirm no shipped type identifier or legacy alias was removed or repurposed.
3. BLOCK — Loader wiring: new block imported into load_blocks(); new kind into load_kinds() + defined in types.py with *_KIND_DOCS; serializable data-kind into KINDS_SERIALIZERS/KINDS_DESERIALIZERS.
4. BLOCK — __init__.py exists in every new/nested block package (guarded by test_init_files.py).
5. BLOCK — sv.Detections.data density: any new data key holds dense N-d numpy arrays (no dtype="object" ragged) and has serializer/deserializer parity if it must round-trip.
6. FLAG — Tests: unit test under tests/workflows/unit_tests/core_steps/{category}/; integration test under tests/workflows/integration_tests/execution/test_workflow_with_*_block.py for new or output-changing blocks.
7. FLAG — EE compatibility floor: actively verify get_execution_engine_compatibility() against the EE capabilities the block actually uses — contributors habitually leave the copy-pasted default range unexamined. Derive the true minimum from the EE changelog in the roboflow/docs repo (workflows/developer-guide/execution-engine-changelog.md) and require the floor raised when the block relies on a capability introduced after the declared minimum. If a capability the block relies on exists only under ## Unreleased (including one added in the same PR), the EE version must be placed and bumped so the block can gate on it — see the EE-compatibility standard.
8. FLAG — EE behavior change: if the block forces an EE behavior change, require a user-facing entry under ## Unreleased in the EE changelog in the roboflow/docs repo (workflows/developer-guide/execution-engine-changelog.md) and surface the maintainer notice Execution Engine requires a version bump for release — never ask the contributor to bump EXECUTION_ENGINE_V1_VERSION or update version-assert tests (maintainers do both in the release PR), except in the unreleased-dependency case above. Outside that exception, a contributor PR touching that constant is a red flag.
9. FLAG — Cross-block kinds: output kinds upstream ⊆ input kinds accepted downstream (esp. detections/segmentation/keypoint interplay).
10. FLAG — Model-version changes update defaults/enums (not silent deletion) and keep old identifiers loadable.
11. FLAG — Batch handling for model/foundation blocks: run iterates Batch[...] correctly and declares get_parameters_accepting_batches(). 11a. FLAG — Dependent-resource declaration: any block that loads a model or external resource overrides discover_dependent_resources() (helpers roboflow_platform_model() / third_party_model()) so resource discovery and model pre-loading see it. A general block contract — not tensor-specific; a missing declaration means video pipelines pay the cold model load on the first frame.
12. FLAG — UQL correctness over speed: be skeptical of micro-optimizations to common/query_language/** ops that change behavior (#1112/#1161/#1162 were reverted).
13. NIT — Field metadata: new inputs have title/description/examples; default= correct (no stray default=None on required fields); json_schema_extra (relevant_for, always_visible) sensible.

Tensor-native parity (blocks with vN_tensor.py siblings)

14. BLOCK — Sibling drift: a PR editing a numpy block's behavior without the mirrored edit in its _tensor sibling (or vice versa) is a finding — main-merge compensation sweeps found dozens of drifted siblings; drift ships divergent flag-on behavior silently.
15. BLOCK — Manifest surface identical across flags: field set, supported variants, batch-parameter declarations, AND output-kind list order must match between siblings — serializer selection is first-kind-that-serializes-wins, so kind-order divergence changes the wire payload per deployment flag (instance-seg v1–v3 RLE-vs-polygon lesson).
16. BLOCK — Both-sibling companions: general contracts hold on BOTH manifests — discover_dependent_resources() (the general rule is 11a; the tensor trap is that a numpy-side-only declaration silently disappears flag-on and pre-loading skips the block); sink siblings both honor disable_sinks.
17. BLOCK — Host-mirror hygiene: any tensor sibling rewriting xyxy/class_id/confidence strips the per-box host mirror (strip_host_mirror_metadata/HOST_MIRROR_KEYS) — a carried mirror makes downstream visualizations draw stale geometry with no error.
18. FLAG — Mutation declarations: every copy_image=False in-place draw calls declare_numpy_image_mutated()/declare_tensor_image_mutated() — including NEW mainline blocks arriving via merge without the retrofit (label v2 / rich_label class).
19. FLAG — CPU-only import safety: every module registered in the flag-on loader branch must import without CUDA (no module-level device initialization); the flag-on suite runs on CPU-only CI runners.
20. FLAG — Flag-off registration is frozen: the loader's flag-off block list and import bindings must stay byte-identical to pre-tensor main behavior — flag-off is the compatibility contract; any flag-off divergence riding a tensor PR is a finding.

Not blocking

• Do NOT demand an inference/core/version.py bump — inference releases are versioned separately from feature/bugfix PRs.
• Do NOT demand a new vN.py for a pure in-place bug-fix; surgical patches to a shipped version are correct (#2346, #1339, #1368).
• Never demand an EXECUTION_ENGINE_V1_VERSION bump from a contributor — maintainers bump it at release. Sole exception: a block gating on unreleased EE capabilities (see the EE-compatibility standard), where the bump must be maintainer-coordinated before the block ships. The ## Unreleased EE changelog entry is required only when the block forces an EE behavior change; most block changes don't.
• Do NOT demand serializer/deserializer entries for a data key that is purely intra-block and never crosses the wire.
• Do NOT demand hand-written docs/workflows/blocks/<block>.md — those are generated from Field(description=, examples=) metadata.

Standards

The one canonical statement of each rule. Past regressions are cited inline; see ## Reference PRs.

• Output contract: every dict run() emits contains EXACTLY the keys declared by describe_outputs(), on every path including empty/zero-area/error branches (return [], {"x": None}) (#2346 — dynamic_crop zero-area branch dropped predictions; #1341 — empty-zone branch).
• Versioning: bug-fixes patch the existing version in place; behavior changes / new inputs-outputs go into a NEW vN.py module — even at the cost of code duplication (docs/workflows/versioning.md, blocks_bundling.md). Never break type identifiers or legacy aliases; e.g. dynamic_crop/v1.py keeps Literal["roboflow_core/dynamic_crop@v1", "DynamicCrop", "Crop"]. Versioning foundation: #565.
• New block = full package: new dir with __init__.py + v1.py, imported and added to load_blocks() in loader.py, plus a unit test AND an integration test (#2362 — overlap_analysis added all of these + a new kind).
• __init__.py in every package (including nested): its absence breaks docs generation and block import; guarded by test_init_files.py (#2351, #830).
• New kind discipline: define the kind in types.py with a *_KIND_DOCS string, return it from load_kinds(), and route serialization through KINDS_SERIALIZERS/KINDS_DESERIALIZERS (#2362 — DETECTIONS_OVERLAPS_KIND; #804 — INFERENCE_ID_KIND as a distinct semantic kind across blocks).
• sv.Detections.data must be dense numpy arrays: object-dtype ragged arrays break supervision indexing / is_data_equal; pad to uniform shape (#2170 — keypoints; #1368 — velocity elements must be np.array).
• Serializer/deserializer parity for round-tripping data keys: a data key that must survive serialization needs matching entries in common/serializers.py + common/deserializers.py (#1368 — velocity/speed; #1156 — timestamps).
• New input fields keep backward compat: add with a default= and update describe_outputs()/run(); behavior-changing additions warrant a new version. Question default=None on non-optional inputs (#2384 — added in-place across v1/v2/v3 with defaults; #658 — removed wrong default=None).
• EE compatibility: get_execution_engine_compatibility() declares the minimum EE version the block relies on (e.g. ">=1.3.0,<2.0.0"); the EE version is a single source of truth in EXECUTION_ENGINE_V1_VERSION. The floor is a contract, not boilerplate: derive it from the EE changelog in the roboflow/docs repo (workflows/developer-guide/execution-engine-changelog.md) — a block needing a capability added in 1.3.7 declares ">=1.3.7,<2.0.0", not the copy-pasted default. A block that forces an EE behavior change makes an EE version bump mandatory at release: the contributor adds the ## Unreleased changelog entry only; maintainers apply the bump + final changelog heading + version-assert test updates in the release PR (#2527 — revert rolled back both the block and the EE version and re-asserted it in tests). Unreleased-dependency exception: when the block relies on a capability still under ## Unreleased, there is no released floor to declare — the EE version must be placed (the ## Unreleased entries get their final vX.Y.Z heading) and EXECUTION_ENGINE_V1_VERSION bumped with mirrored version-assert tests, maintainer-coordinated, in or before this PR, so the block's range can reference it. This is the one case where the bump cannot be deferred to a release PR.
• Model-version deprecation: when a hosted model version is deprecated, update the block default and enum; do not just delete, keep old identifiers loadable (#1668/#1669 — Claude 3.5→4.5; #500 — GPT-4→GPT-4o; #2395 — Gemini default).
• Batch handling for model/foundation blocks: run iterates Batch[...] and declares get_parameters_accepting_batches() (#718 — VLM blocks failed on batch inference).
• Segmentation/mask edge cases: mask→polygon must handle 0/1/degenerate contours (returns None, duplicates points); empty masks must serialize (#1339; #697 — SAM2 masks as points/lines).
• Visualization blocks are fragile across supervision versions: verify against the pinned supervision public API, not internals (#814/#811/#1868 — label/line regressions; #1725/#1951 — sv 0.27.0 handling).
• Query language (UQL): common/query_language/** ops must preserve exact behavior; reject micro-optimizations that alter results (#1112/#1161/#1162 — automated speed-up PRs reverted).
• Rich Field metadata: title, description, examples, and json_schema_extra (relevant_for, always_visible) drive the FE app; block docs are generated from these (#623/#704; #2058 — property options propagated to FE).
• Tensor-sibling contract (#2357): siblings are standalone verbatim re-implementations against native objects — no delegation/method-binding to the numpy manifest (wholesale subclassing OK). Same-name registration in the loader's flag-on branch; the flag swaps implementations, never the registry (block-name lists identical both directions). Testing convention: per-file _TENSOR_ONLY/_NUMPY_ONLY skipif markers, mirrored assertions, and the shared image_as_workflow_input integration fixture parametrizing image inputs as numpy AND torch.Tensor — new tensor-native integration tests should take it.

Required companions

Block a functional change that lacks these (canonical rules above; conditions here):

• Loader registration — new block in load_blocks(); new kind in load_kinds() + types.py; new serializable data-kind in KINDS_SERIALIZERS/KINDS_DESERIALIZERS, all in loader.py (#2362, #804).
• Tests — unit test always; integration test for new or output-changing blocks (#2362, #718).
• __init__.py — in every new/nested block package (#2351, #830).
• EE changelog — ONLY when the block forces an EE behavior change: add an entry under ## Unreleased in the EE changelog in the roboflow/docs repo (workflows/developer-guide/execution-engine-changelog.md); maintainers bump EXECUTION_ENGINE_V1_VERSION at release time.

Key files & entry points

• inference/core/workflows/core_steps/loader.py — load_blocks(), load_kinds(), KINDS_SERIALIZERS/KINDS_DESERIALIZERS, REGISTERED_INITIALIZERS.
• inference/core/workflows/core_steps/{category}/{block}/vN.py — block modules (manifest + block class).
• inference/core/workflows/core_steps/common/{serializers,deserializers,utils}.py — sv.Detections <-> wire conversion, keypoints/mask helpers.
• inference/core/workflows/core_steps/common/query_language/** — UQL operations/evaluation engine.
• inference/core/workflows/execution_engine/entities/types.py — Kind definitions + *_KIND_DOCS.
• inference/core/workflows/execution_engine/entities/base.py — WorkflowImageData, Batch, OutputDefinition.
• inference/core/workflows/prototypes/block.py — WorkflowBlock, WorkflowBlockManifest, BlockResult.
• inference/core/workflows/execution_engine/v1/core.py — EXECUTION_ENGINE_V1_VERSION.
• Docs: the Workflows developer guide on docs.roboflow.com (create a block, versioning, block bundling, EE changelog): https://docs.roboflow.com/workflows/developer-guide/developer-guide.
• Tests: tests/workflows/unit_tests/core_steps/**, tests/workflows/integration_tests/execution/**, test_init_files.py.

Reference PRs

• #2362 — new fusion block + new kind + both test tiers (canonical new-block shape).
• #2346 — missing declared output key on empty branch (output-contract).
• #2170 — object-dtype ragged keypoint arrays break supervision.
• #1368 — sv.Detections.data must hold numpy arrays + serializer parity.
• #2384 — backward-compat input field added in-place to v1/v2/v3 with defaults.
• #804 — new INFERENCE_ID_KIND across blocks (kinds discipline).
• #565 — Workflows block versioning foundation.
• #2351 — missing __init__.py (packaging/docs).
• #2527 — revert: block removal + EE version rollback + test re-assert.
• #658 — stray default=None in manifests.
• #718 — VLM blocks failing on batch inference.
• #1339 — empty-segmentation serialization edge cases.

Related topic skills

When the PR also exhibits these cross-cutting concerns, load the matching topic skill too (see each skill's description for the trigger):

• review-topic-workflow-state-management
• review-topic-local-vs-remote-execution
• review-topic-prediction-integrity
• review-topic-backward-compat-and-versioning
• review-topic-test-hygiene
• create-workflow-block