Merge branch 'main' into feat/add-recap

# Conflicts:
#	uv.lock

tests/annotations/_helpers.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Helpers shared across annotation-pipeline tests."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from lerobot.annotations.steerable_pipeline.vlm_client import StubVlmClient
+
+
+def make_canned_responder(
+    responses_by_marker: dict[str, Any],
+    default: Any = None,
+) -> StubVlmClient:
+    """Return a stub that picks a response by inspecting the user prompt.
+
+    For each call the responder examines the last user-message text and
+    returns the response keyed by the first marker substring it contains.
+    Falls back to ``default`` if no marker matches.
+    """
+
+    def responder(messages: list[dict[str, Any]]) -> Any:
+        last_user_text = ""
+        for message in messages:
+            if message.get("role") != "user":
+                continue
+            content = message.get("content")
+            if isinstance(content, str):
+                last_user_text = content
+            elif isinstance(content, list):
+                for block in content:
+                    if isinstance(block, dict) and block.get("type") == "text":
+                        last_user_text = block.get("text", "")
+        for marker, response in responses_by_marker.items():
+            if marker in last_user_text:
+                return response
+        return default
+
+    return StubVlmClient(responder=responder)
+
+
+def encode_vqa_answer(payload: dict[str, Any]) -> str:
+    return json.dumps(payload, sort_keys=True)

tests/annotations/conftest.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Shared fixtures for annotation-pipeline tests.
+
+The on-disk dataset builder lives with the other dataset factories in
+``tests/fixtures/dataset_factories.py`` (:func:`build_annotation_dataset`);
+these fixtures only wire it into pytest.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+# ``build_annotation_dataset`` pulls in ``lerobot.datasets`` (HF ``datasets``
+# + ``pandas``, only in the ``dataset`` extra), so it's imported lazily inside
+# each fixture — this conftest stays importable without that extra. The test
+# modules ``pytest.importorskip("datasets")`` so they skip rather than error.
+
+
+@pytest.fixture
+def fixture_dataset_root(tmp_path: Path) -> Path:
+    """A tiny dataset with two episodes, 12 frames each at 10 fps."""
+    from tests.fixtures.dataset_factories import build_annotation_dataset
+
+    return build_annotation_dataset(
+        tmp_path / "ds",
+        episode_specs=[
+            (0, 12, "Could you tidy the kitchen please?"),
+            (1, 12, "Please clean up the kitchen"),
+        ],
+        fps=10,
+    )
+
+
+@pytest.fixture
+def single_episode_root(tmp_path: Path) -> Path:
+    from tests.fixtures.dataset_factories import build_annotation_dataset
+
+    return build_annotation_dataset(
+        tmp_path / "ds_one",
+        episode_specs=[(0, 30, "Pour water from the bottle into the cup.")],
+        fps=10,
+    )

tests/annotations/run_e2e_smoke.py

@@ -0,0 +1,116 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Opt-in E2E smoke run for ``make annotation-e2e``.
+
+Builds the shared annotation fixture (:func:`build_annotation_dataset`),
+runs the full annotation pipeline against it with a stub VLM, and prints a
+short report. This is intentionally not a pytest test — it exercises the
+CLI plumbing — but it reuses the same on-disk dataset builder as the pytest
+fixtures so there is no duplicated fixture code.
+"""
+
+from __future__ import annotations
+
+import sys
+import tempfile
+from pathlib import Path
+
+from lerobot.annotations.steerable_pipeline.config import AnnotationPipelineConfig
+from lerobot.annotations.steerable_pipeline.executor import Executor
+from lerobot.annotations.steerable_pipeline.modules import (
+    GeneralVqaModule,
+    InterjectionsAndSpeechModule,
+    PlanSubtasksMemoryModule,
+)
+from lerobot.annotations.steerable_pipeline.validator import StagingValidator
+from lerobot.annotations.steerable_pipeline.vlm_client import StubVlmClient
+from lerobot.annotations.steerable_pipeline.writer import LanguageColumnsWriter
+from tests.fixtures.dataset_factories import build_annotation_dataset
+
+
+def _stub_responder(messages):
+    text = ""
+    for m in messages:
+        if m.get("role") == "user":
+            content = m.get("content")
+            if isinstance(content, list):
+                for block in content:
+                    if isinstance(block, dict) and block.get("type") == "text":
+                        text = block.get("text", "")
+            elif isinstance(content, str):
+                text = content
+    if "atomic subtasks" in text:
+        return {
+            "subtasks": [
+                {"text": "grasp the bottle", "start": 0.0, "end": 1.0},
+                {"text": "pour into the cup", "start": 1.0, "end": 2.0},
+                {"text": "place the bottle down", "start": 2.0, "end": 3.0},
+            ]
+        }
+    if "compressed semantic memory" in text:
+        return {"memory": "poured once"}
+    if "acknowledgement the robot" in text:
+        return {"text": "Sure."}
+    if "compact interjection" in text:
+        return {"interjection": "use less water", "speech": "Using less water."}
+    if "frame-grounded visual question" in text:
+        return {"question": "How many cups?", "answer": {"label": "cup", "count": 1}}
+    return None
+
+
+def main() -> int:
+    with tempfile.TemporaryDirectory() as tmp:
+        root = build_annotation_dataset(
+            Path(tmp) / "ds",
+            episode_specs=[(0, 30, "Pour water into the cup.")],
+            fps=10,
+        )
+        vlm = StubVlmClient(responder=_stub_responder)
+        cfg = AnnotationPipelineConfig()
+        executor = Executor(
+            config=cfg,
+            plan=PlanSubtasksMemoryModule(vlm=vlm, config=cfg.plan),
+            interjections=InterjectionsAndSpeechModule(vlm=vlm, config=cfg.interjections, seed=cfg.seed),
+            vqa=GeneralVqaModule(vlm=vlm, config=cfg.vqa, seed=cfg.seed),
+            writer=LanguageColumnsWriter(),
+            validator=StagingValidator(),
+        )
+        summary = executor.run(root)
+        print(f"phases={[(p.name, p.episodes_processed) for p in summary.phases]}")
+        print(f"validation: {summary.validation_report.summary()}")
+        print(f"shards rewritten: {len(summary.written_paths)}")
+
+        # Assert the interjection code path actually fired — otherwise a stale
+        # canned-VLM marker would silently produce zero interjections and this
+        # smoke run would still "pass" by only printing.
+        import pyarrow.parquet as pq  # noqa: PLC0415
+
+        events = [
+            r
+            for shard in summary.written_paths
+            for ev in pq.read_table(shard).column("language_events").to_pylist()
+            for r in ev
+        ]
+        n_interjections = sum(1 for r in events if r.get("style") == "interjection")
+        n_speech = sum(1 for r in events if r.get("style") is None and r.get("role") == "assistant")
+        print(f"interjections={n_interjections} speech_atoms={n_speech}")
+        assert n_interjections > 0, "no interjection rows produced — check the interjection prompt marker"
+        assert n_speech > 0, "no speech tool-call atoms produced — check the speech prompt marker"
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tests/annotations/test_frames.py

@@ -0,0 +1,246 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Unit tests for :class:`VideoFrameProvider` method bindings.
+
+These were prompted by a real regression: ``video_for_episode`` was once
+indented one level too deep so it ended up nested *inside* a module-level
+helper (after that function's ``return`` statement) — silently dead code
+that meant production runs with ``use_video_url=False`` would
+``AttributeError`` on ``self.frame_provider.video_for_episode(...)``. The
+existing module tests didn't catch it because they exercise stub providers.
+
+The tests below assert on the class itself (not on an instance), so a
+future reindent regression flips them to red without needing a real
+LeRobot dataset on disk.
+"""
+
+from __future__ import annotations
+
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+import torch
+
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+
+from lerobot.annotations.steerable_pipeline.frames import VideoFrameProvider  # noqa: E402
+
+
+class _FakeMeta:
+    """Minimal metadata stub exposing ``video_keys`` / ``camera_keys``."""
+
+    def __init__(self, video_keys: list[str], image_keys: list[str], video_path: Path | None = None) -> None:
+        self.video_keys = video_keys
+        self.camera_keys = [*video_keys, *image_keys]
+        self._video_path = video_path
+        self.episodes = {0: {f"videos/{key}/from_timestamp": 0.0 for key in video_keys}}
+
+    def get_video_file_path(self, episode_index: int, camera_key: str) -> Path:
+        return self._video_path
+
+
+def test_default_camera_key_skips_image_only_cameras(tmp_path: Path, monkeypatch) -> None:
+    """The default camera must be a *video* key — image-stored cameras have no
+    ``videos/<key>/from_timestamp`` and would KeyError in the clip/decode path.
+
+    Regression: a dataset whose first ``camera_keys`` entry was an image-stored
+    camera (e.g. ``observation.images.wrist``) crashed at clip extraction.
+    """
+    fake = _FakeMeta(
+        video_keys=["observation.images.robot0_agentview_right"],
+        image_keys=["observation.images.wrist"],
+    )
+    import lerobot.datasets.dataset_metadata as meta_mod
+
+    monkeypatch.setattr(meta_mod, "LeRobotDatasetMetadata", lambda *a, **k: fake, raising=True)
+    provider = VideoFrameProvider(root=tmp_path)
+    assert provider.camera_key == "observation.images.robot0_agentview_right"
+    assert "observation.images.wrist" not in provider.camera_keys
+
+
+def test_video_for_episode_is_a_method_of_videoframeprovider():
+    """``video_for_episode`` must be a bound method, not nested dead code."""
+    assert callable(getattr(VideoFrameProvider, "video_for_episode", None))
+
+
+def test_episode_clip_path_is_a_method_of_videoframeprovider():
+    """``episode_clip_path`` is now a method (was a free function reaching
+    into ``provider._meta`` from outside the class)."""
+    assert callable(getattr(VideoFrameProvider, "episode_clip_path", None))
+
+
+def test_videoframeprovider_has_a_lock_for_concurrent_use():
+    """A ``ThreadPoolExecutor`` runs the plan / interjections / vqa phases
+    concurrently; the cache + warn-flag accesses must be guarded.
+    """
+    import threading
+
+    # Fresh-instance check via a minimal fake to avoid touching the hub.
+    # The lock is declared with ``init=False`` and has a default factory,
+    # so a constructed instance must own a real ``threading.Lock``.
+    lock_field = next(
+        (f for f in VideoFrameProvider.__dataclass_fields__.values() if f.name == "_lock"),
+        None,
+    )
+    assert lock_field is not None
+    assert lock_field.default_factory is threading.Lock
+
+
+@pytest.fixture
+def sample_video(tmp_path: Path) -> Path:
+    """A 3 s 10 fps test-pattern mp4, written with ffmpeg."""
+    if shutil.which("ffmpeg") is None:
+        pytest.skip("ffmpeg not available")
+    out = tmp_path / "sample.mp4"
+    subprocess.run(
+        [
+            "ffmpeg",
+            "-y",
+            "-f",
+            "lavfi",
+            "-i",
+            "testsrc=duration=3:size=160x120:rate=10",
+            "-pix_fmt",
+            "yuv420p",
+            str(out),
+        ],
+        check=True,
+        capture_output=True,
+    )
+    return out
+
+
+def _provider_for_video(tmp_path: Path, video: Path, monkeypatch) -> VideoFrameProvider:
+    """A provider whose single camera resolves to ``video`` via fake metadata."""
+    fake = _FakeMeta(video_keys=["observation.images.cam"], image_keys=[], video_path=video)
+    import lerobot.datasets.dataset_metadata as meta_mod
+
+    monkeypatch.setattr(meta_mod, "LeRobotDatasetMetadata", lambda *a, **k: fake, raising=True)
+    return VideoFrameProvider(root=tmp_path, tolerance_s=0.2)
+
+
+def test_decode_returns_one_uint8_frame_per_timestamp(
+    sample_video: Path, tmp_path: Path, monkeypatch
+) -> None:
+    """``_decode`` routes through ``decode_video_frames`` (torchcodec when
+    available, PyAV otherwise) — no subprocess fallback.
+    """
+    provider = _provider_for_video(tmp_path, sample_video, monkeypatch)
+    timestamps = [0.0, 1.0, 2.5]
+    frames = provider._decode(0, timestamps, "observation.images.cam")
+
+    assert len(frames) == len(timestamps)
+    for frame in frames:
+        assert isinstance(frame, torch.Tensor)
+        assert frame.dtype == torch.uint8
+        assert frame.shape == (3, 120, 160)
+
+
+def test_frames_at_snaps_mid_frame_grid_to_real_frames(
+    sample_video: Path, tmp_path: Path, monkeypatch
+) -> None:
+    """Uniform sampling grids land mid-frame; ``frames_at`` must snap them to
+    real frame timestamps before decoding.
+
+    Regression: ``decode_video_frames`` rejects queries farther than
+    ``tolerance_s`` (default 10 ms) from a decodable frame, so un-snapped
+    mid-frame queries raised ``FrameTimestampError`` wholesale and the plan
+    module silently lost its contact sheets for most episodes.
+    """
+    from types import SimpleNamespace
+
+    fake = _FakeMeta(video_keys=["observation.images.cam"], image_keys=[], video_path=sample_video)
+    import lerobot.datasets.dataset_metadata as meta_mod
+
+    monkeypatch.setattr(meta_mod, "LeRobotDatasetMetadata", lambda *a, **k: fake, raising=True)
+    provider = VideoFrameProvider(root=tmp_path)  # default 10 ms tolerance
+    # 10 fps fixture -> frames at 0.0, 0.1, ...; queries sit mid-frame.
+    record = SimpleNamespace(episode_index=0, frame_timestamps=[i / 10 for i in range(30)])
+
+    frames = provider.frames_at(record, [0.149, 1.234, 2.04], camera_key="observation.images.cam")
+
+    assert len(frames) == 3
+    for frame in frames:
+        assert isinstance(frame, torch.Tensor)
+        assert frame.shape == (3, 120, 160)
+
+
+def test_decode_returns_empty_list_on_missing_file(tmp_path: Path, monkeypatch) -> None:
+    """A missing video is a recoverable no-frames condition, never a crash."""
+    provider = _provider_for_video(tmp_path, tmp_path / "does_not_exist.mp4", monkeypatch)
+    assert provider._decode(0, [0.0], "observation.images.cam") == []
+
+
+def test_episode_clip_path_trims_via_reencode_video(tmp_path: Path, monkeypatch) -> None:
+    """Clip extraction delegates to ``video_utils.reencode_video`` with the
+    episode's ``[from_timestamp, to_timestamp)`` trim window — no subprocess.
+    """
+    from types import SimpleNamespace
+
+    import lerobot.annotations.steerable_pipeline.frames as frames_mod
+
+    src = tmp_path / "src.mp4"
+    src.write_bytes(b"src")
+    fake = _FakeMeta(video_keys=["observation.images.cam"], image_keys=[], video_path=src)
+    fake.episodes[0]["videos/observation.images.cam/from_timestamp"] = 1.5
+    fake.episodes[0]["videos/observation.images.cam/to_timestamp"] = 4.0
+    import lerobot.datasets.dataset_metadata as meta_mod
+
+    monkeypatch.setattr(meta_mod, "LeRobotDatasetMetadata", lambda *a, **k: fake, raising=True)
+
+    captured = {}
+
+    def fake_reencode(
+        input_video_path,
+        output_video_path,
+        camera_encoder=None,
+        overwrite=False,
+        start_time_s=None,
+        end_time_s=None,
+    ):
+        captured.update(
+            src=Path(input_video_path),
+            encoder=camera_encoder,
+            start_time_s=start_time_s,
+            end_time_s=end_time_s,
+        )
+        Path(output_video_path).write_bytes(b"clip")
+
+    monkeypatch.setattr(frames_mod, "reencode_video", fake_reencode, raising=True)
+    provider = VideoFrameProvider(root=tmp_path)
+    record = SimpleNamespace(episode_index=0, frame_timestamps=[0.0, 1.0])
+
+    out = provider.episode_clip_path(record, tmp_path / "clips")
+
+    assert out == tmp_path / "clips" / "ep_000000.mp4"
+    assert captured["src"] == src
+    assert captured["start_time_s"] == 1.5
+    assert captured["end_time_s"] == 4.0
+    # H.264 so the clip is decodable by vllm's libav build (sources are often AV1).
+    assert captured["encoder"].vcodec == "h264"
+
+
+def test_videoframeprovider_serializes_decodes_with_a_lock() -> None:
+    """torchcodec's cached per-file decoder is single-threaded; the provider
+    must own a dedicated lock that ``_decode`` holds around the decoder call.
+    """
+    import threading
+
+    lock_field = VideoFrameProvider.__dataclass_fields__.get("_decode_lock")
+    assert lock_field is not None
+    assert lock_field.default_factory is threading.Lock

tests/annotations/test_modules.py

@@ -0,0 +1,390 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Module 1/2/3 unit tests with stubbed VLMs."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+import PIL.Image
+import pytest
+
+# ``lerobot.annotations`` imports pull in ``lerobot.datasets`` (-> the HF
+# ``datasets`` library), which only ships under the ``dataset`` extra. Skip
+# this module in tiers without it instead of erroring at import.
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
+
+from lerobot.annotations.steerable_pipeline.config import (  # noqa: E402
+    InterjectionsConfig,
+    PlanConfig,
+    VqaConfig,
+)
+from lerobot.annotations.steerable_pipeline.modules import (  # noqa: E402
+    GeneralVqaModule,
+    InterjectionsAndSpeechModule,
+    PlanSubtasksMemoryModule,
+)
+from lerobot.annotations.steerable_pipeline.reader import iter_episodes  # noqa: E402
+from lerobot.annotations.steerable_pipeline.staging import EpisodeStaging  # noqa: E402
+from lerobot.annotations.steerable_pipeline.vlm_client import StubVlmClient  # noqa: E402
+
+from ._helpers import make_canned_responder  # noqa: E402
+
+
+@dataclass
+class _StubFrameProvider:
+    """Returns one sentinel object per requested timestamp."""
+
+    # A real (tiny) PIL image so the contact-sheet builder, which resizes and
+    # tiles frames, has something to draw. VQA still passes it through by
+    # identity via ``to_image_blocks``.
+    sentinel: Any = field(default_factory=lambda: PIL.Image.new("RGB", (32, 24)))
+    cameras: tuple[str, ...] = ("observation.images.top",)
+    calls: list[tuple[int, tuple[float, ...], str | None]] = field(default_factory=list)
+    video_calls: list[tuple[int, int, str | None]] = field(default_factory=list)
+
+    @property
+    def camera_keys(self) -> list[str]:
+        return list(self.cameras)
+
+    def frames_at(self, record, timestamps, camera_key=None):
+        self.calls.append((record.episode_index, tuple(timestamps), camera_key))
+        return [self.sentinel] * len(timestamps)
+
+    def video_for_episode(self, record, max_frames, camera_key=None):
+        self.video_calls.append((record.episode_index, max_frames, camera_key))
+        n = min(max_frames, len(record.frame_timestamps))
+        return [self.sentinel] * n
+
+
+def _spy_responder(captured: list[list[dict[str, Any]]], reply: Any):
+    def responder(messages):
+        captured.append(list(messages))
+        return reply
+
+    return StubVlmClient(responder=responder)
+
+
+def test_module1_plan_memory_subtask_smoke(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    vlm = make_canned_responder(
+        {
+            "atomic subtasks": {
+                "subtasks": [
+                    {"text": "grasp the handle of the sponge", "start": 0.0, "end": 0.4},
+                    {"text": "wipe the counter from left to right", "start": 0.4, "end": 0.8},
+                    {"text": "place the sponge into the sink", "start": 0.8, "end": 1.1},
+                ]
+            },
+            "compressed semantic memory": {"memory": "wiped the counter once"},
+        },
+    )
+    module = PlanSubtasksMemoryModule(vlm=vlm, config=PlanConfig())
+    record = next(iter_episodes(fixture_dataset_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    module.run_episode(record, staging)
+    rows = staging.read("plan")
+
+    styles = {r["style"] for r in rows}
+    assert {"subtask", "plan", "memory"}.issubset(styles)
+    # subtask timestamps must be exact frame timestamps
+    frame_set = set(record.frame_timestamps)
+    for row in rows:
+        assert row["timestamp"] in frame_set
+    # one plan row per subtask boundary; the first lands at t0 and each
+    # plan is the deterministic numbered list of still-todo subtasks
+    plan_rows = sorted((r for r in rows if r["style"] == "plan"), key=lambda r: r["timestamp"])
+    subtask_rows = [r for r in rows if r["style"] == "subtask"]
+    assert len(plan_rows) == len(subtask_rows)
+    assert plan_rows[0]["timestamp"] == record.frame_timestamps[0]
+    # the t0 plan enumerates all subtasks; later plans shrink
+    assert plan_rows[0]["content"].startswith("1. ")
+    assert len(plan_rows[0]["content"].splitlines()) == len(subtask_rows)
+    assert len(plan_rows[-1]["content"].splitlines()) == 1
+
+
+def test_module1_emit_memory_false_skips_memory_keeps_subtasks_and_plan(
+    fixture_dataset_root: Path, tmp_path: Path
+) -> None:
+    """``emit_memory=False`` drops ``memory`` rows (and their VLM calls) while
+    leaving subtask + plan generation intact — symmetric to ``emit_plan``."""
+    vlm = make_canned_responder(
+        {
+            "atomic subtasks": {
+                "subtasks": [
+                    {"text": "grasp the handle of the sponge", "start": 0.0, "end": 0.4},
+                    {"text": "wipe the counter from left to right", "start": 0.4, "end": 0.8},
+                    {"text": "place the sponge into the sink", "start": 0.8, "end": 1.1},
+                ]
+            },
+            "compressed semantic memory": {"memory": "wiped the counter once"},
+        },
+    )
+    module = PlanSubtasksMemoryModule(vlm=vlm, config=PlanConfig(emit_memory=False))
+    record = next(iter_episodes(fixture_dataset_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    module.run_episode(record, staging)
+    rows = staging.read("plan")
+
+    styles = {r["style"] for r in rows}
+    assert "memory" not in styles
+    assert {"subtask", "plan"}.issubset(styles)
+
+
+def test_module2_at_t0_emits_speech_only_no_interjection(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    vlm = make_canned_responder(
+        {"acknowledgement the robot": {"text": "Sure, on it."}},
+    )
+    module = InterjectionsAndSpeechModule(
+        vlm=vlm,
+        config=InterjectionsConfig(max_interjections_per_episode=0),
+    )
+    record = next(iter_episodes(fixture_dataset_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    module.run_episode(record, staging)
+    rows = staging.read("interjections")
+    assert len(rows) == 1
+    only = rows[0]
+    assert only["role"] == "assistant"
+    assert only["style"] is None
+    assert only["content"] is None
+    assert only["timestamp"] == record.frame_timestamps[0]
+    assert only["tool_calls"][0]["function"]["name"] == "say"
+
+
+def test_module2_mid_episode_emits_paired_interjection_and_speech(
+    fixture_dataset_root: Path, tmp_path: Path
+) -> None:
+    """Module 2 anchors interjections on Module 1's subtask boundaries.
+
+    The executor runs Module 1 first, then Module 2 reads the subtask
+    rows back from the same staging tree (see
+    ``_mid_episode_interjections``). Reproduce that contract here by
+    seeding the staging with two subtask rows so a single ``0 → 1``
+    boundary exists for Module 2 to anchor on.
+    """
+    vlm = make_canned_responder(
+        {
+            "acknowledgement the robot": {"text": "OK."},
+            # Marker matches the distinctive line of
+            # ``interjections_interjection.txt`` ("Write ONE compact
+            # interjection ..."). Keep this in sync with that prompt's
+            # wording — the canned responder matches on substring.
+            "Write ONE compact interjection": {
+                "interjection": "now wipe the counter please",
+                "speech": "On it.",
+            },
+        },
+    )
+    module = InterjectionsAndSpeechModule(
+        vlm=vlm,
+        config=InterjectionsConfig(max_interjections_per_episode=1, interjection_min_t=0.2),
+        seed=7,
+    )
+    record = next(iter_episodes(fixture_dataset_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    # Seed Module 1's subtask staging so Module 2 has a boundary to
+    # anchor on (it bails with zero rows when no spans exist — the
+    # production executor guarantees Module 1 ran first).
+    boundary_ts = float(record.frame_timestamps[len(record.frame_timestamps) // 2])
+    staging.write(
+        "plan",
+        [
+            {
+                "role": "assistant",
+                "content": "grasp the sponge",
+                "style": "subtask",
+                "timestamp": float(record.frame_timestamps[0]),
+                "tool_calls": None,
+            },
+            {
+                "role": "assistant",
+                "content": "wipe the counter",
+                "style": "subtask",
+                "timestamp": boundary_ts,
+                "tool_calls": None,
+            },
+        ],
+    )
+    module.run_episode(record, staging)
+    rows = staging.read("interjections")
+
+    interjections = [r for r in rows if r["style"] == "interjection"]
+    speeches = [r for r in rows if r["style"] is None and r["role"] == "assistant"]
+    assert len(interjections) == 1
+    assert len(speeches) >= 2  # initial t=0 + one paired with the interjection
+    inter_t = interjections[0]["timestamp"]
+    assert any(abs(s["timestamp"] - inter_t) < 1e-9 for s in speeches)
+
+
+def test_module3_vqa_unique_per_frame_and_camera(single_episode_root: Path, tmp_path: Path) -> None:
+    payload = {
+        "question": "How many cups?",
+        "answer": {"label": "cup", "count": 2, "note": "white & blue"},
+    }
+    vlm = make_canned_responder({"frame-grounded visual question": payload})
+    module = GeneralVqaModule(
+        vlm=vlm,
+        config=VqaConfig(vqa_emission_hz=1.0, K=3),
+        seed=1,
+        frame_provider=_StubFrameProvider(cameras=("observation.images.top", "observation.images.wrist")),
+    )
+    record = next(iter_episodes(single_episode_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    module.run_episode(record, staging)
+    rows = staging.read("vqa")
+    # every vqa row must carry a camera tag and one of the configured cameras
+    for r in rows:
+        assert r["style"] == "vqa"
+        assert r.get("camera") in {"observation.images.top", "observation.images.wrist"}
+    # at most one (vqa, user) and one (vqa, assistant) per (timestamp, camera)
+    user_keys = [(r["timestamp"], r["camera"]) for r in rows if r["role"] == "user" and r["style"] == "vqa"]
+    assistant_keys = [
+        (r["timestamp"], r["camera"]) for r in rows if r["role"] == "assistant" and r["style"] == "vqa"
+    ]
+    assert len(user_keys) == len(set(user_keys))
+    assert len(assistant_keys) == len(set(assistant_keys))
+    # both cameras must be represented
+    assert {c for _, c in user_keys} == {"observation.images.top", "observation.images.wrist"}
+    # every emitted timestamp must be an exact source frame timestamp
+    frame_set = set(record.frame_timestamps)
+    for ts, _ in user_keys + assistant_keys:
+        assert ts in frame_set
+
+
+def test_module1_attaches_contact_sheets_to_subtask_prompt(
+    fixture_dataset_root: Path, tmp_path: Path
+) -> None:
+    """Module 1 sends timestamped contact-sheet image blocks (not a raw video block)."""
+    captured: list[list[dict[str, Any]]] = []
+    payload = {
+        "subtasks": [
+            {"text": "grasp the handle of the sponge", "start": 0.0, "end": 0.5},
+            {"text": "wipe the counter", "start": 0.5, "end": 1.1},
+        ]
+    }
+    memory_payload = {"memory": "wiped once"}
+
+    def responder(messages):
+        captured.append(list(messages))
+        text = ""
+        for m in messages:
+            for block in m.get("content", []):
+                if isinstance(block, dict) and block.get("type") == "text":
+                    text = block.get("text", "")
+        if "compressed semantic memory" in text:
+            return memory_payload
+        return payload
+
+    provider = _StubFrameProvider()
+    module = PlanSubtasksMemoryModule(
+        vlm=StubVlmClient(responder=responder),
+        # Disable the rephrasings sub-prompt so the test's only video-bearing
+        # call is the subtask one — keeps the assertions below focused on
+        # ``_generate_subtasks`` rather than fighting the order of unrelated
+        # text-only Module-1 sub-prompts.
+        config=PlanConfig(frames_per_second=2.0, max_frames_per_prompt=60, n_task_rephrasings=0),
+        frame_provider=provider,
+    )
+    record = next(iter_episodes(fixture_dataset_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    module.run_episode(record, staging)
+
+    # Find the call carrying the subtask prompt rather than blindly taking
+    # captured[0] — Module 1 issues several sub-prompts and their order is
+    # not part of the contract.
+    assert captured, "no VLM calls made"
+
+    def _prompt_text(messages):
+        for m in messages:
+            for block in m.get("content", []):
+                if isinstance(block, dict) and block.get("type") == "text":
+                    return block.get("text", "")
+        return ""
+
+    subtask_calls = [m for m in captured if "atomic subtasks" in _prompt_text(m)]
+    assert len(subtask_calls) == 1, "expected exactly one subtask-prompt VLM call"
+    content = subtask_calls[0][0]["content"]
+    video_blocks = [b for b in content if isinstance(b, dict) and b.get("type") == "video"]
+    image_blocks = [b for b in content if isinstance(b, dict) and b.get("type") == "image"]
+    text_blocks = [b for b in content if isinstance(b, dict) and b.get("type") == "text"]
+    assert video_blocks == [], "contact-sheet mode must not emit a raw video block"
+    assert len(image_blocks) >= 1, f"expected >=1 contact-sheet image block, got {content}"
+    assert all(isinstance(b["image"], PIL.Image.Image) for b in image_blocks)
+    assert len(text_blocks) == 1
+    # the prompt is prefixed with the contact-sheet reading instructions
+    assert text_blocks[0]["text"].startswith("CONTACT SHEETS")
+    # frames were decoded for this episode at episode-relative timestamps
+    assert provider.calls and provider.calls[0][0] == record.episode_index
+
+
+def test_module3_attaches_frame_image_block_to_prompt(single_episode_root: Path, tmp_path: Path) -> None:
+    """Each VQA prompt must carry a single image block at the emission frame."""
+    captured: list[list[dict[str, Any]]] = []
+    payload = {
+        "question": "How many cups?",
+        "answer": {"label": "cup", "count": 1},
+    }
+    provider = _StubFrameProvider()
+    module = GeneralVqaModule(
+        vlm=_spy_responder(captured, payload),
+        config=VqaConfig(vqa_emission_hz=1.0, K=1),
+        seed=0,
+        frame_provider=provider,
+    )
+    record = next(iter_episodes(single_episode_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    module.run_episode(record, staging)
+
+    assert captured, "no VLM calls made"
+    for messages in captured:
+        content = messages[0]["content"]
+        image_blocks = [b for b in content if isinstance(b, dict) and b.get("type") == "image"]
+        text_blocks = [b for b in content if isinstance(b, dict) and b.get("type") == "text"]
+        assert len(image_blocks) == 1, f"expected 1 image block per VQA prompt, got {content}"
+        assert image_blocks[0]["image"] is provider.sentinel
+        assert len(text_blocks) == 1
+    # provider was called once per emission per camera with the exact emission timestamp
+    for ep_idx, ts_tuple, camera in provider.calls:
+        assert ep_idx == record.episode_index
+        assert len(ts_tuple) == 1
+        assert ts_tuple[0] in record.frame_timestamps
+        assert camera in provider.cameras
+
+
+def test_module3_assistant_content_is_valid_json(single_episode_root: Path, tmp_path: Path) -> None:
+    payload = {
+        "question": "Where is the cup?",
+        "answer": {"detections": [{"label": "cup", "bbox_format": "xyxy", "bbox": [10, 20, 50, 80]}]},
+    }
+    vlm = make_canned_responder({"frame-grounded visual question": payload})
+    module = GeneralVqaModule(
+        vlm=vlm,
+        config=VqaConfig(vqa_emission_hz=1.0, K=2),
+        seed=2,
+        frame_provider=_StubFrameProvider(),
+    )
+    record = next(iter_episodes(single_episode_root))
+    staging = EpisodeStaging(tmp_path / "stage", record.episode_index)
+    module.run_episode(record, staging)
+    rows = staging.read("vqa")
+    for row in rows:
+        if row["role"] == "assistant" and row["style"] == "vqa":
+            decoded = json.loads(row["content"])
+            assert "detections" in decoded

tests/annotations/test_pipeline_recipe_render.py

@@ -0,0 +1,183 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""End-to-end smoke: pipeline output → canonical recipe rendering."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+# ``pyarrow`` and the ``lerobot.datasets`` chain (-> the HF ``datasets``
+# library) only ship under the ``dataset`` extra. Skip this module in
+# tiers without it instead of erroring at import.
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
+
+import pyarrow.parquet as pq  # noqa: E402
+
+from lerobot.annotations.steerable_pipeline.config import (  # noqa: E402
+    AnnotationPipelineConfig,
+    InterjectionsConfig,
+    PlanConfig,
+    VqaConfig,
+)
+from lerobot.annotations.steerable_pipeline.executor import Executor  # noqa: E402
+from lerobot.annotations.steerable_pipeline.modules import (  # noqa: E402
+    GeneralVqaModule,
+    InterjectionsAndSpeechModule,
+    PlanSubtasksMemoryModule,
+)
+from lerobot.annotations.steerable_pipeline.validator import StagingValidator  # noqa: E402
+from lerobot.annotations.steerable_pipeline.writer import LanguageColumnsWriter  # noqa: E402
+from lerobot.configs.recipe import MessageTurn, TrainingRecipe  # noqa: E402
+from lerobot.datasets.language_render import render_sample  # noqa: E402
+
+from ._helpers import make_canned_responder  # noqa: E402
+
+
+def _build_style_blend_recipe() -> TrainingRecipe:
+    """Inline blend recipe that consumes every style this pipeline produces.
+
+    The language schema/DSL work used to ship
+    ``src/lerobot/configs/recipes/pi05_hirobot.yaml`` as a canonical
+    example, but that file was dropped during review. The contract this
+    test guards is "the recipe DSL can render non-empty messages from
+    pipeline output", which doesn't require a specific YAML — so we build
+    the equivalent blend in code.
+    """
+    return TrainingRecipe(
+        blend={
+            "low_level_execution": TrainingRecipe(
+                weight=0.35,
+                messages=[
+                    MessageTurn(
+                        role="user",
+                        content="${task}\nPlan: ${plan}\nMemory: ${memory}",
+                        stream="high_level",
+                    ),
+                    MessageTurn(role="assistant", content="${subtask}", stream="low_level", target=True),
+                ],
+            ),
+            "user_interjection_response": TrainingRecipe(
+                weight=0.16,
+                bindings={
+                    "speech": "emitted_at(t, role=assistant, tool_name=say)",
+                    "interjection": "emitted_at(t, style=interjection)",
+                },
+                messages=[
+                    MessageTurn(role="user", content="${task}", stream="high_level"),
+                    MessageTurn(
+                        role="user",
+                        content="${interjection}",
+                        stream="high_level",
+                        if_present="interjection",
+                    ),
+                    MessageTurn(
+                        role="assistant",
+                        content="${plan}",
+                        stream="high_level",
+                        target=True,
+                        if_present="plan",
+                        tool_calls_from="speech",
+                    ),
+                ],
+            ),
+        }
+    )
+
+
+def _build_executor() -> Executor:
+    vlm = make_canned_responder(
+        {
+            "atomic subtasks": {
+                "subtasks": [
+                    {"text": "grasp the bottle", "start": 0.0, "end": 0.5},
+                    {"text": "pour into the cup", "start": 0.5, "end": 1.0},
+                    {"text": "place the bottle down", "start": 1.0, "end": 1.5},
+                ]
+            },
+            "compressed semantic memory": {"memory": "poured once"},
+            "acknowledgement the robot": {"text": "Sure."},
+            "compact interjection": {
+                "interjection": "use less water",
+                "speech": "Using less water.",
+            },
+            "frame-grounded visual question": {
+                "question": "How many cups?",
+                "answer": {"label": "cup", "count": 1},
+            },
+        },
+    )
+    config = AnnotationPipelineConfig(
+        plan=PlanConfig(),
+        interjections=InterjectionsConfig(max_interjections_per_episode=1, interjection_min_t=0.5),
+        vqa=VqaConfig(vqa_emission_hz=1.0, K=2),
+    )
+    return Executor(
+        config=config,
+        plan=PlanSubtasksMemoryModule(vlm=vlm, config=config.plan),
+        interjections=InterjectionsAndSpeechModule(vlm=vlm, config=config.interjections, seed=config.seed),
+        vqa=GeneralVqaModule(vlm=vlm, config=config.vqa, seed=config.seed),
+        writer=LanguageColumnsWriter(),
+        validator=StagingValidator(),
+    )
+
+
+def test_canonical_recipe_renders_nonempty_from_pipeline_output(
+    single_episode_root: Path,
+) -> None:
+    executor = _build_executor()
+    summary = executor.run(single_episode_root)
+    # validator may emit warnings but no errors for the synthetic fixture
+    assert summary.validation_report.ok, summary.validation_report.summary()
+
+    table = pq.read_table(single_episode_root / "data" / "chunk-000" / "file-000.parquet")
+    persistent_lists = table.column("language_persistent").to_pylist()
+    events_lists = table.column("language_events").to_pylist()
+    timestamps = table.column("timestamp").to_pylist()
+
+    recipe = _build_style_blend_recipe()
+
+    rendered_any = False
+    for ts, persistent, events in zip(timestamps, persistent_lists, events_lists, strict=True):
+        result = render_sample(
+            recipe=recipe,
+            persistent=persistent,
+            events=events,
+            t=float(ts),
+            sample_idx=0,
+            dataset_ctx={"task": "Pour water from the bottle into the cup."},
+        )
+        if result is None:
+            continue
+        if result["messages"]:
+            rendered_any = True
+            assert result["target_message_indices"]
+            break
+    assert rendered_any, "recipe rendered no messages from pipeline output"
+
+    # Sanity: speech atom appears in events column intact
+    flat_events = [r for ev in events_lists for r in ev]
+    speech_rows = [r for r in flat_events if r.get("style") is None and r.get("role") == "assistant"]
+    assert speech_rows
+    say = speech_rows[0]["tool_calls"][0]
+    assert say["function"]["name"] == "say"
+    assert isinstance(say["function"]["arguments"]["text"], str)
+    # The pipeline does not write a ``tools`` column — the say schema lives
+    # as a constant (``SAY_TOOL_SCHEMA``) so the language row struct is the
+    # single source of truth for the v3.1 schema.
+    assert "tools" not in table.column_names

tests/annotations/test_validator.py

@@ -0,0 +1,133 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Validator behavior tests."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+# ``lerobot.annotations`` imports pull in ``lerobot.datasets`` (-> the HF
+# ``datasets`` library), which only ships under the ``dataset`` extra. Skip
+# this module in tiers without it instead of erroring at import.
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
+
+from lerobot.annotations.steerable_pipeline.reader import iter_episodes  # noqa: E402
+from lerobot.annotations.steerable_pipeline.staging import EpisodeStaging  # noqa: E402
+from lerobot.annotations.steerable_pipeline.validator import StagingValidator  # noqa: E402
+from lerobot.annotations.steerable_pipeline.writer import speech_atom  # noqa: E402
+
+
+def _validate(root: Path, staging_dir: Path):
+    records = list(iter_episodes(root))
+    return StagingValidator().validate(records, staging_dir)
+
+
+def test_validator_catches_misaligned_timestamps(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    staging_dir = tmp_path / "stage"
+    EpisodeStaging(staging_dir, 0).write(
+        "vqa",
+        [
+            {
+                "role": "assistant",
+                "content": json.dumps({"label": "cup", "count": 2}, sort_keys=True),
+                "style": "vqa",
+                "timestamp": 9.999,  # not on any 10 fps frame
+                "tool_calls": None,
+            }
+        ],
+    )
+    report = _validate(fixture_dataset_root, staging_dir)
+    assert not report.ok
+    assert any("does not match any source frame timestamp" in e for e in report.errors)
+
+
+def test_validator_catches_orphan_speech(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    staging_dir = tmp_path / "stage"
+    EpisodeStaging(staging_dir, 0).write(
+        "interjections",
+        [
+            speech_atom(0.0, "Got it."),
+            # interjection at 0.3s with NO paired speech
+            {
+                "role": "user",
+                "content": "skip it",
+                "style": "interjection",
+                "timestamp": 0.3,
+                "tool_calls": None,
+            },
+        ],
+    )
+    report = _validate(fixture_dataset_root, staging_dir)
+    assert not report.ok
+    assert any("paired speech" in e for e in report.errors)
+
+
+def test_validator_catches_inconsistent_plan_memory(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    staging_dir = tmp_path / "stage"
+    EpisodeStaging(staging_dir, 0).write(
+        "plan",
+        [
+            {
+                "role": "assistant",
+                "content": "1. do x",
+                "style": "plan",
+                "timestamp": 0.0,
+                "tool_calls": None,
+            },
+            {
+                "role": "assistant",
+                "content": "do x",
+                "style": "subtask",
+                "timestamp": 0.0,
+                "tool_calls": None,
+            },
+        ],
+    )
+    EpisodeStaging(staging_dir, 0).write(
+        "interjections",
+        [
+            speech_atom(0.0, "Got it."),
+            speech_atom(0.4, "Replanning."),
+            {
+                "role": "user",
+                "content": "replan",
+                "style": "interjection",
+                "timestamp": 0.4,
+                "tool_calls": None,
+            },
+        ],
+    )
+    report = _validate(fixture_dataset_root, staging_dir)
+    # missing co-timestamped plan refresh at 0.4s → error
+    assert not report.ok
+    assert any("co-timestamped plan update" in e for e in report.errors)
+
+
+def test_validator_catches_wrong_column(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    staging_dir = tmp_path / "stage"
+    EpisodeStaging(staging_dir, 0).write(
+        "plan",
+        [
+            {"role": "user", "content": "where?", "style": "vqa", "timestamp": 0.0, "tool_calls": None},
+        ],
+    )
+    report = _validate(fixture_dataset_root, staging_dir)
+    assert not report.ok
+    assert any("plan emitted style 'vqa'" in e or "must be persistent" in e for e in report.errors)

tests/annotations/test_vlm_client.py

@@ -0,0 +1,41 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Unit tests for ``vlm_client`` helpers."""
+
+from __future__ import annotations
+
+import pytest
+
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+
+from lerobot.annotations.steerable_pipeline.vlm_client import _bind_serve_port  # noqa: E402
+
+
+def test_bind_serve_port_substitutes_placeholder() -> None:
+    # The {port} placeholder is replaced everywhere it appears, regardless of
+    # parallel vs single server — the bug was the single-server path passing
+    # it through unsubstituted.
+    cmd = "vllm serve M --max-model-len 32768 --port {port}"
+    assert _bind_serve_port(cmd, 8000) == "vllm serve M --max-model-len 32768 --port 8000"
+
+
+def test_bind_serve_port_appends_when_missing() -> None:
+    assert _bind_serve_port("vllm serve M", 8001) == "vllm serve M --port 8001"
+
+
+def test_bind_serve_port_leaves_explicit_port_untouched() -> None:
+    cmd = "vllm serve M --port 9000"
+    assert _bind_serve_port(cmd, 8000) == cmd

tests/annotations/test_writer.py

@@ -0,0 +1,430 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Writer correctness tests."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+# ``pyarrow`` and the ``lerobot.annotations`` -> ``lerobot.datasets`` chain
+# (-> the HF ``datasets`` library) only ship under the ``dataset`` extra.
+# Skip this module in tiers without it instead of erroring at import.
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
+
+import pandas as pd  # noqa: E402
+import pyarrow.parquet as pq  # noqa: E402
+
+from lerobot.annotations.steerable_pipeline.reader import iter_episodes  # noqa: E402
+from lerobot.annotations.steerable_pipeline.staging import EpisodeStaging  # noqa: E402
+from lerobot.annotations.steerable_pipeline.writer import (  # noqa: E402
+    LanguageColumnsWriter,
+    speech_atom,
+)
+
+
+def _stage_episode(
+    staging_dir: Path,
+    episode_index: int,
+    *,
+    plan: list[dict] | None = None,
+    interjections: list[dict] | None = None,
+    vqa: list[dict] | None = None,
+) -> None:
+    staging = EpisodeStaging(staging_dir, episode_index)
+    if plan is not None:
+        staging.write("plan", plan)
+    if interjections is not None:
+        staging.write("interjections", interjections)
+    if vqa is not None:
+        staging.write("vqa", vqa)
+
+
+def test_writer_persistence_identity(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    """Every frame in an episode has a byte-identical persistent list."""
+    staging_dir = tmp_path / "stage"
+    _stage_episode(
+        staging_dir,
+        0,
+        plan=[
+            {
+                "role": "assistant",
+                "content": "grasp the sponge",
+                "style": "subtask",
+                "timestamp": 0.0,
+                "tool_calls": None,
+            },
+            {
+                "role": "assistant",
+                "content": "1. wipe\n2. dry",
+                "style": "plan",
+                "timestamp": 0.0,
+                "tool_calls": None,
+            },
+            {
+                "role": "assistant",
+                "content": "wiped the counter",
+                "style": "memory",
+                "timestamp": 0.5,
+                "tool_calls": None,
+            },
+        ],
+    )
+    records = list(iter_episodes(fixture_dataset_root))
+    LanguageColumnsWriter().write_all(records, staging_dir, fixture_dataset_root)
+
+    table = pq.read_table(fixture_dataset_root / "data" / "chunk-000" / "file-000.parquet")
+    persistent = table.column("language_persistent").to_pylist()
+    first = persistent[0]
+    assert first  # non-empty
+    for row in persistent:
+        assert row == first, "persistent slice must be byte-identical across all frames"
+
+
+def test_writer_events_exact_timestamp(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    staging_dir = tmp_path / "stage"
+    _stage_episode(
+        staging_dir,
+        0,
+        interjections=[
+            speech_atom(0.0, "Got it."),
+            {
+                "role": "user",
+                "content": "skip the dishes",
+                "style": "interjection",
+                "timestamp": 0.5,
+                "tool_calls": None,
+            },
+            speech_atom(0.5, "Skipping the dishes."),
+        ],
+    )
+    records = list(iter_episodes(fixture_dataset_root))
+    LanguageColumnsWriter().write_all(records, staging_dir, fixture_dataset_root)
+
+    table = pq.read_table(fixture_dataset_root / "data" / "chunk-000" / "file-000.parquet")
+    timestamps = table.column("timestamp").to_pylist()
+    events = table.column("language_events").to_pylist()
+    for ts, ev in zip(timestamps, events, strict=True):
+        if abs(ts - 0.0) < 1e-9:
+            assert any(r["role"] == "assistant" and r.get("style") is None for r in ev), ev
+        elif abs(ts - 0.5) < 1e-9:
+            assert any(r.get("style") == "interjection" for r in ev), ev
+            assert any(r.get("style") is None for r in ev), ev
+        else:
+            assert ev == []
+
+
+def test_writer_column_routing(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    staging_dir = tmp_path / "stage"
+    _stage_episode(
+        staging_dir,
+        0,
+        plan=[
+            {
+                "role": "assistant",
+                "content": "do X",
+                "style": "subtask",
+                "timestamp": 0.0,
+                "tool_calls": None,
+            },
+            {
+                "role": "assistant",
+                "content": "1. do X",
+                "style": "plan",
+                "timestamp": 0.0,
+                "tool_calls": None,
+            },
+            {
+                "role": "assistant",
+                "content": "did X",
+                "style": "memory",
+                "timestamp": 0.3,
+                "tool_calls": None,
+            },
+        ],
+        interjections=[
+            speech_atom(0.0, "OK"),
+            {
+                "role": "user",
+                "content": "wait",
+                "style": "interjection",
+                "timestamp": 0.2,
+                "tool_calls": None,
+            },
+            speech_atom(0.2, "Waiting"),
+        ],
+        vqa=[
+            {
+                "role": "user",
+                "content": "where is the cup?",
+                "style": "vqa",
+                "timestamp": 0.4,
+                "camera": "observation.images.front",
+                "tool_calls": None,
+            },
+            {
+                "role": "assistant",
+                "content": json.dumps(
+                    {"detections": [{"label": "cup", "bbox_format": "xyxy", "bbox": [1, 2, 3, 4]}]},
+                    sort_keys=True,
+                ),
+                "style": "vqa",
+                "timestamp": 0.4,
+                "camera": "observation.images.front",
+                "tool_calls": None,
+            },
+        ],
+    )
+    records = list(iter_episodes(fixture_dataset_root))
+    LanguageColumnsWriter().write_all(records, staging_dir, fixture_dataset_root)
+    table = pq.read_table(fixture_dataset_root / "data" / "chunk-000" / "file-000.parquet")
+
+    persistent = table.column("language_persistent").to_pylist()[0]
+    persistent_styles = {r["style"] for r in persistent}
+    assert persistent_styles == {"subtask", "plan", "memory"}
+
+    all_events = [r for ev in table.column("language_events").to_pylist() for r in ev]
+    event_styles = {r.get("style") for r in all_events}
+    assert event_styles == {None, "interjection", "vqa"}
+
+
+def test_writer_drops_subtask_index_idempotent(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    staging_dir = tmp_path / "stage"
+    _stage_episode(
+        staging_dir,
+        0,
+        plan=[
+            {
+                "role": "assistant",
+                "content": "do X",
+                "style": "subtask",
+                "timestamp": 0.0,
+                "tool_calls": None,
+            },
+        ],
+    )
+    records = list(iter_episodes(fixture_dataset_root))
+    writer = LanguageColumnsWriter()
+    writer.write_all(records, staging_dir, fixture_dataset_root)
+
+    path = fixture_dataset_root / "data" / "chunk-000" / "file-000.parquet"
+    table_a = pq.read_table(path)
+    assert "subtask_index" not in table_a.column_names
+    assert "language_persistent" in table_a.column_names
+    assert "language_events" in table_a.column_names
+    # The writer no longer emits a dataset-level ``tools`` column; the
+    # ``say`` tool schema lives as a code constant (``SAY_TOOL_SCHEMA``)
+    # so the parquet stays small and the pipeline doesn't extend the schema.
+    assert "tools" not in table_a.column_names
+
+    # second pass — must produce identical bytes for the language columns
+    records_again = list(iter_episodes(fixture_dataset_root))
+    writer.write_all(records_again, staging_dir, fixture_dataset_root)
+    table_b = pq.read_table(path)
+    assert (
+        table_a.column("language_persistent").to_pylist() == table_b.column("language_persistent").to_pylist()
+    )
+    assert table_a.column("language_events").to_pylist() == table_b.column("language_events").to_pylist()
+
+
+def test_writer_normalize_rejects_misrouted_persistent_style() -> None:
+    """``_normalize_persistent_row`` must reject any non-persistent style."""
+    from lerobot.annotations.steerable_pipeline.writer import _normalize_persistent_row
+
+    with pytest.raises(ValueError, match="non-persistent style"):
+        _normalize_persistent_row(
+            {"role": "assistant", "content": "oops", "style": "vqa", "timestamp": 0.0, "tool_calls": None}
+        )
+
+
+def test_writer_normalize_rejects_misrouted_event_style() -> None:
+    """``_normalize_event_row`` must reject any persistent style."""
+    from lerobot.annotations.steerable_pipeline.writer import _normalize_event_row
+
+    with pytest.raises(ValueError):
+        _normalize_event_row({"role": "assistant", "content": "oops", "style": "subtask", "tool_calls": None})
+
+
+def test_say_tool_schema_constant_is_well_formed() -> None:
+    """``SAY_TOOL_SCHEMA`` (and ``DEFAULT_TOOLS``) replace the parquet
+    ``tools`` column — chat-template consumers import them directly.
+    """
+    from lerobot.annotations.steerable_pipeline.writer import (
+        DEFAULT_TOOLS,
+        SAY_TOOL_SCHEMA,
+    )
+
+    assert DEFAULT_TOOLS == [SAY_TOOL_SCHEMA]
+    assert SAY_TOOL_SCHEMA["function"]["name"] == "say"
+    params = SAY_TOOL_SCHEMA["function"]["parameters"]
+    assert params["properties"]["text"]["type"] == "string"
+    assert params["required"] == ["text"]
+
+
+def test_writer_does_not_add_tools_column(fixture_dataset_root: Path, tmp_path: Path) -> None:
+    """Re-running on a parquet that already has a legacy ``tools`` column
+    must drop it cleanly so reruns converge to the v3.1 schema.
+    """
+    staging_dir = tmp_path / "stage"
+    _stage_episode(
+        staging_dir,
+        0,
+        plan=[
+            {"role": "assistant", "content": "x", "style": "subtask", "timestamp": 0.0, "tool_calls": None}
+        ],
+    )
+    records = list(iter_episodes(fixture_dataset_root))
+    LanguageColumnsWriter().write_all(records, staging_dir, fixture_dataset_root)
+    table = pq.read_table(fixture_dataset_root / "data" / "chunk-000" / "file-000.parquet")
+    assert "tools" not in table.column_names
+
+
+def test_annotation_metadata_sync_allows_non_streaming_load(
+    fixture_dataset_root: Path, tmp_path: Path
+) -> None:
+    """Annotated parquet columns must be declared in ``meta/info.json``.
+
+    ``LeRobotDataset`` loads non-streaming datasets by casting parquet
+    against metadata-derived HF features. If the annotation writer adds
+    language columns but metadata stays stale, that cast fails with a column
+    mismatch.
+    """
+    from lerobot.annotations.steerable_pipeline.executor import Executor
+    from lerobot.datasets.feature_utils import get_hf_features_from_features
+    from lerobot.datasets.io_utils import load_info, load_nested_dataset
+    from lerobot.datasets.language import LANGUAGE_EVENTS, LANGUAGE_PERSISTENT, language_feature_info
+
+    info_path = fixture_dataset_root / "meta" / "info.json"
+    info = json.loads(info_path.read_text())
+    info["features"] = {
+        "episode_index": {"dtype": "int64", "shape": (1,), "names": None},
+        "frame_index": {"dtype": "int64", "shape": (1,), "names": None},
+        "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
+        "task_index": {"dtype": "int64", "shape": (1,), "names": None},
+    }
+    info_path.write_text(json.dumps(info, indent=2))
+
+    staging_dir = tmp_path / "stage"
+    _stage_episode(
+        staging_dir,
+        0,
+        plan=[
+            {"role": "assistant", "content": "do X", "style": "subtask", "timestamp": 0.0, "tool_calls": None}
+        ],
+    )
+    records = list(iter_episodes(fixture_dataset_root))
+    LanguageColumnsWriter().write_all(records, staging_dir, fixture_dataset_root)
+
+    Executor._ensure_annotation_metadata_in_info(fixture_dataset_root)
+
+    synced = load_info(fixture_dataset_root)
+    for key, feature in language_feature_info().items():
+        assert synced["features"][key] == feature
+
+    hf_features = get_hf_features_from_features(synced["features"])
+    dataset = load_nested_dataset(fixture_dataset_root / "data", features=hf_features)
+
+    assert LANGUAGE_PERSISTENT in dataset.column_names
+    assert LANGUAGE_EVENTS in dataset.column_names
+    assert len(dataset) == 24
+
+
+def _build_packed_dataset(root: Path, episode_lengths: list[int], *, fps: int = 10) -> Path:
+    """Pack several episodes into a single shard (vs build_annotation_dataset's one-per-file),
+    so the writer's rewrite must re-emit one row group per episode instead of collapsing them."""
+    from lerobot.datasets.io_utils import write_tasks
+    from lerobot.utils.io_utils import write_json
+
+    data_dir = root / "data" / "chunk-000"
+    data_dir.mkdir(parents=True, exist_ok=True)
+
+    episode_index, frame_index, timestamp, task_index, subtask_index = [], [], [], [], []
+    for ep, length in enumerate(episode_lengths):
+        episode_index += [ep] * length
+        frame_index += list(range(length))
+        timestamp += [round(i / fps, 6) for i in range(length)]
+        task_index += [0] * length
+        subtask_index += [0] * length  # legacy column the writer must drop
+    pd.DataFrame(
+        {
+            "episode_index": episode_index,
+            "frame_index": frame_index,
+            "timestamp": timestamp,
+            "task_index": task_index,
+            "subtask_index": subtask_index,
+        }
+    ).to_parquet(data_dir / "file-000.parquet", index=False)
+
+    tasks_df = pd.DataFrame({"task_index": [0]}, index=pd.Index(["do the thing"], name="task"))
+    write_tasks(tasks_df, root)
+    write_json(
+        {"codebase_version": "v3.1", "fps": fps, "features": {}, "total_episodes": len(episode_lengths)},
+        root / "meta" / "info.json",
+    )
+    return root
+
+
+def test_writer_one_row_group_per_episode(tmp_path: Path) -> None:
+    """Rewriting a packed shard must keep one row group per episode, not collapse
+    every episode into a single giant row group."""
+    episode_lengths = [4, 6, 5]  # unequal lengths, all in one shard
+    root = _build_packed_dataset(tmp_path / "ds", episode_lengths)
+    shard = root / "data" / "chunk-000" / "file-000.parquet"
+    assert pq.ParquetFile(shard).metadata.num_row_groups == 1, "fixture should start collapsed"
+
+    staging_dir = tmp_path / "stage"
+    for ep in range(len(episode_lengths)):
+        _stage_episode(
+            staging_dir,
+            ep,
+            plan=[
+                {
+                    "role": "assistant",
+                    "content": f"subtask for ep {ep}",
+                    "style": "subtask",
+                    "timestamp": 0.0,
+                    "tool_calls": None,
+                }
+            ],
+        )
+
+    records = list(iter_episodes(root))
+    LanguageColumnsWriter().write_all(records, staging_dir, root)
+
+    # One row group per episode, with row counts matching the episode lengths.
+    md = pq.ParquetFile(shard).metadata
+    assert md.num_row_groups == len(episode_lengths)
+    assert [md.row_group(i).num_rows for i in range(md.num_row_groups)] == episode_lengths
+    # Language columns are still present after the per-episode rewrite.
+    table = pq.read_table(shard)
+    assert "language_persistent" in table.column_names
+    assert "language_events" in table.column_names
+
+
+def test_speech_atom_shape_matches_plan_spec() -> None:
+    atom = speech_atom(2.5, "I'm cleaning up!")
+    assert atom["role"] == "assistant"
+    assert atom["style"] is None
+    assert atom["content"] is None
+    assert atom["timestamp"] == 2.5
+    assert isinstance(atom["tool_calls"], list)
+    call = atom["tool_calls"][0]
+    assert call["type"] == "function"
+    assert call["function"]["name"] == "say"
+    assert call["function"]["arguments"]["text"] == "I'm cleaning up!"

tests/datasets/test_aggregate.py

@@ -32,6 +32,26 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from tests.fixtures.constants import DUMMY_REPO_ID
 
 
+def assert_data_shards_one_row_group_per_episode(root):
+    """Every aggregated DATA shard must have exactly one parquet row group per episode."""
+    import pyarrow.parquet as pq
+
+    shards = sorted((root / "data").rglob("*.parquet"))
+    assert shards, f"no data shards found under {root}/data"
+    n_episodes = 0
+    for shard in shards:
+        pf = pq.ParquetFile(shard)
+        episodes = pf.read(columns=["episode_index"]).column("episode_index").to_pylist()
+        assert pf.metadata.num_row_groups == len(set(episodes)), shard
+        for i in range(pf.metadata.num_row_groups):
+            rg_episodes = set(
+                pf.read_row_group(i, columns=["episode_index"]).column("episode_index").to_pylist()
+            )
+            assert len(rg_episodes) == 1, f"{shard} row group {i} spans episodes {rg_episodes}"
+        n_episodes += len(set(episodes))
+    return n_episodes
+
+
 def assert_episode_and_frame_counts(aggr_ds, expected_episodes, expected_frames):
     """Test that total number of episodes and frames are correctly aggregated."""
     assert aggr_ds.num_episodes == expected_episodes, (
@@ -289,6 +309,52 @@ def test_aggregate_datasets(tmp_path, lerobot_dataset_factory):
     assert_dataset_iteration_works(aggr_ds)
 
 
+def test_aggregate_datasets_without_concatenation(tmp_path, lerobot_dataset_factory):
+    """With concatenation disabled, each source file is kept as its own destination file."""
+    ds_0 = lerobot_dataset_factory(
+        root=tmp_path / "no_stitch_0",
+        repo_id=f"{DUMMY_REPO_ID}_no_stitch_0",
+        total_episodes=3,
+        total_frames=60,
+    )
+    ds_1 = lerobot_dataset_factory(
+        root=tmp_path / "no_stitch_1",
+        repo_id=f"{DUMMY_REPO_ID}_no_stitch_1",
+        total_episodes=4,
+        total_frames=80,
+    )
+
+    aggr_root = tmp_path / "no_stitch_aggr"
+    aggregate_datasets(
+        repo_ids=[ds_0.repo_id, ds_1.repo_id],
+        roots=[ds_0.root, ds_1.root],
+        aggr_repo_id=f"{DUMMY_REPO_ID}_no_stitch_aggr",
+        aggr_root=aggr_root,
+        concatenate_videos=False,
+        concatenate_data=False,
+    )
+
+    with (
+        patch("lerobot.datasets.dataset_metadata.get_safe_version") as mock_get_safe_version,
+        patch("lerobot.datasets.dataset_metadata.snapshot_download") as mock_snapshot_download,
+    ):
+        mock_get_safe_version.return_value = "v3.0"
+        mock_snapshot_download.return_value = str(aggr_root)
+        aggr_ds = LeRobotDataset(f"{DUMMY_REPO_ID}_no_stitch_aggr", root=aggr_root)
+
+    assert_episode_and_frame_counts(
+        aggr_ds, ds_0.num_episodes + ds_1.num_episodes, ds_0.num_frames + ds_1.num_frames
+    )
+    assert_dataset_iteration_works(aggr_ds)
+    assert_video_timestamps_within_bounds(aggr_ds)
+
+    # Two single-file sources stay as two files each, instead of being packed together.
+    assert len(list((aggr_root / "data").rglob("*.parquet"))) == 2
+    assert aggr_ds.meta.video_keys, "Test fixture should produce at least one video feature"
+    for key in aggr_ds.meta.video_keys:
+        assert len(list((aggr_root / "videos" / key).rglob("*.mp4"))) == 2
+
+
 @pytest.mark.parametrize("mutation", ["mismatched_value", "missing_key"])
 def test_aggregate_incomplete_video_encoder_info_warns_and_nuls_encoders(
     tmp_path, lerobot_dataset_factory, caplog, mutation
@@ -520,6 +586,41 @@ def assert_image_frames_integrity(aggr_ds, ds_0, ds_1):
             )
 
 
+@pytest.mark.parametrize("use_videos", [True, False], ids=["video", "image"])
+def test_aggregate_one_row_group_per_episode(tmp_path, lerobot_dataset_factory, use_videos):
+    """Aggregated DATA shards keep one row group per episode (not one collapsed group).
+
+    Covers both the non-image (``df.to_parquet``) and image
+    (``to_parquet_with_hf_images``) write branches, including the merge-into-
+    existing-file branch via a low file-size threshold that forces packing.
+    """
+    ds_0 = lerobot_dataset_factory(
+        root=tmp_path / "rg_0",
+        repo_id=f"{DUMMY_REPO_ID}_rg_0",
+        total_episodes=3,
+        total_frames=60,
+        use_videos=use_videos,
+    )
+    ds_1 = lerobot_dataset_factory(
+        root=tmp_path / "rg_1",
+        repo_id=f"{DUMMY_REPO_ID}_rg_1",
+        total_episodes=4,
+        total_frames=80,
+        use_videos=use_videos,
+    )
+
+    aggr_root = tmp_path / "rg_aggr"
+    aggregate_datasets(
+        repo_ids=[ds_0.repo_id, ds_1.repo_id],
+        roots=[ds_0.root, ds_1.root],
+        aggr_repo_id=f"{DUMMY_REPO_ID}_rg_aggr",
+        aggr_root=aggr_root,
+    )
+
+    n_episodes = assert_data_shards_one_row_group_per_episode(aggr_root)
+    assert n_episodes == ds_0.num_episodes + ds_1.num_episodes
+
+
 def test_aggregate_image_datasets(tmp_path, lerobot_dataset_factory):
     """Test aggregation of image-based datasets preserves HuggingFace Image schema.
 

tests/datasets/test_compute_stats.py

@@ -83,6 +83,29 @@ def test_get_feature_stats_images():
     assert stats["min"].shape == stats["max"].shape == stats["mean"].shape == stats["std"].shape
 
 
+def test_get_feature_stats_uint8_images_preserves_std():
+    data = np.array(
+        [
+            [
+                [[0, 64], [128, 255]],
+                [[255, 128], [64, 0]],
+                [[32, 96], [160, 224]],
+            ],
+            [
+                [[16, 80], [144, 240]],
+                [[240, 144], [80, 16]],
+                [[48, 112], [176, 208]],
+            ],
+        ],
+        dtype=np.uint8,
+    )
+
+    stats = get_feature_stats(data, axis=(0, 2, 3), keepdims=True)
+
+    expected_std = data.transpose(0, 2, 3, 1).reshape(-1, 3).std(axis=0).reshape(1, 3, 1, 1)
+    np.testing.assert_allclose(stats["std"], expected_std)
+
+
 def test_get_feature_stats_axis_0_keepdims(sample_array):
     expected = {
         "min": np.array([[1, 2, 3]]),

tests/datasets/test_datasets.py

@@ -51,7 +51,7 @@ from lerobot.robots import make_robot_from_config
 from lerobot.transforms import ImageTransforms, ImageTransformsConfig
 from lerobot.utils.constants import ACTION, DONE, OBS_IMAGES, OBS_STATE, OBS_STR, REWARD
 from lerobot.utils.feature_utils import hw_to_dataset_features
-from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_REPO_ID
+from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_MOTOR_FEATURES, DUMMY_REPO_ID
 from tests.mocks.mock_robot import MockRobotConfig
 from tests.utils import require_x86_64_kernel
 
@@ -133,6 +133,21 @@ def test_dataset_feature_with_forward_slash_raises_error():
         )
 
 
+def test_create_does_not_mutate_input_features(tmp_path, empty_lerobot_dataset_factory):
+    # ``create`` must deep-copy features so a dataset built from another's features stays independent.
+    dataset = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds1", features=DUMMY_MOTOR_FEATURES, use_videos=False
+    )
+    dataset_copy = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds2", features=dataset.meta.features, use_videos=False
+    )
+
+    original_shape = dataset.meta.info.features["state"]["shape"]
+    dataset_copy.meta.info.features["state"]["shape"] = (999,)
+
+    assert dataset.meta.info.features["state"]["shape"] == original_shape
+
+
 def test_add_frame_missing_task(tmp_path, empty_lerobot_dataset_factory):
     features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
     dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)

tests/datasets/test_sampler.py

@@ -114,6 +114,19 @@ def test_shuffle():
     assert set(sampler) == {0, 1, 2, 3, 4, 5}
 
 
+def test_shuffle_is_reproducible_across_instances():
+    # The order is a pure function of (seed, epoch), so two fresh samplers (e.g. two ranks)
+    # produce the same permutation without any generator synchronization.
+    sampler_a = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0
+    # Desyncing the global RNG must not affect the permutation.
+    sampler_c = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    torch.randperm(1000)  # consume global RNG, as rank-asymmetric code (e.g. eval) would
+    assert list(sampler_c) == epoch_0
+
+
 def test_negative_drop_first_frames_raises():
     with pytest.raises(ValueError, match="drop_n_first_frames must be >= 0"):
         EpisodeAwareSampler([0], [10], drop_n_first_frames=-1)
@@ -137,3 +150,87 @@ def test_partial_episode_drop_warns(caplog):
     # Episode 0 is skipped (1 frame, drop 1), Episode 1 keeps frames 2-5
     assert sampler.indices == [2, 3, 4, 5]
     assert "Episode 0" in caplog.text
+
+
+# --- seeded (seed, epoch) shuffling, resume, and state ---
+
+from lerobot.datasets.sampler import compute_sampler_state  # noqa: E402
+
+EPISODE_BOUNDS = ([0, 2, 3], [2, 3, 6])  # episodes of 2, 1 and 3 frames
+
+
+@pytest.mark.parametrize("num_frames", [1, 2, 3, 37, 64, 100])
+def test_deterministic_sampler_shuffle_is_permutation(num_frames):
+    for seed in (0, 1, 1234):
+        sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=seed)
+        assert sorted(sampler) == list(range(num_frames))
+
+
+def test_deterministic_sampler_epochs_reproduce_and_differ():
+    sampler_a = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0  # same (seed, epoch) -> same order on any process
+    epoch_1 = list(sampler_a)  # __iter__ auto-advances the epoch
+    assert epoch_1 != epoch_0
+    assert sorted(epoch_1) == sorted(epoch_0)
+    sampler_a.set_epoch(0)
+    assert list(sampler_a) == epoch_0
+    assert list(EpisodeAwareSampler([0], [100], shuffle=True, seed=7)) != epoch_0
+
+
+def test_deterministic_sampler_resume_mid_epoch():
+    reference = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+    epoch_0 = list(reference)
+    epoch_1 = list(reference)
+    for start in (0, 1, 4, len(epoch_0)):
+        resumed = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+        resumed.load_state_dict({"epoch": 0, "start_index": start})
+        assert list(resumed) == epoch_0[start:]
+        # the resumed sampler continues into the same epoch 1 as the uninterrupted one
+        assert list(resumed) == epoch_1
+
+
+def test_deterministic_sampler_construction_stores_only_boundaries():
+    # Construction is O(num_episodes), not O(num_frames): a million-frame single episode
+    # instantiates from just its boundaries without materializing a per-frame index list.
+    num_frames = 1_000_000
+    sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    assert len(sampler) == num_frames
+    assert sampler._starts.shape == (1,) and sampler._cum_lengths.shape == (1,)
+
+
+def test_deterministic_sampler_resume_is_exact_at_scale():
+    # Seeded randperm makes resume sample-exact at non-trivial sizes: regenerating the epoch's
+    # permutation and slicing from the saved offset reproduces the remaining order exactly.
+    num_frames = 100_000
+    reference = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    epoch_0 = list(reference)
+    assert sorted(epoch_0) == list(range(num_frames))
+    start = num_frames - 5
+    resumed = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    resumed.load_state_dict({"epoch": 0, "start_index": start})
+    assert list(resumed) == epoch_0[start:]
+
+
+def test_compute_sampler_state():
+    # 100 frames, batch 10, 2 ranks -> 10 underlying batches, 5 per rank per epoch.
+    assert compute_sampler_state(step=0, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 0,
+        "start_index": 0,
+    }
+    # step 7 -> epoch 1, 2 per-rank batches in = 2 * 10 * 2 = 40 samples in
+    assert compute_sampler_state(step=7, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 40,
+    }
+    # uneven epoch: 95 frames -> 10 underlying batches (last short), still 5 per rank
+    assert compute_sampler_state(step=12, num_frames=95, batch_size=10, num_processes=2) == {
+        "epoch": 2,
+        "start_index": 40,
+    }
+    # uneven sharding: 105 frames -> 11 underlying batches, 6 per rank (even_batches pads)
+    assert compute_sampler_state(step=11, num_frames=105, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 100,
+    }

tests/datasets/test_video_encoding.py

@@ -504,6 +504,19 @@ class TestReencodeVideo:
         assert info["video.g"] == 6
         assert info["video.crf"] == 23
 
+    @require_h264
+    def test_reencode_video_trim_window(self, tmp_path):
+        src = TEST_ARTIFACTS_DIR / "clip_6frames.mp4"
+        out = tmp_path / "trim_window.mp4"
+        cfg = VideoEncoderConfig(vcodec="h264")
+        reencode_video(src, out, camera_encoder=cfg, start_time_s=0.05, end_time_s=0.12, overwrite=True)
+
+        with av.open(str(out)) as container:
+            frames = list(container.decode(video=0))
+        # Only the frames at 0.067 and 0.1 s fall inside [0.05, 0.12).
+        assert len(frames) == 2
+        assert frames[0].time == pytest.approx(0.0, abs=1e-3)
+
 
 class TestConcatenateVideoFiles:
     def test_two_clips_frame_count(self, tmp_path):

tests/fixtures/dataset_factories.py

@@ -552,3 +552,64 @@ def lerobot_dataset_factory(
 @pytest.fixture(scope="session")
 def empty_lerobot_dataset_factory() -> LeRobotDatasetFactory:
     return partial(LeRobotDataset.create, repo_id=DUMMY_REPO_ID, fps=DEFAULT_FPS)
+
+
+def build_annotation_dataset(
+    root: Path,
+    episode_specs: list[tuple[int, int, str]],
+    *,
+    fps: int = 10,
+) -> Path:
+    """Build a minimal LeRobot-shaped dataset on disk for annotation tests.
+
+    ``episode_specs`` is a list of ``(episode_index, num_frames, task_text)``.
+    Each episode is written to its own
+    ``data/chunk-000/file-{ep:03d}.parquet`` so the writer's per-shard
+    rewrite path is exercised. The dataset carries the minimum
+    ``meta/tasks.parquet`` + ``meta/info.json`` the reader / executor need;
+    it has no videos, so the modules fall back to text-only prompts.
+
+    Shared by the annotation-pipeline pytest fixtures (``tests/annotations/
+    conftest.py``) and the opt-in E2E smoke run so the fixture shape lives
+    in exactly one place.
+    """
+    from lerobot.datasets.io_utils import write_tasks
+    from lerobot.utils.io_utils import write_json
+
+    data_dir = root / "data" / "chunk-000"
+    data_dir.mkdir(parents=True, exist_ok=True)
+
+    tasks: dict[int, str] = {}
+    for episode_index, num_frames, task_text in episode_specs:
+        if task_text not in tasks.values():
+            tasks[len(tasks)] = task_text
+        task_index = next(k for k, v in tasks.items() if v == task_text)
+        frame = pd.DataFrame(
+            {
+                "episode_index": [episode_index] * num_frames,
+                "frame_index": list(range(num_frames)),
+                "timestamp": [round(i / fps, 6) for i in range(num_frames)],
+                "task_index": [task_index] * num_frames,
+                "subtask_index": [0] * num_frames,  # legacy column the writer must drop
+            }
+        )
+        frame.to_parquet(data_dir / f"file-{episode_index:03d}.parquet", index=False)
+
+    # Canonical tasks frame: indexed by task string with a ``task_index``
+    # column, matching what ``lerobot.datasets.io_utils.load_tasks`` expects.
+    tasks_df = pd.DataFrame(
+        {"task_index": list(tasks.keys())},
+        index=pd.Index(list(tasks.values()), name="task"),
+    )
+    write_tasks(tasks_df, root)
+
+    write_json(
+        {
+            "codebase_version": "v3.1",
+            "fps": fps,
+            "features": {},
+            "total_episodes": len(episode_specs),
+        },
+        root / "meta" / "info.json",
+    )
+    return root

tests/processor/test_pipeline.py

@@ -2370,14 +2370,32 @@ def test_aggregate_images_when_use_videos_false():
     out = aggregate_pipeline_dataset_features(
         pipeline=rp,
         initial_features={PipelineFeatureType.ACTION: {}, PipelineFeatureType.OBSERVATION: initial},
-        use_videos=False,  # expect "image" dtype
+        use_videos=False,  # images kept, stored as "image" dtype
         patterns=None,
     )
 
     key = f"{OBS_IMAGES}.back"
     key_front = f"{OBS_IMAGES}.front"
-    assert key not in out
-    assert key_front not in out
+    assert key in out
+    assert key_front in out
+    assert out[key]["dtype"] == "image"
+    assert out[key_front]["dtype"] == "image"
+    assert out[key]["shape"] == initial["back"]
+
+
+def test_aggregate_images_excluded():
+    rp = DataProcessorPipeline([AddObservationStateFeatures(add_front_image=True)])
+    initial = {"back": (480, 640, 3)}
+
+    out = aggregate_pipeline_dataset_features(
+        pipeline=rp,
+        initial_features={PipelineFeatureType.ACTION: {}, PipelineFeatureType.OBSERVATION: initial},
+        exclude_images=True,
+        patterns=None,
+    )
+
+    assert f"{OBS_IMAGES}.back" not in out
+    assert f"{OBS_IMAGES}.front" not in out
 
 
 def test_aggregate_images_when_use_videos_true():

tests/scripts/test_edit_dataset_parsing.py

@@ -66,6 +66,20 @@ class TestOperationTypeParsing:
         with pytest.raises(ValueError, match="--new_repo_id is required for merge"):
             _validate_config(cfg)
 
+    @pytest.mark.parametrize("flag", ["concatenate_videos", "concatenate_data"])
+    def test_merge_concatenate_flag_defaults_true(self, flag):
+        cfg = parse_cfg(["--new_repo_id", "test/merged", "--operation.type", "merge"])
+        assert isinstance(cfg.operation, MergeConfig)
+        assert getattr(cfg.operation, flag) is True
+
+    @pytest.mark.parametrize("flag", ["concatenate_videos", "concatenate_data"])
+    def test_merge_concatenate_flag_can_be_disabled(self, flag):
+        cfg = parse_cfg(
+            ["--new_repo_id", "test/merged", "--operation.type", "merge", f"--operation.{flag}", "false"]
+        )
+        assert isinstance(cfg.operation, MergeConfig)
+        assert getattr(cfg.operation, flag) is False
+
     def test_non_merge_requires_repo_id(self):
         cfg = parse_cfg(["--operation.type", "delete_episodes"])
         with pytest.raises(ValueError, match="--repo_id is required for delete_episodes"):

tests/scripts/test_lerobot_annotate.py

@@ -0,0 +1,86 @@
+#!/usr/bin/env python
+
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+from types import SimpleNamespace
+
+import pytest
+
+# ``lerobot.scripts.lerobot_annotate`` (and the ``_push_to_hub`` path it
+# exercises) imports ``lerobot.datasets``, which only ships under the
+# ``dataset`` extra. Skip in tiers without it instead of erroring.
+pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
+
+
+def test_push_to_hub_tags_uploaded_dataset_revision(tmp_path, monkeypatch):
+    from lerobot.scripts.lerobot_annotate import _push_to_hub
+
+    root = tmp_path / "dataset"
+    (root / "meta").mkdir(parents=True)
+    (root / "meta" / "info.json").write_text(json.dumps({"codebase_version": "v3.0"}))
+
+    calls = {}
+
+    class FakeHfApi:
+        def create_repo(self, **kwargs):
+            calls["create_repo"] = kwargs
+
+        def upload_folder(self, **kwargs):
+            calls["upload_folder"] = kwargs
+            return SimpleNamespace(oid="abc123")
+
+        def delete_tag(self, repo_id, **kwargs):
+            import requests
+            from huggingface_hub.errors import RevisionNotFoundError
+
+            calls["delete_tag"] = {"repo_id": repo_id, **kwargs}
+            # Simulate the common case: no stale tag to delete.
+            raise RevisionNotFoundError("no such tag", response=requests.Response())
+
+        def create_tag(self, **kwargs):
+            calls["create_tag"] = kwargs
+
+    monkeypatch.setattr("huggingface_hub.HfApi", FakeHfApi)
+
+    cfg = SimpleNamespace(
+        repo_id="source/dataset",
+        new_repo_id="annotated/dataset",
+        push_private=True,
+        push_commit_message=None,
+    )
+
+    _push_to_hub(root, cfg)
+
+    assert calls["create_repo"] == {
+        "repo_id": "annotated/dataset",
+        "repo_type": "dataset",
+        "private": True,
+        "exist_ok": True,
+    }
+    assert calls["upload_folder"]["repo_id"] == "annotated/dataset"
+    # A stale tag (e.g. from a previous annotation run) is deleted first so
+    # the new tag always points at the upload we just made.
+    assert calls["delete_tag"] == {
+        "repo_id": "annotated/dataset",
+        "tag": "v3.0",
+        "repo_type": "dataset",
+    }
+    assert calls["create_tag"] == {
+        "repo_id": "annotated/dataset",
+        "tag": "v3.0",
+        "repo_type": "dataset",
+        "revision": "abc123",
+    }

tests/teleoperators/test_rebot_102_leader.py

@@ -18,7 +18,7 @@ from unittest.mock import MagicMock, patch
 
 import pytest
 
-from lerobot.teleoperators.bi_rebot_102_leader import BiRebotArm102Leader, BiRebotArm102LeaderConfig
+from lerobot.teleoperators.bi_rebot_102_leader import BiRebot102Leader, BiRebot102LeaderConfig
 from lerobot.teleoperators.rebot_102_leader import (
     RebotArm102Leader,
     RebotArm102LeaderConfig,
@@ -91,11 +91,11 @@ def test_send_feedback_not_implemented(leader):
 
 def test_bimanual_prefixes_features():
     with patch(f"{_MODULE}.require_package", lambda *a, **kw: None):
-        cfg = BiRebotArm102LeaderConfig(
+        cfg = BiRebot102LeaderConfig(
             left_arm_config=RebotArm102LeaderConfig(port="/dev/null0"),
             right_arm_config=RebotArm102LeaderConfig(port="/dev/null1"),
         )
-        teleop = BiRebotArm102Leader(cfg)
+        teleop = BiRebot102Leader(cfg)
     assert any(k.startswith("left_") for k in teleop.action_features)
     assert any(k.startswith("right_") for k in teleop.action_features)
     assert "left_gripper.pos" in teleop.action_features

tests/utils/test_logging_utils.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 
 import pytest
+import torch
 
 from lerobot.utils.logging_utils import AverageMeter, MetricsTracker
 
@@ -25,8 +26,16 @@ def mock_metrics():
 
 
 class MockAccelerator:
-    def __init__(self, num_processes: int):
+    def __init__(self, num_processes: int, reduce_fn=None):
         self.num_processes = num_processes
+        self.device = torch.device("cpu")
+        self._reduce_fn = reduce_fn
+
+    def reduce(self, tensor, reduction="mean"):
+        # In single-process tests we just want a deterministic stand-in for accelerate's reduce.
+        if self._reduce_fn is not None:
+            return self._reduce_fn(tensor, reduction)
+        return tensor
 
 
 def test_average_meter_initialization():
@@ -157,3 +166,70 @@ def test_metrics_tracker_reset_averages(mock_metrics):
     tracker.reset_averages()
     assert tracker.loss.avg == 0.0
     assert tracker.accuracy.avg == 0.0
+
+
+def test_average_meter_invalid_reduction():
+    with pytest.raises(ValueError):
+        AverageMeter("loss", reduction="median")
+
+
+def test_average_meter_reduction_stored():
+    meter = AverageMeter("updt_s", reduction="max")
+    assert meter.reduction == "max"
+
+
+def test_metrics_tracker_reduce_across_ranks_no_accelerator():
+    metrics = {"update_s": AverageMeter("update_s", reduction="max")}
+    tracker = MetricsTracker(batch_size=32, num_frames=1000, num_episodes=50, metrics=metrics)
+    tracker.update_s = 0.5
+    tracker.reduce_across_ranks()  # no-op without accelerator
+    assert tracker.update_s.avg == 0.5
+
+
+def test_metrics_tracker_reduce_across_ranks_single_process():
+    metrics = {"update_s": AverageMeter("update_s", reduction="max")}
+    tracker = MetricsTracker(
+        batch_size=32,
+        num_frames=1000,
+        num_episodes=50,
+        metrics=metrics,
+        accelerator=MockAccelerator(num_processes=1),
+    )
+    tracker.update_s = 0.5
+    tracker.reduce_across_ranks()  # no-op when world size is 1
+    assert tracker.update_s.avg == 0.5
+
+
+def test_metrics_tracker_reduce_across_ranks_invokes_reduce():
+    captured = {}
+
+    def fake_reduce(tensor, reduction):
+        captured["reduction"] = reduction
+        captured["values"] = tensor.clone()
+        # Pretend the slowest rank reported 0.9 instead of this rank's 0.4.
+        return torch.tensor([0.9], dtype=tensor.dtype, device=tensor.device)
+
+    metrics = {
+        "loss": AverageMeter("loss"),  # reduction="none" -> not touched
+        "update_s": AverageMeter("update_s", reduction="max"),
+    }
+    tracker = MetricsTracker(
+        batch_size=32,
+        num_frames=1000,
+        num_episodes=50,
+        metrics=metrics,
+        accelerator=MockAccelerator(num_processes=4, reduce_fn=fake_reduce),
+    )
+    tracker.loss = 1.0
+    tracker.update_s = 0.4
+    tracker.reduce_across_ranks()
+
+    assert captured["reduction"] == "max"
+    assert torch.allclose(captured["values"], torch.tensor([0.4]))
+    assert tracker.update_s.avg == pytest.approx(0.9)
+    # Metrics without a reduction stay untouched.
+    assert tracker.loss.avg == 1.0
+    # Invariant: avg == sum / count must hold after reduce, so subsequent .update() calls
+    # accumulate against the cluster view rather than the stale per-rank sum.
+    meter = tracker.update_s
+    assert meter.sum / meter.count == pytest.approx(meter.avg)

tests/utils/test_train_utils.py

@@ -20,6 +20,8 @@ from unittest.mock import Mock, patch
 from lerobot.common.train_utils import (
     get_step_checkpoint_dir,
     get_step_identifier,
+    load_training_batch_size,
+    load_training_num_processes,
     load_training_state,
     load_training_step,
     save_checkpoint,
@@ -63,6 +65,28 @@ def test_load_training_step(tmp_path):
     assert loaded_step == step
 
 
+def test_save_training_state_records_num_processes(tmp_path, optimizer, scheduler):
+    save_training_state(tmp_path, 10, optimizer, scheduler, num_processes=4)
+    assert load_training_num_processes(tmp_path) == 4
+
+
+def test_load_training_num_processes_absent_returns_none(tmp_path, optimizer, scheduler):
+    # Checkpoints written before the world size was recorded must still load (back-compat).
+    save_training_state(tmp_path, 10, optimizer, scheduler)
+    assert load_training_num_processes(tmp_path) is None
+
+
+def test_save_training_state_records_batch_size(tmp_path, optimizer, scheduler):
+    save_training_state(tmp_path, 10, optimizer, scheduler, batch_size=32)
+    assert load_training_batch_size(tmp_path) == 32
+
+
+def test_load_training_batch_size_absent_returns_none(tmp_path, optimizer, scheduler):
+    # Checkpoints written before the batch size was recorded must still load (back-compat).
+    save_training_state(tmp_path, 10, optimizer, scheduler)
+    assert load_training_batch_size(tmp_path) is None
+
+
 def test_update_last_checkpoint(tmp_path):
     checkpoint = tmp_path / "0005"
     checkpoint.mkdir()