sequence-base 0.3.18__tar.gz → 0.3.21__tar.gz

{sequence_base-0.3.18 → sequence_base-0.3.21}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: sequence-base
-Version: 0.3.18
+Version: 0.3.21
 Summary: Shared foundation for the Sequences platform — wire contract, correlation, error taxonomy, constants. Used by the robot SDK, gateway, and container harness.
 Requires-Python: >=3.10
 Requires-Dist: pydantic>=2.0

{sequence_base-0.3.18 → sequence_base-0.3.21}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sequence-base"
-version = "0.3.18"
+version = "0.3.21"
 description = "Shared foundation for the Sequences platform — wire contract, correlation, error taxonomy, constants. Used by the robot SDK, gateway, and container harness."
 # 3.10, not 3.11: `seq` (which imports this) must run on ROS 2 Humble's Python 3.10, and nothing here needs 3.11.
 # {2026-10-03 the full suite on CPython 3.10.20: "147 PASSED"; vermin --eval-annotations: "MINIMUM REQUIRED VERSIONS: 3.9"}

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/constants/version.py

@@ -4,7 +4,10 @@
 # across deployments / repos can tell which version of a shape it is validating against. MUST change
 # whenever any shared wire shape changes (a field added / removed / retyped) — that is what makes
 # contract evolution mechanically detectable instead of eyeballed. Bump on any shape change.
-CONTRACT_VERSION = "0.4.15"  # 0.4.14 -> 0.4.15: limits count GPUs and CPU/memory are billed like Modal — Resources.cpu / memory_gb optional floats + cpu_limit / memory_limit_gb; Autoscaler.max_gpus / on_full / queue_timeout_s replace max_containers; Concurrency fields optional + replicas_per_gpu / gpus_per_replica; DeploymentStatusResponse.reason; EvalRequest.policy_instances / policy_gpu / benchmark_instances / benchmark_gpu / envs replace gpus / envs_per_gpu; EvalStatus.QUEUED; EvalCards / EvalQueue
+CONTRACT_VERSION = "0.4.18"  # 0.4.17 -> 0.4.18: DeploymentPolicy.models — the platform models a policy calls while it runs (catalogue ids), which the act key driving it may call on /v1/chat and /v1/decide (gs-plans models-api/11)
+# previous: 0.4.17  # 0.4.16 -> 0.4.17: weights optional — PolicySpec.weights, DeploymentRequest.weights and BootstrapResponse.weights_uri may be None, for a policy without a checkpoint (nothing copied or mounted; gs-plans storage T13)
+# previous: 0.4.16  # 0.4.15 -> 0.4.16: batched benchmarks — SeqFlags.OBSERVE / STEP, BenchmarkSpec / BenchmarkRegisterRequest.envs and .batched; per-arm eval reports — EvalReport.arms / paired / timing / stop_reason, EvalStatus.CANCELED, Estimate.breakdown / default_timing, EvalResponse.arms, EvalDryRunResponse.layout / notes
+# previous: 0.4.15  # 0.4.14 -> 0.4.15: limits count GPUs and CPU/memory are billed like Modal — Resources.cpu / memory_gb optional floats + cpu_limit / memory_limit_gb; Autoscaler.max_gpus / on_full / queue_timeout_s replace max_containers; Concurrency fields optional + replicas_per_gpu / gpus_per_replica; DeploymentStatusResponse.reason; EvalRequest.policy_instances / policy_gpu / benchmark_instances / benchmark_gpu / envs replace gpus / envs_per_gpu; EvalStatus.QUEUED; EvalCards / EvalQueue
 # previous: 0.4.14  # 0.4.13 -> 0.4.14: the /act wire is opaque (ActRequest.inputs / ActResponse.outputs); action / observation contracts optional + DeploymentPolicy/PolicySpec/ConnectResponse .io; the robot-arm types live in sequence_base.presets.robot_arm (contract.act / contract.io re-export them); ImageStepAddLocal (seq.Image.add_local_*); ActionSpace.JOINT_VELOCITY (pi0.5 DROID's joint velocities + gripper position)
 # previous: 0.4.13  # 0.4.12 -> 0.4.13: template://<name>[@v<n>] — official template weights (parse_template); gs:// / s3:// / https:// still accepted until their removal
 # previous: 0.4.12  # 0.4.11 -> 0.4.12: one error definition — ErrorBody is every non-2xx body (detail, code, retryable, retry_after_s, request_id, hint, details), ErrorCode covers every case, ERRORS / ErrorCodes.json give each code its status, retryability, SDK class and CLI exit code

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/contract/accounts.py

@@ -76,7 +76,8 @@ class BootstrapResponse(BaseModel):
     tls_key: str = Field(..., description="Self-signed TLS private key (down-direction platform secret)")
     tls_cert: str = Field(..., description="Self-signed TLS certificate (public)")
     env: dict[str, str] = Field(..., description="Policy's user secrets, expanded ENV_KEY → value")
-    weights_uri: str
+    # None: a policy without a checkpoint — nothing was copied or mounted. {gs-plans/storage T13}
+    weights_uri: str | None = None
     lease: Lease
 
 

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/contract/benchmarks.py

@@ -16,7 +16,7 @@ from pydantic import BaseModel, ConfigDict, Field, model_validator
 
 from sequence_base.contract.accounts import SecretRef
 from sequence_base.contract.deployments import CodeBundle, DeploymentState
-from sequence_base.contract.flags import BENCHMARK_BINDABLE, SeqFlags
+from sequence_base.contract.flags import BENCHMARK_BATCHED, BENCHMARK_BINDABLE, SeqFlags
 from sequence_base.contract.policy import ActionContract, ImageSpec, VolumeRef
 
 __all__ = [
@@ -92,8 +92,10 @@ class BenchmarkSpec(BaseModel):
     conditions: Conditions
     max_steps: int = Field(..., description="Platform safety cap; truncates and records truncated")
     handler_ref: str
-    bindings: dict[SeqFlags, str] = Field(..., description="SETUP / RESET / CHECK / SCORE → method name")
+    bindings: dict[SeqFlags, str] = Field(..., description="SETUP / RESET / CHECK / SCORE (+ OBSERVE / STEP) → method name")
     gpu: str | list[str] | None = None
+    # How many environments one benchmark instance runs at once when an eval does not say (the author's default).
+    envs: int | None = Field(default=None, ge=1, le=65536)
     assets: list[VolumeRef] = Field(default_factory=list)
     render: RenderSpec | None = None
     cpu: int | None = None
@@ -108,9 +110,12 @@ class BenchmarkSpec(BaseModel):
         if bad:
             names = ", ".join(sorted(f.name or str(f.value) for f in bad))
             raise ValueError(
-                f"benchmark bindings may only cover SETUP / RESET / CHECK / SCORE; got disallowed {names}"
+                f"benchmark bindings may only cover SETUP / RESET / CHECK / SCORE / OBSERVE / STEP; got disallowed {names}"
                 " (ROLLOUT is deprecated and must not be bound)"
             )
+        batched = set(self.bindings) & set(BENCHMARK_BATCHED)
+        if batched and batched != set(BENCHMARK_BATCHED):
+            raise ValueError("a batched benchmark binds both OBSERVE and STEP")
         return self
 
 
@@ -166,8 +171,13 @@ class BenchmarkRegisterRequest(BaseModel):
     timeout_s: float | None = Field(default=None, gt=0)
     # The adapter (module:Class in the benchmark's own code) an eval uses when it names none.
     default_adapter_ref: str | None = None
-    # A GPU for each sim sandbox (a renderer or a GPU physics engine): a card name the platform offers. None: CPU.
+    # The card each benchmark instance runs on (L4 or RTX-PRO-6000); None takes the platform's default (L4). The
+    # simulation never runs on a CPU. {gs-plans/benchmark-internals/design.md "一、规则"}
     gpu: str | None = Field(default=None, max_length=32)
+    # How many environments one instance runs at once when an eval does not say; None takes 1.
+    envs: int | None = Field(default=None, ge=1, le=65536)
+    # Whether the benchmark is batched (binds OBSERVE and STEP and steps many environments in one call).
+    batched: bool = False
     # The Dockerfile's build context, as for a deployment.
     image_context: CodeBundle | None = None
 

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/contract/deployments.py

@@ -157,12 +157,23 @@ class DeploymentPolicy(BaseModel):
     # The inference parameters @seq.infer takes beyond the observation (name -> type, default). None: an older
     # SDK that did not look; {}: the policy declares none.
     params: dict[str, InferParam] | None = Field(default=None, max_length=32)
+    # The platform models the policy calls while it runs (`@seq.policy(models=[...])`), by full catalogue id
+    # (accounts/sequences/models/<name>): the act key that drives this deployment may call exactly these on /v1/chat
+    # and /v1/decide, and nothing else it could spend on. None: the policy calls none (and an older SDK never sent it).
+    # {gs-plans models-api/11, user 2026-10-04 "act key should have ability to call any model involved in the policy"}
+    models: list[str] | None = Field(default=None, max_length=16)
 
     @model_validator(mode="after")
     def _param_names(self) -> "DeploymentPolicy":
         for name in self.params or {}:
             if not name.isidentifier() or name.startswith("_"):
                 raise ValueError(f"inference parameter {name!r} is not a public Python identifier")
+        # A declared model is a catalogue id: no blanks, no duplicates (the gateway checks each one exists).
+        for model in self.models or []:
+            if not model or model != model.strip() or "/" not in model:
+                raise ValueError(f"policy model {model!r} is not a catalogue id (accounts/sequences/models/<name>)")
+        if self.models and len(set(self.models)) != len(self.models):
+            raise ValueError("policy models lists a model twice")
         return self
 
 
@@ -206,14 +217,15 @@ class CodeBundle(BaseModel):
 
 
 class DeploymentRequest(BaseModel):
-    """The full `POST /v1/deployments` body. Required: name, image, weights, resources, policy (goal 18).
-    autoscaler / concurrency / timeouts are typed optional knobs with sane defaults."""
+    """The full `POST /v1/deployments` body. Required: name, image, resources, policy (goal 18); weights unless the
+    policy has no checkpoint. autoscaler / concurrency / timeouts are typed optional knobs with sane defaults."""
 
     model_config = ConfigDict(extra="forbid")
 
     name: str
     image: ImageSpec
-    weights: VolumeMount
+    # None: a policy without a checkpoint — no weights volume is made, copied or mounted. {gs-plans/storage T13}
+    weights: VolumeMount | None = None
     resources: Resources
     policy: DeploymentPolicy
     # The handler's source. None only when `handler_ref` is importable from the image itself (a package the
@@ -252,6 +264,10 @@ class DeploymentRequest(BaseModel):
         and, until they are removed, a gs:// or s3:// prefix or an https:// file — never empty. PolicySpec already enforces this for SDK-authored deploys; this holds it for any POSTed body."""
         from sequence_base.contract.policy import check_weights_parts
 
+        if self.weights is None:
+            if self.weights_name or self.extra_weights:
+                raise ValueError("weights parts need weights: name the first part's origin too")
+            return self
         check_weights_parts(self.weights_name, dict(self.extra_weights or {}))
         for what, ref in [("weights.uri", self.weights)] + [(f"extra_weights[{n!r}].uri", r)
                                                             for n, r in (self.extra_weights or {}).items()]:

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/contract/eval.py

@@ -17,6 +17,9 @@ from pydantic import BaseModel, ConfigDict, Field, model_validator
 __all__ = [
     "EvalRequest",
     "ClampInfo",
+    "ArmReport",
+    "PairedDiff",
+    "EvalTiming",
     "Estimate",
     "EvalResponse",
     "EvalDryRunResponse",
@@ -57,7 +60,10 @@ class EvalRequest(BaseModel):
     envs: int | None = Field(default=None, ge=1, le=100000)
     vlm_max_concurrency: int | None = None
     adapter: str | None = None
-    compare: str | None = Field(default=None, description="Another model id to compare against")
+    # Arms run side by side in one eval: other models on the same episodes (each its own policy instances), and/or a
+    # grid of inference-parameter values (every combination is one arm).
+    compare: str | list[str] | None = Field(default=None, description="Other model ids to compare against")
+    sweep: dict[str, list[int | float | bool | str]] | None = Field(default=None, max_length=8)
     dry_run: bool = False
     # Only these tasks (of the selected suite / trials). {gs-plans/customization/benchmark.md "子集"}
     tasks: list[str] | None = Field(default=None, max_length=1000)
@@ -90,6 +96,10 @@ class Estimate(BaseModel):
     wall_clock_s: float
     cost_usd: float
     vlm_calls: int
+    # The cost split by machine and by resource (GPU / CPU / memory), when the gateway itemizes it.
+    breakdown: dict[str, float] | None = None
+    # True when no run of this benchmark version has been measured yet and the estimate uses the default step time.
+    default_timing: bool | None = None
 
 
 class EvalResponse(BaseModel):
@@ -103,6 +113,8 @@ class EvalResponse(BaseModel):
     clamped_by: ClampInfo | None = None
     # Set when the account was short of GPUs: the eval waits (up to its timeout) instead of being refused.
     queue: EvalQueue | None = None
+    # How many arms (model x swept parameters) the eval runs side by side.
+    arms: int | None = None
 
 
 class CardSide(BaseModel):
@@ -160,6 +172,10 @@ class EvalDryRunResponse(BaseModel):
 
     estimate: Estimate
     cards: EvalCards | None = None
+    # The eval's layout as it would run: per side the instances, card and environments (the user's numbers or the
+    # fixed defaults), and the arms. Advice only ever comes as `notes`; the platform never changes the numbers.
+    layout: dict | None = None
+    notes: list[str] = Field(default_factory=list)
 
 
 class RolloutResult(BaseModel):
@@ -190,6 +206,7 @@ class EvalStatus(str, Enum):
     RUNNING = "running"
     COMPLETED = "completed"
     FAILED = "failed"
+    CANCELED = "canceled"
 
 
 class SuiteAgg(BaseModel):
@@ -236,6 +253,49 @@ class LossyTransform(BaseModel):
     reason: str
 
 
+class ArmReport(BaseModel):
+    """One arm of an eval (a model x one combination of swept inference parameters): its own metrics and episode
+    counts, so arms are compared side by side and never pooled."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    arm: int
+    model: str
+    params: dict[str, int | float | bool | str] = Field(default_factory=dict)
+    metrics: dict[str, MetricAgg] = Field(default_factory=dict)
+    # Each metric's 95% confidence interval as [low, high].
+    intervals: dict[str, list[float]] = Field(default_factory=dict)
+    episodes: int = 0
+    failed: int = 0
+    not_run: int = 0
+    budget_stopped: bool = False
+
+
+class PairedDiff(BaseModel):
+    """`compare`: one metric's difference between two arms over the seeds both ran (b - a), with its interval."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    metric: str
+    arm_a: int
+    arm_b: int
+    n: int
+    mean_diff: float
+    interval: list[float]
+
+
+class EvalTiming(BaseModel):
+    """Where an eval's time went: cold starts, and per step the simulation, rendering, the policy's batched forward and
+    the network around it, plus the forward's average batch size."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    cold_start_s: dict[str, float] = Field(default_factory=dict)
+    step_ms: dict[str, float] = Field(default_factory=dict)
+    batch_size_mean: float | None = None
+    steps: int = 0
+
+
 class EvalReport(BaseModel):
     """GET /v1/eval/{id} (eval-raw revision). NO hard-coded `overall_sr` / `success_rate` top-level field —
     `metrics` is a passthrough dict keyed by the user's metric names (from `@seq.check`'s `Step.metrics`),
@@ -257,6 +317,12 @@ class EvalReport(BaseModel):
     errors: list[str] = Field(
         default_factory=list, description="Why episodes or the run failed, so an incomplete report says why"
     )
+    # Per arm (one entry for an eval with one arm), paired comparisons for `compare`, the timing breakdown, and why
+    # the run stopped early (canceled / budget), when it did.
+    arms: list[ArmReport] | None = None
+    paired: list[PairedDiff] | None = None
+    timing: EvalTiming | None = None
+    stop_reason: str | None = None
 
 
 class AdapterKind(str, Enum):

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/contract/flags.py

@@ -41,18 +41,25 @@ class SeqFlags(IntFlag):
     SCORE = 16384       # cross-episode aggregate (optional; default = mean of each numeric metric)
     RESET = 32768       # build one episode's env (gym-style handle) — eval-raw new hook, bit 15
     CHECK = 65536       # per step: read raw env state → done + user-named metrics — eval-raw new hook, bit 16
+    # A batched benchmark (a GPU sim that steps many environments at once) also binds these two; its RESET then
+    # takes (slots, conditions) and its CHECK (slots). {gs-plans/benchmark-internals/design.md "三、两种写法"}
+    OBSERVE = 131072    # read the observation of the given slots, one per slot — bit 17
+    STEP = 262144       # apply one action to each given slot — bit 18
     # DEPRECATED — the old single-hook benchmark model. Bit 8192 stays reserved (never reused for another
     # member) but is NOT bound in any registration; setup/reset/check/score replaces it. Kept only so the
     # value cannot be recycled and an old serialized 8192 still decodes to a known name.
     ROLLOUT = 8192
 
 
-# The benchmark lifecycle bits a live registration may bind (eval-raw): SETUP / RESET / CHECK / SCORE.
-# ROLLOUT is deliberately absent — deprecated, never bound.
+# The benchmark lifecycle bits a live registration may bind (eval-raw): SETUP / RESET / CHECK / SCORE, and OBSERVE /
+# STEP for a batched benchmark. ROLLOUT is deliberately absent — deprecated, never bound.
 BENCHMARK_BINDABLE: frozenset[SeqFlags] = frozenset(
-    {SeqFlags.SETUP, SeqFlags.RESET, SeqFlags.CHECK, SeqFlags.SCORE}
+    {SeqFlags.SETUP, SeqFlags.RESET, SeqFlags.CHECK, SeqFlags.SCORE, SeqFlags.OBSERVE, SeqFlags.STEP}
 )
 
+# The two bits that make a benchmark batched: present together or not at all.
+BENCHMARK_BATCHED: frozenset[SeqFlags] = frozenset({SeqFlags.OBSERVE, SeqFlags.STEP})
+
 # Deprecated benchmark bits: reserved value, must not appear in any bindings.
 DEPRECATED_FLAGS: frozenset[SeqFlags] = frozenset({SeqFlags.ROLLOUT})
 

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/contract/policy.py

@@ -185,8 +185,9 @@ class PolicySpec(BaseModel):
     name: str
     gpu: str | list[str] = Field(..., description="A GPU type, or a list = cross-provider fallback")
     image: ImageSpec
-    weights: str = Field(..., description="Weights URI: template:// (an official template), hf:// or vol:// (an "
-                                          "uploaded volume)")
+    # None: the policy has no checkpoint (a closed-form controller); nothing is copied or mounted for it.
+    weights: str | None = Field(None, description="Weights URI: template:// (an official template), hf:// or vol:// "
+                                                  "(an uploaded volume); none for a policy without a checkpoint")
     volumes: dict[str, VolumeRef] = Field(default_factory=dict)
     secrets: list[SecretRef] = Field(default_factory=list)
     min_containers: int = 0
@@ -217,11 +218,13 @@ class PolicySpec(BaseModel):
 
     @field_validator("weights")
     @classmethod
-    def _weights_scheme(cls, v: str) -> str:
-        return check_weights_uri(v, "weights")
+    def _weights_scheme(cls, v: str | None) -> str | None:
+        return None if v is None else check_weights_uri(v, "weights")
 
     @model_validator(mode="after")
     def _weights_parts(self) -> "PolicySpec":
+        if self.weights is None and (self.weights_name or self.extra_weights):
+            raise ValueError("weights parts need weights: name the first part's origin too")
         check_weights_parts(self.weights_name, {k: v for k, v in (self.extra_weights or {}).items()})
         for name, uri in (self.extra_weights or {}).items():
             check_weights_uri(uri, f"weights[{name!r}]")

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ActRequest.json

@@ -65,5 +65,5 @@
   ],
   "title": "ActRequest",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ActResponse.json

@@ -75,5 +75,5 @@
   ],
   "title": "ActResponse",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ActionChunk.json

@@ -99,5 +99,5 @@
   ],
   "title": "ActionChunk",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ActionContract.json

@@ -100,5 +100,5 @@
   ],
   "title": "ActionContract",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ActionStep.json

@@ -19,5 +19,5 @@
   ],
   "title": "ActionStep",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/AdapterTransform.json

@@ -38,5 +38,5 @@
   ],
   "title": "AdapterTransform",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/AddLocal.json

@@ -23,5 +23,5 @@
   ],
   "title": "AddLocal",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/Autoscaler.json

@@ -69,5 +69,5 @@
   },
   "title": "Autoscaler",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/BenchmarkConditionSpec.json

@@ -62,5 +62,5 @@
   ],
   "title": "BenchmarkConditionSpec",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/BenchmarkRegisterRequest.json

@@ -523,6 +523,11 @@
       "title": "Assets",
       "type": "array"
     },
+    "batched": {
+      "default": false,
+      "title": "Batched",
+      "type": "boolean"
+    },
     "code": {
       "anyOf": [
         {
@@ -577,6 +582,20 @@
       "default": null,
       "title": "Default Adapter Ref"
     },
+    "envs": {
+      "anyOf": [
+        {
+          "maximum": 65536,
+          "minimum": 1,
+          "type": "integer"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Envs"
+    },
     "gpu": {
       "anyOf": [
         {
@@ -681,5 +700,5 @@
   ],
   "title": "BenchmarkRegisterRequest",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/BenchmarkSpec.json

@@ -273,15 +273,18 @@
         8192,
         16384,
         32768,
-        65536
+        65536,
+        131072,
+        262144
       ],
       "title": "SeqFlags",
       "type": "integer",
-      "x-contract-version": "0.4.15",
+      "x-contract-version": "0.4.18",
       "x-flag-names": {
         "CHECK": 65536,
         "INFER": 16,
         "LOAD": 2,
+        "OBSERVE": 131072,
         "ON_EPISODE": 1024,
         "PLAN": 4,
         "RESET": 32768,
@@ -289,6 +292,7 @@
         "SCORE": 16384,
         "SETUP": 4096,
         "SHUTDOWN": 8,
+        "STEP": 262144,
         "WARMUP": 1
       }
     },
@@ -357,7 +361,7 @@
       "additionalProperties": {
         "type": "string"
       },
-      "description": "SETUP / RESET / CHECK / SCORE → method name",
+      "description": "SETUP / RESET / CHECK / SCORE (+ OBSERVE / STEP) → method name",
       "propertyNames": {
         "enum": [
           "1",
@@ -370,7 +374,9 @@
           "8192",
           "16384",
           "32768",
-          "65536"
+          "65536",
+          "131072",
+          "262144"
         ],
         "type": "string"
       },
@@ -392,6 +398,20 @@
       "default": null,
       "title": "Cpu"
     },
+    "envs": {
+      "anyOf": [
+        {
+          "maximum": 65536,
+          "minimum": 1,
+          "type": "integer"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Envs"
+    },
     "gpu": {
       "anyOf": [
         {
@@ -467,5 +487,5 @@
   ],
   "title": "BenchmarkSpec",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/BenchmarkStatusResponse.json

@@ -40,5 +40,5 @@
   ],
   "title": "BenchmarkStatusResponse",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/BootstrapResponse.json

@@ -58,8 +58,16 @@
       "type": "string"
     },
     "weights_uri": {
-      "title": "Weights Uri",
-      "type": "string"
+      "anyOf": [
+        {
+          "type": "string"
+        },
+        {
+          "type": "null"
+        }
+      ],
+      "default": null,
+      "title": "Weights Uri"
     }
   },
   "required": [
@@ -67,10 +75,9 @@
     "tls_key",
     "tls_cert",
     "env",
-    "weights_uri",
     "lease"
   ],
   "title": "BootstrapResponse",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/Camera.json

@@ -34,5 +34,5 @@
   ],
   "title": "Camera",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ClampInfo.json

@@ -23,5 +23,5 @@
   ],
   "title": "ClampInfo",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/CodeBundle.json

@@ -23,5 +23,5 @@
   ],
   "title": "CodeBundle",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/Concurrency.json

@@ -61,5 +61,5 @@
   },
   "title": "Concurrency",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/Conditions.json

@@ -55,5 +55,5 @@
   ],
   "title": "Conditions",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ConnectRequest.json

@@ -25,5 +25,5 @@
   ],
   "title": "ConnectRequest",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/ConnectResponse.json

@@ -307,5 +307,5 @@
   ],
   "title": "ConnectResponse",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }

{sequence_base-0.3.18 → sequence_base-0.3.21}/sequence_base/schemas/DeploymentAccepted.json

@@ -18,5 +18,5 @@
   ],
   "title": "DeploymentAccepted",
   "type": "object",
-  "x-contract-version": "0.4.15"
+  "x-contract-version": "0.4.18"
 }