kiteml-cli 0.4.0__tar.gz → 0.5.0__tar.gz

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/PKG-INFO

@@ -1,10 +1,10 @@
 Metadata-Version: 2.5
 Name: kiteml-cli
-Version: 0.4.0
+Version: 0.5.0
 Summary: Kite CLI and MCP server: train robot policies with reinforcement learning in simulation, augment LeRobot datasets, and build MuJoCo digital twins, from your terminal or an AI agent.
 Project-URL: Homepage, https://kiteml.com
-Project-URL: Documentation, https://kite-ml.mintlify.app/tools/cli-and-mcp
-Project-URL: API reference, https://kite-ml.mintlify.app/platform-api/overview
+Project-URL: Documentation, https://docs.kiteml.com/tools/cli-and-mcp
+Project-URL: API reference, https://docs.kiteml.com/platform-api/overview
 Project-URL: Dashboard, https://app.kiteml.com
 Project-URL: MCP server, https://mcp.kiteml.com/mcp
 Project-URL: Blog, https://kiteml.com/blog
@@ -43,8 +43,8 @@ Description-Content-Type: text/markdown
 # Kite CLI and MCP server
 
 Robot learning on cloud GPUs, from your terminal or your AI agent. Kite trains robot policies with
-reinforcement learning in simulation, augments LeRobot datasets, and rebuilds robot episodes as MuJoCo digital
-twins. This package gives you the `kite` command line and the `kite-mcp` server for Claude, Cursor, and any
+reinforcement learning in simulation, evaluates policies on Kite SDK specs, augments LeRobot datasets, and rebuilds
+robot episodes as MuJoCo digital twins. This package gives you the `kite` command line and the `kite-mcp` server for Claude, Cursor, and any
 Model Context Protocol (MCP) client.
 
 <!-- mcp-name: com.kiteml/kite -->
@@ -52,11 +52,11 @@ Model Context Protocol (MCP) client.
 - **RL training in simulation.** Describe a behavior in plain English. Kite writes the task spec, checks its
   rewards for free, and trains the policy on a GPU. You get the policy as ONNX, the MuJoCo scene it trained in, and
   clips.
-- **Policy evaluation.** Every RL run ends with a report: a pass, needs_review or fail verdict from measured checks
+- **RL verdicts.** Every RL run ends with a report: a pass, needs_review or fail verdict from measured checks
   (falls per minute, survival, command tracking, gait, posture), plus a vision model's read of the clip.
-- **Dataset augmentation.** Relight a LeRobot dataset to match a deployment scene, or generate new variations of
-  it with video augmentation. Output is a standard LeRobot dataset (Parquet + MP4), downloaded or pushed to
-  Hugging Face.
+- **Dataset augmentation.** Relight a LeRobot dataset on Hugging Face or in a public S3 bucket to match a deployment
+  scene, or generate new variations of it with video augmentation. Output is a standard LeRobot dataset (Parquet +
+  MP4), downloaded or pushed to Hugging Face.
 - **Digital twins (beta).** Point Kite at one episode of a LeRobot dataset and get an interactable MuJoCo scene of the
   room, with the objects the robot handles built to size.
 - **Evaluations (beta).** Evaluate a policy by running a Kite SDK spec on Kite's machines, on an NVIDIA L4 for a GPU
@@ -106,7 +106,7 @@ lerobot/pusht to look like this photo".
 | --- | --- |
 | `kite_rl_catalog`, `kite_rl_plan`, `kite_rl_validate` | See which robots and objectives RL runs support, and plan and check a training spec for free |
 | `kite_rl_train`, `kite_rl_fork`, `kite_rl_cancel` | Start, branch from, and stop RL runs |
-| `kite_rl_status`, `kite_rl_metrics`, `kite_rl_report`, `kite_rl_list` | Follow training and read each run's evaluation verdict |
+| `kite_rl_status`, `kite_rl_metrics`, `kite_rl_report`, `kite_rl_list` | Follow training and read each run's verdict |
 | `kite_augment_create`, `kite_augment_status`, `kite_augment_list`, `kite_augment_cancel` | Augment LeRobot datasets |
 | `kite_twin_validate`, `kite_twin_create`, `kite_twin_status`, `kite_twin_list`, `kite_twin_cancel`, `kite_twin_resume` | Build MuJoCo digital twins (beta) |
 | `kite_eval_validate`, `kite_eval_run`, `kite_eval_status`, `kite_eval_list`, `kite_eval_report`, `kite_eval_logs`, `kite_eval_cancel`, `kite_eval_resume` | Evaluate a policy on a Kite SDK spec on Kite's machines and read the scorecard (beta) |
@@ -132,6 +132,9 @@ kite augment create --repo-id lerobot/pusht \
 kite augment download aug_... -o ./pusht-marble
 ```
 
+`--repo-id` takes a Hugging Face id (`lerobot/pusht`), its `huggingface.co/datasets/` URL, or an S3 link to the
+dataset's folder in a public bucket (`s3://bucket/dataset/`).
+
 Build a digital twin (beta, enabled per account):
 
 ```bash
@@ -145,16 +148,20 @@ kite eval validate specs/smoke.yaml            # free: the plan and what a run r
 kite eval run specs/smoke.yaml --out ./runs    # -> ./runs/eval_.../scorecard.json
 ```
 
+A file too big to send, such as a checkpoint's weights, can come from the Hugging Face Hub or a public S3 bucket:
+`--hf PATH=OWNER/NAME@COMMIT:FILE` or `--hf PATH=<the file's huggingface.co URL>`, and `--s3 PATH=s3://BUCKET/KEY`,
+PATH being the path the spec names.
+
 Every command prints JSON (`{"ok": true, "data": ...}`) so scripts and agents can parse it. Run `kite --help` for
 the full list.
 
 ## Links
 
-- Docs: [CLI and MCP](https://kite-ml.mintlify.app/tools/cli-and-mcp) ·
-  [RL runs](https://kite-ml.mintlify.app/platform-api/rl-runs) ·
-  [Augmentations](https://kite-ml.mintlify.app/platform-api/augmentation) ·
-  [Twins](https://kite-ml.mintlify.app/platform-api/twins) ·
-  [Evaluations](https://kite-ml.mintlify.app/platform-api/evaluations)
+- Docs: [CLI and MCP](https://docs.kiteml.com/tools/cli-and-mcp) ·
+  [RL runs](https://docs.kiteml.com/platform-api/rl-runs) ·
+  [Augmentations](https://docs.kiteml.com/platform-api/augmentation) ·
+  [Twins](https://docs.kiteml.com/platform-api/twins) ·
+  [Evaluations](https://docs.kiteml.com/platform-api/evaluations)
 - Dashboard: [app.kiteml.com](https://app.kiteml.com)
 - Website: [kiteml.com](https://kiteml.com)
 - Support: [raul@kiteml.com](mailto:raul@kiteml.com)

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/README.md

@@ -1,8 +1,8 @@
 # Kite CLI and MCP server
 
 Robot learning on cloud GPUs, from your terminal or your AI agent. Kite trains robot policies with
-reinforcement learning in simulation, augments LeRobot datasets, and rebuilds robot episodes as MuJoCo digital
-twins. This package gives you the `kite` command line and the `kite-mcp` server for Claude, Cursor, and any
+reinforcement learning in simulation, evaluates policies on Kite SDK specs, augments LeRobot datasets, and rebuilds
+robot episodes as MuJoCo digital twins. This package gives you the `kite` command line and the `kite-mcp` server for Claude, Cursor, and any
 Model Context Protocol (MCP) client.
 
 <!-- mcp-name: com.kiteml/kite -->
@@ -10,11 +10,11 @@ Model Context Protocol (MCP) client.
 - **RL training in simulation.** Describe a behavior in plain English. Kite writes the task spec, checks its
   rewards for free, and trains the policy on a GPU. You get the policy as ONNX, the MuJoCo scene it trained in, and
   clips.
-- **Policy evaluation.** Every RL run ends with a report: a pass, needs_review or fail verdict from measured checks
+- **RL verdicts.** Every RL run ends with a report: a pass, needs_review or fail verdict from measured checks
   (falls per minute, survival, command tracking, gait, posture), plus a vision model's read of the clip.
-- **Dataset augmentation.** Relight a LeRobot dataset to match a deployment scene, or generate new variations of
-  it with video augmentation. Output is a standard LeRobot dataset (Parquet + MP4), downloaded or pushed to
-  Hugging Face.
+- **Dataset augmentation.** Relight a LeRobot dataset on Hugging Face or in a public S3 bucket to match a deployment
+  scene, or generate new variations of it with video augmentation. Output is a standard LeRobot dataset (Parquet +
+  MP4), downloaded or pushed to Hugging Face.
 - **Digital twins (beta).** Point Kite at one episode of a LeRobot dataset and get an interactable MuJoCo scene of the
   room, with the objects the robot handles built to size.
 - **Evaluations (beta).** Evaluate a policy by running a Kite SDK spec on Kite's machines, on an NVIDIA L4 for a GPU
@@ -64,7 +64,7 @@ lerobot/pusht to look like this photo".
 | --- | --- |
 | `kite_rl_catalog`, `kite_rl_plan`, `kite_rl_validate` | See which robots and objectives RL runs support, and plan and check a training spec for free |
 | `kite_rl_train`, `kite_rl_fork`, `kite_rl_cancel` | Start, branch from, and stop RL runs |
-| `kite_rl_status`, `kite_rl_metrics`, `kite_rl_report`, `kite_rl_list` | Follow training and read each run's evaluation verdict |
+| `kite_rl_status`, `kite_rl_metrics`, `kite_rl_report`, `kite_rl_list` | Follow training and read each run's verdict |
 | `kite_augment_create`, `kite_augment_status`, `kite_augment_list`, `kite_augment_cancel` | Augment LeRobot datasets |
 | `kite_twin_validate`, `kite_twin_create`, `kite_twin_status`, `kite_twin_list`, `kite_twin_cancel`, `kite_twin_resume` | Build MuJoCo digital twins (beta) |
 | `kite_eval_validate`, `kite_eval_run`, `kite_eval_status`, `kite_eval_list`, `kite_eval_report`, `kite_eval_logs`, `kite_eval_cancel`, `kite_eval_resume` | Evaluate a policy on a Kite SDK spec on Kite's machines and read the scorecard (beta) |
@@ -90,6 +90,9 @@ kite augment create --repo-id lerobot/pusht \
 kite augment download aug_... -o ./pusht-marble
 ```
 
+`--repo-id` takes a Hugging Face id (`lerobot/pusht`), its `huggingface.co/datasets/` URL, or an S3 link to the
+dataset's folder in a public bucket (`s3://bucket/dataset/`).
+
 Build a digital twin (beta, enabled per account):
 
 ```bash
@@ -103,16 +106,20 @@ kite eval validate specs/smoke.yaml            # free: the plan and what a run r
 kite eval run specs/smoke.yaml --out ./runs    # -> ./runs/eval_.../scorecard.json
 ```
 
+A file too big to send, such as a checkpoint's weights, can come from the Hugging Face Hub or a public S3 bucket:
+`--hf PATH=OWNER/NAME@COMMIT:FILE` or `--hf PATH=<the file's huggingface.co URL>`, and `--s3 PATH=s3://BUCKET/KEY`,
+PATH being the path the spec names.
+
 Every command prints JSON (`{"ok": true, "data": ...}`) so scripts and agents can parse it. Run `kite --help` for
 the full list.
 
 ## Links
 
-- Docs: [CLI and MCP](https://kite-ml.mintlify.app/tools/cli-and-mcp) ·
-  [RL runs](https://kite-ml.mintlify.app/platform-api/rl-runs) ·
-  [Augmentations](https://kite-ml.mintlify.app/platform-api/augmentation) ·
-  [Twins](https://kite-ml.mintlify.app/platform-api/twins) ·
-  [Evaluations](https://kite-ml.mintlify.app/platform-api/evaluations)
+- Docs: [CLI and MCP](https://docs.kiteml.com/tools/cli-and-mcp) ·
+  [RL runs](https://docs.kiteml.com/platform-api/rl-runs) ·
+  [Augmentations](https://docs.kiteml.com/platform-api/augmentation) ·
+  [Twins](https://docs.kiteml.com/platform-api/twins) ·
+  [Evaluations](https://docs.kiteml.com/platform-api/evaluations)
 - Dashboard: [app.kiteml.com](https://app.kiteml.com)
 - Website: [kiteml.com](https://kiteml.com)
 - Support: [raul@kiteml.com](mailto:raul@kiteml.com)

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/pyproject.toml

@@ -43,8 +43,8 @@ dependencies = [
 
 [project.urls]
 Homepage = "https://kiteml.com"
-Documentation = "https://kite-ml.mintlify.app/tools/cli-and-mcp"
-"API reference" = "https://kite-ml.mintlify.app/platform-api/overview"
+Documentation = "https://docs.kiteml.com/tools/cli-and-mcp"
+"API reference" = "https://docs.kiteml.com/platform-api/overview"
 Dashboard = "https://app.kiteml.com"
 "MCP server" = "https://mcp.kiteml.com/mcp"
 Blog = "https://kiteml.com/blog"

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/src/kite_cli/__init__.py

@@ -1,3 +1,3 @@
 """Kite CLI and MCP server for the Kite robotics training platform."""
 
-__version__ = "0.4.0"
+__version__ = "0.5.0"

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/src/kite_cli/cli/commands/augment.py

@@ -38,10 +38,11 @@ def reference_image(path: str, camera: Optional[str] = None) -> dict:
 
 
 def scene_config(episodes, model: Optional[str], variants: Optional[int] = None,
-                 preview: Optional[bool] = None, allow_video_fallback: Optional[bool] = None) -> dict:
-    """The /v1 ``config`` for an episode selection ("all" or a count) and the scene engines' options: background /
-    world augment whole datasets and take ``episodes``; every other model takes a count as ``episode_count`` ("all"
-    sends none: relight covers every episode, video and auto keep their default)."""
+                 preview: Optional[bool] = None, allow_video_fallback: Optional[bool] = None,
+                 video_check: Optional[str] = None) -> dict:
+    """The /v1 ``config`` for an episode selection ("all" or a count), the scene engines' options and the video
+    check's mode: background / world augment whole datasets and take ``episodes``; every other model takes a count as
+    ``episode_count`` ("all" sends none: relight covers every episode, video and auto keep their default)."""
     config: dict = {}
     if episodes is not None:
         if str(episodes) != "all" and not str(episodes).isdigit():
@@ -56,12 +57,14 @@ def scene_config(episodes, model: Optional[str], variants: Optional[int] = None,
         config["preview_first"] = preview
     if allow_video_fallback:
         config["allow_video_fallback"] = True
+    if video_check:
+        config["video_check"] = video_check
     return config
 
 
 @app.command()
 def create(
-    repo_id: str = typer.Option(..., "--repo-id", help="Source HF dataset id, e.g. lerobot/pusht"),
+    repo_id: str = typer.Option(..., "--repo-id", help="Source LeRobot dataset: a HF id (lerobot/pusht) or huggingface.co/datasets/ URL, or an S3 link (s3://bucket/dataset/ or its https URL; public buckets)"),
     instructions: str = typer.Option(..., "-i", "--instructions", help="Augmentation instruction"),
     episodes: Optional[str] = typer.Option(
         None, "-n", "--episodes",
@@ -89,6 +92,10 @@ def create(
         help="auto: use generative video (far more tokens, no pixel guarantee) where no scene engine can run, "
         "instead of failing",
     ),
+    video_check: Optional[str] = typer.Option(
+        None, "--video-check", help="Generative video: standard (default: rejects a changed robot or a missing object; "
+        "the rest ship as warnings) | strict (every problem rejects) | off (records warnings, never rejects)",
+    ),
     reference_images: Optional[List[str]] = typer.Option(
         None, "--reference-image", help="Photo of the target look, [camera=]PATH (repeatable)"
     ),
@@ -107,7 +114,7 @@ def create(
     body: dict = {
         "source": {"type": "huggingface", "repo_id": repo_id},
         "instructions": instructions,
-        "config": scene_config(episodes, model, variants, preview, allow_video_fallback),
+        "config": scene_config(episodes, model, variants, preview, allow_video_fallback, video_check),
         "output": {"type": output, "private": private},
     }
     if hf_repo:

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/src/kite_cli/cli/commands/evaluations.py

@@ -16,7 +16,8 @@ machine otherwise. (Until 2026-10-18 this was `kite sim`, over /v1/sim_runs.)
 SPEC is the file `kite run` takes (kite_spec_version "2"). The command sends its text, asks the API which paths it
 names (`files_required`), and sends each from this disk: a file or folder Kite ships by its sha256 alone, anything
 else inline (about 15 MB in all). A file on the Hugging Face Hub, such as a checkpoint's weights, goes as a
-reference: `--hf PATH=OWNER/NAME@COMMIT:FILE`, PATH as the spec names it (a folder's file as FOLDER/FILE).
+reference: `--hf PATH=OWNER/NAME@COMMIT:FILE`, PATH as the spec names it (a folder's file as FOLDER/FILE); an object
+in a public S3 bucket as `--s3 PATH=s3://BUCKET/KEY`.
 Where runs go: --out DIR, else $KITE_EVAL_RUNS_DIR, else ./runs (`~` works). Paths printed are absolute. A run
 directory extracts to <out>/<run_id>/; the SDK's `kite verify <out>/<run_id> --evals DIR` checks it, DIR being the
 evaluators the spec named (the run records where its job read them, not your folder).
@@ -56,7 +57,10 @@ _HF = re.compile(r"(?P<repo>[A-Za-z0-9][A-Za-z0-9._-]*/[A-Za-z0-9][A-Za-z0-9._-]
                  r":(?P<path>.+)")
 
 _SET_HELP = "FIELD=VALUE, as `kite run --set` (repeatable)"
-_HF_HELP = "PATH=OWNER/NAME@COMMIT:FILE: the spec's PATH is that file of a public Hugging Face repo (repeatable)"
+_HF_HELP = ("PATH=OWNER/NAME@COMMIT:FILE, or PATH=URL (https://huggingface.co/OWNER/NAME/blob/REVISION/FILE; Kite pins a "
+            "branch to its commit): the spec's PATH is that file of a public Hugging Face repo (repeatable)")
+_S3_HELP = ("PATH=s3://BUCKET/KEY (or its https://BUCKET.s3.amazonaws.com/KEY URL): the spec's PATH is that object of a "
+            "public S3 bucket, pinned to the object it is when the run is created (repeatable)")
 
 
 @app.command()
@@ -64,12 +68,13 @@ def validate(
     spec: Path = typer.Argument(..., help="The spec file, e.g. specs/smoke.yaml"),
     set_: Optional[List[str]] = typer.Option(None, "--set", help=_SET_HELP),
     hf: Optional[List[str]] = typer.Option(None, "--hf", help=_HF_HELP),
+    s3: Optional[List[str]] = typer.Option(None, "--s3", help=_S3_HELP),
     kite_version: Optional[str] = typer.Option(None, "--kite-version", help="The Kite release you expect"),
 ):
     """Free dry run of a run: the spec plans, every file it names is sent, and the price. Creates nothing."""
     try:
         client = KiteClient(timeout=_SLOW_S)
-        _, validation, sent = request(client, spec, set_ or [], hf_refs(hf or []), kite_version)
+        _, validation, sent = request(client, spec, set_ or [], {**hf_refs(hf or []), **s3_refs(s3 or [])}, kite_version)
         print_json({"ok": True, "data": validation, "files": sent})
     except KiteAPIError as e:
         print_error(e.message, code=e.code, details=e.details)
@@ -80,6 +85,7 @@ def run(
     spec: Path = typer.Argument(..., help="The spec file, e.g. specs/smoke.yaml"),
     set_: Optional[List[str]] = typer.Option(None, "--set", help=_SET_HELP),
     hf: Optional[List[str]] = typer.Option(None, "--hf", help=_HF_HELP),
+    s3: Optional[List[str]] = typer.Option(None, "--s3", help=_S3_HELP),
     max_minutes: Optional[int] = typer.Option(
         None, "--max-minutes", help="The run's cap, 1-720: what it reserves and the most it can cost "
                                     "(default 60 on CPU, 120 on an L4)"),
@@ -100,7 +106,7 @@ def run(
     key = idempotency_key or str(uuid.uuid4())
     try:
         client = KiteClient(timeout=_SLOW_S)
-        body, _, _ = request(client, spec, set_ or [], hf_refs(hf or []), kite_version)
+        body, _, _ = request(client, spec, set_ or [], {**hf_refs(hf or []), **s3_refs(s3 or [])}, kite_version)
         options = {k: v for k, v in (("max_minutes", max_minutes), ("no_eval", no_eval),
                                      ("verify_determinism", verify_determinism)) if v}
         body = {**body, **({"options": options} if options else {}), **({"label": label} if label else {})}
@@ -242,10 +248,11 @@ def resume(run_id: str = typer.Argument(..., help="Evaluation id (eval_...)")):
 
 # --- the request: the spec and every path it names ------------------------------------------------------------------
 
-def request(client: KiteClient, spec: Path, overrides: List[str], hf: Dict[str, dict],
+def request(client: KiteClient, spec: Path, overrides: List[str], remote: Dict[str, dict],
             kite_version: Optional[str] = None) -> Tuple[dict, dict, Dict[str, str]]:
     """The request body for SPEC with every path it names, the validation that accepted it, and how each path went
-    (`sha256`, `inline`, `hf`).
+    (`sha256`, `inline`, `hf`, `s3`). `remote` maps the paths that are an `hf` file or an `s3` object to their
+    reference ({"hf": …} or {"s3": …}).
 
     The API names what it lacks: `files_required` lists the paths the spec names that the request does not carry,
     and `file_unavailable` on `files.<path>` says a hash Kite does not ship. A path first goes by its hash alone, so
@@ -253,12 +260,12 @@ def request(client: KiteClient, spec: Path, overrides: List[str], hf: Dict[str,
     text, spec_path, base = _spec(spec)
     files: Dict[str, dict] = {}
     sent: Dict[str, str] = {}
-    for key, ref in hf.items():
-        files[key] = {"hf": ref}
+    for key, ref in remote.items():
+        files[key] = dict(ref)
         local = _local(base, key)
-        if local.is_file():  # pin the bytes this machine has: the Hub's file must be the same
+        if local.is_file():  # pin the bytes this machine has: the Hub's file or the object must be the same
             files[key]["sha256"] = _file_sha256(local)
-        sent[key] = "hf"
+        sent[key] = next(iter(ref))
     body: dict = {"spec_yaml": text, "spec_path": spec_path, "files": files}
     if overrides:
         body["overrides"] = list(overrides)
@@ -356,8 +363,8 @@ def _fits(key: str, raw_bytes: int, files: Dict[str, dict], biggest: List[Tuple[
     if size > _INLINE_MAX or total > _BODY_MAX:
         hint = "".join(f", {posixpath.join(key, rel)} {p.stat().st_size >> 20} MB" for rel, p in biggest)
         raise KiteAPIError(413, f"{key} is not a file Kite ships, and it is too big to send inline "
-                                f"({raw_bytes >> 20} MB{hint}): put its large files on the Hugging Face Hub and "
-                                "pass each with --hf PATH=OWNER/NAME@COMMIT:FILE", code="file_too_large",
+                                f"({raw_bytes >> 20} MB{hint}): put its large files on the Hugging Face Hub or in a public "
+                                "S3 bucket and pass each with --hf PATH=OWNER/NAME@COMMIT:FILE or --s3 PATH=s3://BUCKET/KEY", code="file_too_large",
                            details={"path": key})
 
 
@@ -401,11 +408,28 @@ def hf_refs(pairs: List[str]) -> Dict[str, dict]:
     refs: Dict[str, dict] = {}
     for pair in pairs:
         key, _, ref = pair.partition("=")
-        m = _HF.fullmatch(ref.strip())
+        ref = ref.strip()
+        if key and re.match(r"(https?://)?(www\.)?(huggingface|hf)\.co/", ref):  # the API reads the URL
+            refs[key.strip()] = {"hf": ref}
+            continue
+        m = _HF.fullmatch(ref)
         if not key or not m:
             raise KiteAPIError(400, f"--hf {pair!r}: expected PATH=OWNER/NAME@COMMIT:FILE, with the full 40-character "
-                                    "commit", code="parameter_invalid", details={"param": "hf"})
-        refs[key.strip()] = {"repo": m["repo"], "revision": m["revision"], "path": m["path"]}
+                                    "commit, or PATH=the file's huggingface.co URL", code="parameter_invalid",
+                               details={"param": "hf"})
+        refs[key.strip()] = {"hf": {"repo": m["repo"], "revision": m["revision"], "path": m["path"]}}
+    return refs
+
+
+def s3_refs(pairs: List[str]) -> Dict[str, dict]:
+    """PATH=S3_LINK pairs as {"s3": link} references; the API reads the link."""
+    refs: Dict[str, dict] = {}
+    for pair in pairs:
+        key, _, link = pair.partition("=")
+        if not key or not re.match(r"(?i)(s3://|https://)\S+", link.strip()):
+            raise KiteAPIError(400, f"--s3 {pair!r}: expected PATH=s3://BUCKET/KEY", code="parameter_invalid",
+                               details={"param": "s3"})
+        refs[key.strip()] = {"s3": link.strip()}
     return refs
 
 

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/src/kite_cli/client.py

@@ -161,6 +161,17 @@ class KiteClient:
                 for chunk in resp.iter_bytes(chunk_size=8192):
                     f.write(chunk)
 
+    def redirect_target(self, path: str) -> Optional[str]:
+        """Where an API download path (a run's or a twin's archive) redirects: a fresh signed link, or None when the
+        API streams the file itself. Only the response's headers are read."""
+        with self._client.stream("GET", path, headers=self._base_headers(), follow_redirects=False) as resp:
+            if resp.is_redirect:
+                return resp.headers.get("location")
+            if resp.status_code >= 400:
+                resp.read()
+                raise _api_error(resp)
+            return None
+
     def close(self) -> None:
         self._client.close()
 

{kiteml_cli-0.4.0 → kiteml_cli-0.5.0}/src/kite_cli/mcp/server.py

@@ -56,6 +56,23 @@ def _twin_source(dataset: str) -> dict:
     return {"type": "huggingface", "repo_id": dataset}
 
 
+def _fetch_command(d: dict, fname: str, client: KiteClient) -> str:
+    """The command that fetches a finished resource's archive. From Kite-Version 2026-10-12 `download.url` is this
+    API's path, which needs the key: ask it for the fresh signed link it redirects to, so an agent without the key
+    (the hosted server's OAuth) can fetch it directly. Without one, the API path with the key."""
+    if d.get("requires_auth"):
+        resolve = getattr(client, "redirect_target", None)
+        try:
+            signed = resolve(d["url"]) if resolve else None
+        except Exception:  # the link is a convenience: the API path below still works
+            signed = None
+        if signed:
+            return f'curl -fL "{signed}" -o {fname}'
+        url = f"{client.base_url.rstrip('/')}{d['url']}"
+        return f'curl -fL -H "Authorization: Bearer $KITE_API_KEY" "{url}" -o {fname}'
+    return f'curl -fL "{d["url"]}" -o {fname}'
+
+
 def _twin_view(twin: dict, client: KiteClient) -> dict:
     """A twin as an agent should see it: the resource plus `next`, the one thing to do now. A hosted MCP server
     cannot write to the agent's disk, so a finished twin's `next` is the exact command that fetches and
@@ -65,11 +82,7 @@ def _twin_view(twin: dict, client: KiteClient) -> dict:
     d = twin.get("download") or {}
     if status == "succeeded" and d:
         fname = d.get("filename") or f"{tid}.tar.gz"
-        if d.get("requires_auth"):
-            url = f"{client.base_url.rstrip('/')}{d['url']}"
-            fetch = f'curl -fL -H "Authorization: Bearer $KITE_API_KEY" "{url}" -o {fname}'
-        else:
-            fetch = f'curl -fL "{d["url"]}" -o {fname}'
+        fetch = _fetch_command(d, fname, client)
         out["next"] = (f"{fetch} && tar xzf {fname}   # -> {d.get('scene')}"
                        + (f" (link valid until {d['expires_at']}; call kite_twin_status again for a fresh one)"
                           if d.get("expires_at") else ""))
@@ -95,11 +108,7 @@ def _eval_view(run: dict, client: KiteClient) -> dict:
     d = run.get("download") or {}
     if status == "succeeded" and d:
         fname = d.get("filename") or f"{rid}.tar.gz"
-        if d.get("requires_auth"):
-            url = f"{client.base_url.rstrip('/')}{d['url']}"
-            fetch = f'curl -fL -H "Authorization: Bearer $KITE_API_KEY" "{url}" -o {fname}'
-        else:
-            fetch = f'curl -fL "{d["url"]}" -o {fname}'
+        fetch = _fetch_command(d, fname, client)
         out["next"] = (f"{fetch} && tar xzf {fname}   # -> {rid}/, the run directory (`kite verify {rid} --evals <the spec's evaluators>` checks it)"
                        + (f" (link valid until {d['expires_at']}; call kite_eval_status again for a fresh one)"
                           if d.get("expires_at") else "") + f"; kite_eval_report {rid} has the scorecard")
@@ -131,10 +140,11 @@ def _eval_request(args: dict, client: KiteClient) -> tuple[dict, dict | None]:
         raise ValueError("`spec` only works with the local (stdio) Kite MCP server; send `spec_yaml` and `files`")
     from pathlib import Path
 
-    from kite_cli.cli.commands.evaluations import hf_refs, request
+    from kite_cli.cli.commands.evaluations import hf_refs, request, s3_refs
 
-    hf = hf_refs([f"{k}={v}" for k, v in (args.get("hf") or {}).items()])
-    body, validation, _ = request(client, Path(args["spec"]), args.get("overrides") or [], hf, args.get("kite_version"))
+    remote = {**hf_refs([f"{k}={v}" for k, v in (args.get("hf") or {}).items()]),
+              **s3_refs([f"{k}={v}" for k, v in (args.get("s3") or {}).items()])}
+    body, validation, _ = request(client, Path(args["spec"]), args.get("overrides") or [], remote, args.get("kite_version"))
     return body, validation
 
 
@@ -194,9 +204,10 @@ _EVAL_SPEC = {
     "spec": {"type": "string", "description": "A Kite SDK spec file on this machine, best as an absolute path (the local server only): every path it names is sent from this disk, a file Kite ships by its hash alone"},
     "spec_yaml": {"type": "string", "description": "The spec's text, instead of `spec` (the hosted server): send `spec_path` and `files` with it"},
     "spec_path": {"type": "string", "description": "With spec_yaml: its path in your workspace, e.g. specs/smoke.yaml. Paths in the spec are read relative to it"},
-    "files": {"type": "object", "description": "With spec_yaml: each path the spec names, as it names it, to where its bytes come from: {\"kind\": \"directory\", \"shipped\": \"evals\"} (a path Kite ships), {\"sha256\", \"kind\"} (a shipped file by hash), {\"text\"} or {\"base64\"} (a folder as a .tar), or {\"hf\": {\"repo\", \"revision\" (a full commit), \"path\"}}. A missing one is files_required, listing each"},
+    "files": {"type": "object", "description": "With spec_yaml: each path the spec names, as it names it, to where its bytes come from: {\"kind\": \"directory\", \"shipped\": \"evals\"} (a path Kite ships), {\"sha256\", \"kind\"} (a shipped file by hash), {\"text\"} or {\"base64\"} (a folder as a .tar), or {\"hf\": {\"repo\", \"revision\" (a full commit), \"path\"}} or {\"hf\": \"<the file's huggingface.co URL>\"} (a branch is pinned to its commit), or {\"s3\": \"s3://BUCKET/KEY\"} (an object in a public bucket, pinned to its ETag). A missing one is files_required, listing each"},
     "overrides": {"type": "array", "items": {"type": "string"}, "description": "FIELD=VALUE pairs, as `kite run --set`, e.g. episodes.count=20. Optional"},
-    "hf": {"type": "object", "description": "With spec: {PATH: \"OWNER/NAME@COMMIT:FILE\"}, the paths that are a public Hugging Face file (a checkpoint's weights). Optional"},
+    "hf": {"type": "object", "description": "With spec: {PATH: \"OWNER/NAME@COMMIT:FILE\" or the file's huggingface.co URL}, the paths that are a public Hugging Face file (a checkpoint's weights). Optional"},
+    "s3": {"type": "object", "description": "With spec: {PATH: \"s3://BUCKET/KEY\"}, the paths that are an object in a public S3 bucket. Optional"},
     "kite_version": {"type": "string", "description": "The Kite release you expect; refused if it is not the one served. Optional"},
 }
 _EVAL_ID = {"run_id": {"type": "string", "description": "Evaluation id (eval_...)"}}
@@ -208,9 +219,9 @@ _EVAL_ID = {"run_id": {"type": "string", "description": "Evaluation id (eval_...
 TOOLS = [
     # Augmentations (public /v1 API — relighting + video augmentation)
     Tool(name="kite_augment_create", title="Create Augmentation",
-         description="Create a dataset augmentation run from a HuggingFace LeRobot dataset. model 'auto' (default) lets a planner pick per camera: 'relight' (cheap, exact colour/exposure/warmth transform of every episode — e.g. match a darker, warmer deployment scene), 'background' (a new place behind the work cell from a panorama; the robot, objects and table keep their recorded pixels), 'world' (a 3D world rendered from each camera's pose, for moving cameras) or 'video' (generative video augmentation of content: textures, objects). background and world run on whole datasets and may not be enabled for every account. Pass reference_images of the target look. Returns the augmentation resource with its id (aug_...) and, once planned, the chosen plan.",
+         description="Create a dataset augmentation run from a LeRobot dataset on HuggingFace or in a public S3 bucket. model 'auto' (default) lets a planner pick per camera: 'relight' (cheap, exact colour/exposure/warmth transform of every episode — e.g. match a darker, warmer deployment scene), 'background' (a new place behind the work cell from a panorama; the robot, objects and table keep their recorded pixels), 'world' (a 3D world rendered from each camera's pose, for moving cameras) or 'video' (generative video augmentation of content: textures, objects). background and world run on whole datasets and may not be enabled for every account. Pass reference_images of the target look. Returns the augmentation resource with its id (aug_...) and, once planned, the chosen plan.",
          inputSchema={"type": "object", "properties": {
-             "repo_id": {"type": "string", "description": "Source HF dataset id, e.g. lerobot/pusht"},
+             "repo_id": {"type": "string", "description": "Source LeRobot dataset: a HuggingFace id (lerobot/pusht) or its huggingface.co/datasets/ URL, or an S3 link to its folder in a public bucket (s3://bucket/dataset/ or its https://….amazonaws.com URL)"},
              "instructions": {"type": "string", "description": "Augmentation instruction"},
              "episode_count": {"type": "integer", "description": "Video: episodes to generate (1-50, default 3)"},
              "episodes": {"anyOf": [{"type": "integer", "minimum": 1}, {"type": "string", "enum": ["all"]}],
@@ -228,6 +239,7 @@ TOOLS = [
              "variants": {"type": "integer", "description": "background/world: environments per source episode (default 1)"},
              "preview_first": {"type": "boolean", "description": "background/world: check one episode's quality gates (windows at its start, middle and end) before charging for the rest (default true)"},
              "allow_video_fallback": {"type": "boolean", "description": "auto: where no scene engine can run, use generative video (far more tokens, no pixel guarantee) instead of failing with the reasons and the video price (default false)"},
+             "video_check": {"type": "string", "enum": ["standard", "strict", "off"], "description": "Generative video: what the check of each clip rejects. standard (default): a changed robot or a missing, duplicated or swapped object; an object redrawn or the table or view redrawn ship as warnings. strict: every problem rejects. off: records warnings, never rejects"},
              "reference_images": {"type": "array", "maxItems": 6, "description": "Photos of the target look", "items": {
                  "type": "object", "properties": {
                      "camera": {"type": "string", "description": "Camera it shows (e.g. 'top'); omit for all cameras"},
@@ -419,7 +431,8 @@ async def _dispatch(name: str, args: dict, client: KiteClient) -> list[TextConte
             "source": {"type": "huggingface", "repo_id": args["repo_id"]},
             "instructions": args["instructions"],
             "config": scene_config(args.get("episodes", args.get("episode_count")), args.get("model"),
-                                   args.get("variants"), args.get("preview_first"), args.get("allow_video_fallback")),
+                                   args.get("variants"), args.get("preview_first"), args.get("allow_video_fallback"),
+                                   args.get("video_check")),
             "output": {"type": args.get("output_type", "download")},
             **{k: args[k] for k in ("environments", "must_keep") if args.get(k)},
         }