urun-cli 0.6.0__tar.gz → 0.6.2__tar.gz

{urun_cli-0.6.0 → urun_cli-0.6.2}/.gitignore

@@ -21,3 +21,6 @@ build/
 # Vendored urun-openai proxy binary (scripts/vendor_openai_proxy.py puts it here
 # at release-build time; hatch `artifacts` ships it in the wheel despite this ignore)
 src/urun/_bin/urun-openai
+
+# Agent tooling (Serena MCP) — never ships with the CLI.
+.serena/

{urun_cli-0.6.0 → urun_cli-0.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: urun-cli
-Version: 0.6.0
+Version: 0.6.2
 Summary: End-user CLI for deploying apps to urun
 Project-URL: Homepage, https://urun.sh
 Project-URL: Repository, https://github.com/urun-sh/urun-cli
@@ -98,14 +98,14 @@ In this release, `urun run` uses the same deploy pipeline as `urun deploy`.
 `deploy` remains available as the lower-level command while the full
 deploy/run/monitor workflow is being built.
 
-### Launch an official browser demo
+### Attach an official browser frontend
 
 For official `urun-examples` apps, the shortest browser stream path is:
 
 ```bash
 cd ~/workspace/urun-examples
 urun deploy prompt_canary/backend/app.py --name prompt-canary --timeout 900
-urun demo prompt-canary --examples-root ~/workspace/urun-examples
+urun dev --attach prompt-canary --examples-root ~/workspace/urun-examples
 ```
 
 The positional argument is the **deployed app slug** (as shown by `urun app
@@ -118,7 +118,7 @@ list`). Nothing is declared in a manifest — every value is discovered:
 | local port | the `--port N` in that package's `scripts.dev`, else 3000 |
 | function | read from the deployed app itself |
 
-`urun demo` uses your stored API key server-side to request a short-lived,
+`urun dev --attach` uses your resolved API key server-side to request a short-lived,
 function-scoped, origin-pinned browser JWT, sets the frontend's
 `NEXT_PUBLIC_*` environment variables, starts the local Next.js frontend, and
 opens it in your browser. The API key itself never reaches the page.
@@ -129,15 +129,15 @@ the right directory yourself with `--frontend-dir` (relative to the examples
 root):
 
 ```bash
-urun demo test-matrix-game --frontend-dir matrix-game-3/frontend
+urun dev --attach test-matrix-game --frontend-dir matrix-game-3/frontend
 ```
 
-#### There is no curated list of demoable examples
+#### There is no curated list of attachable examples
 
-`urun demo` will launch **any** `<example>/frontend` it can match, and it does
-not know which examples actually work. Earlier releases read a manifest that
-carried a `supported` flag with a reason, and refused the examples marked
-unsupported; that allowlist is gone along with the manifest.
+`urun dev --attach` will launch **any** `<example>/frontend` it can match, and
+it does not know which examples actually work. Earlier releases read a
+manifest that carried a `supported` flag with a reason, and refused the
+examples marked unsupported; that allowlist is gone along with the manifest.
 
 The practical consequence: a frontend that is a stub, half-built, or known
 broken now starts a dev server and opens a browser instead of failing up
@@ -371,6 +371,17 @@ Enable a disabled app and bring it back online:
 urun app enable lingbot-handle
 ```
 
+If `urun deploy` or `urun dev --attach` reports that a function was paused
+after repeated startup failures, inspect its runtime logs before overriding
+the safety pause. Then explicitly resume every paused function:
+
+```bash
+urun app enable rapid-clips --force
+```
+
+`--force` may restart a known-crashing function; it does not change the
+platform's automatic pause policy.
+
 Stop a queued, starting, or live session:
 
 ```bash

{urun_cli-0.6.0 → urun_cli-0.6.2}/README.md

@@ -76,14 +76,14 @@ In this release, `urun run` uses the same deploy pipeline as `urun deploy`.
 `deploy` remains available as the lower-level command while the full
 deploy/run/monitor workflow is being built.
 
-### Launch an official browser demo
+### Attach an official browser frontend
 
 For official `urun-examples` apps, the shortest browser stream path is:
 
 ```bash
 cd ~/workspace/urun-examples
 urun deploy prompt_canary/backend/app.py --name prompt-canary --timeout 900
-urun demo prompt-canary --examples-root ~/workspace/urun-examples
+urun dev --attach prompt-canary --examples-root ~/workspace/urun-examples
 ```
 
 The positional argument is the **deployed app slug** (as shown by `urun app
@@ -96,7 +96,7 @@ list`). Nothing is declared in a manifest — every value is discovered:
 | local port | the `--port N` in that package's `scripts.dev`, else 3000 |
 | function | read from the deployed app itself |
 
-`urun demo` uses your stored API key server-side to request a short-lived,
+`urun dev --attach` uses your resolved API key server-side to request a short-lived,
 function-scoped, origin-pinned browser JWT, sets the frontend's
 `NEXT_PUBLIC_*` environment variables, starts the local Next.js frontend, and
 opens it in your browser. The API key itself never reaches the page.
@@ -107,15 +107,15 @@ the right directory yourself with `--frontend-dir` (relative to the examples
 root):
 
 ```bash
-urun demo test-matrix-game --frontend-dir matrix-game-3/frontend
+urun dev --attach test-matrix-game --frontend-dir matrix-game-3/frontend
 ```
 
-#### There is no curated list of demoable examples
+#### There is no curated list of attachable examples
 
-`urun demo` will launch **any** `<example>/frontend` it can match, and it does
-not know which examples actually work. Earlier releases read a manifest that
-carried a `supported` flag with a reason, and refused the examples marked
-unsupported; that allowlist is gone along with the manifest.
+`urun dev --attach` will launch **any** `<example>/frontend` it can match, and
+it does not know which examples actually work. Earlier releases read a
+manifest that carried a `supported` flag with a reason, and refused the
+examples marked unsupported; that allowlist is gone along with the manifest.
 
 The practical consequence: a frontend that is a stub, half-built, or known
 broken now starts a dev server and opens a browser instead of failing up
@@ -349,6 +349,17 @@ Enable a disabled app and bring it back online:
 urun app enable lingbot-handle
 ```
 
+If `urun deploy` or `urun dev --attach` reports that a function was paused
+after repeated startup failures, inspect its runtime logs before overriding
+the safety pause. Then explicitly resume every paused function:
+
+```bash
+urun app enable rapid-clips --force
+```
+
+`--force` may restart a known-crashing function; it does not change the
+platform's automatic pause policy.
+
 Stop a queued, starting, or live session:
 
 ```bash

{urun_cli-0.6.0 → urun_cli-0.6.2}/pyproject.toml

@@ -10,24 +10,17 @@ build-backend = "hatchling.build"
 
 [project]
 name = "urun-cli"
-# Cut a release for the 11 commits stranded since v0.5.32 (2026-08-20).
+# Cut the paused-function recovery fixes merged after v0.6.0.
 #
 # WHY THIS IS EDITED BY HAND. [tool.semantic_release] below is fully
 # configured -- conventional commit parser, version_toml pointing here, tag
 # format, changelog templates -- but NOTHING invokes it: `grep -rn
 # semantic.release .github/workflows/` matches nothing. release.yaml only
 # READS this value and skips publishing when the matching tag already exists,
-# so the version has been frozen since v0.5.32 was tagged and every merge
-# after it published nothing. CHANGELOG.md stopping at 0.5.4 is the same gap
-# seen from the other end. Wiring semantic-release into CI is the real fix and
-# is deliberately NOT bundled here; this bump is the reversible unblock.
-#
-# 0.6.0 is semantic-release's OWN answer, not a guess:
-#   $ uvx --from python-semantic-release semantic-release --noop version --print
-#   0.6.0
-# (several feat: commits, and major_on_zero = false keeps 0.x on a minor bump)
-# so choosing it leaves the numbering where automation would resume.
-version = "0.6.0"
+# so every merge remains on the already-published version until this value is
+# advanced. Wiring semantic-release into CI remains separate work; this patch
+# bump is the minimal release unblock for the three post-v0.6.0 fixes.
+version = "0.6.2"
 description = "End-user CLI for deploying apps to urun"
 readme = "README.md"
 requires-python = ">=3.11"

{urun_cli-0.6.0 → urun_cli-0.6.2}/src/urun/api.py

@@ -607,6 +607,23 @@ class ApiClient:
         body: dict[str, Any] | None = {"environment": environment} if environment else None
         return self._json("POST", f"/app/{urllib.parse.quote(app, safe='')}/activate", body)
 
+    def set_function_paused(
+        self,
+        app: str,
+        function_name: str,
+        paused: bool,
+        environment: str | None = None,
+    ) -> dict[str, Any]:
+        """Pause or resume one function of an app.
+
+        Hits ``POST /app/<slug>/pause`` with the function name and explicit
+        pause state. The org is derived from the API key.
+        """
+        body: dict[str, Any] = {"paused": paused, "function_name": function_name}
+        if environment:
+            body["environment"] = environment
+        return self._json("POST", f"/app/{urllib.parse.quote(app, safe='')}/pause", body)
+
     def retire_app(self, app: str, environment: str | None = None) -> dict[str, Any]:
         """Retire an app through the CLI edge route (backs ``urun app delete``).
 

{urun_cli-0.6.0 → urun_cli-0.6.2}/src/urun/catalog.py

@@ -115,10 +115,11 @@ ENGINE_DEPS: dict[str, list[str]] = {
         "safetensors",
         "huggingface_hub>=1.5.0,<2.0",
         "opencv-python-headless==4.11.0.86",
-        "nvidia-cuda-nvcc==13.3.73",
-        "nvidia-cuda-crt==13.3.73",
-        "nvidia-cuda-cccl==13.3.3.4.1",
-        "nvidia-cuda-runtime==13.3.29",
+        "nvidia-cuda-nvcc==13.0.88",
+        "nvidia-cuda-crt==13.0.88",
+        "nvidia-cuda-cccl==13.0.85",
+        "nvidia-nvvm==13.0.88",
+        "nvidia-cuda-runtime==13.0.96",
         "numpy<2",
         "pillow>=10.0.0",
         "orjson",
@@ -155,10 +156,11 @@ ENGINE_DEPS: dict[str, list[str]] = {
         "accelerate",
         "safetensors",
         "huggingface_hub>=1.5.0,<2.0",
-        "nvidia-cuda-nvcc==13.3.73",
-        "nvidia-cuda-crt==13.3.73",
-        "nvidia-cuda-cccl==13.3.3.4.1",
-        "nvidia-cuda-runtime==13.3.29",
+        "nvidia-cuda-nvcc==13.0.88",
+        "nvidia-cuda-crt==13.0.88",
+        "nvidia-cuda-cccl==13.0.85",
+        "nvidia-nvvm==13.0.88",
+        "nvidia-cuda-runtime==13.0.96",
         "numpy<2",
         "pillow>=10.0.0",
         "orjson",
@@ -357,10 +359,11 @@ ENGINE_DEPS: dict[str, list[str]] = {
         "safetensors",
         "huggingface_hub>=1.5.0,<2.0",
         "flashinfer_python[cu13]==0.6.14",
-        "nvidia-cuda-nvcc==13.2.78",
-        "nvidia-cuda-crt==13.2.78",
-        "nvidia-cuda-cccl==13.2.75",
-        "nvidia-cuda-runtime==13.2.86",
+        "nvidia-cuda-nvcc==13.0.88",
+        "nvidia-cuda-crt==13.0.88",
+        "nvidia-cuda-cccl==13.0.85",
+        "nvidia-nvvm==13.0.88",
+        "nvidia-cuda-runtime==13.0.96",
         "numpy<2",
         "pillow>=10.0.0",
         "orjson",
@@ -670,13 +673,21 @@ def _published_cuda_for_torch(minor: tuple[int, int]) -> tuple[int, ...] | None:
     return None
 
 
-def torch_backend_for_serve(engine: str | None, gpu_spec: str) -> str:
+def torch_backend_for_serve(
+    engine: str | None,
+    gpu_spec: str,
+    *,
+    engine_deps: list[str] | None = None,
+) -> str:
     """The ``deps.torch_backend`` value to bake for a catalog serve.
 
     ARCH-ANCHORED mirror of urun-python ``resolve_torch_backend`` (KEEP IN
     SYNC): the GPU-arch map value (:func:`torch_backend_for_gpu`) encodes the
     runtime base image's CUDA ABI and is kept byte-identical whenever it
-    publishes the ENGINE's pinned torch (from :data:`ENGINE_DEPS`). When the
+    publishes the pinned torch. ``engine_deps`` (a MODEL-level override) is
+    the pin SOURCE when provided — the SAME list the shim's deps literal and
+    ``URUN_SERVE_DEPS`` carry, so one derivation anchors the backend; None
+    falls back to the ENGINE's group (from :data:`ENGINE_DEPS`). When the
     arch lane provably cannot publish the pin, repair to a lane that does AND
     meets the GPU's kernel floor — preferring the ABI-compatible direction:
     DOWN (newest below the anchor) for the pre-cxx11 torch<=2.6 era, the
@@ -712,7 +723,8 @@ def torch_backend_for_serve(engine: str | None, gpu_spec: str) -> str:
             "verdict. Deploy a plain @app.function on gh200 instead."
         )
     arch_backend = torch_backend_for_gpu(gpu_spec)
-    minor = _torch_minor(pinned_torch_version(serve_deps_for_engine(engine)))
+    dep_list = engine_deps if engine_deps is not None else serve_deps_for_engine(engine)
+    minor = _torch_minor(pinned_torch_version(dep_list))
     if minor is None:
         return arch_backend
     published = _published_cuda_for_torch(minor)
@@ -735,6 +747,24 @@ def torch_backend_for_serve(engine: str | None, gpu_spec: str) -> str:
     return f"cu{chosen}"
 
 
+def serve_deps_for_row(row: CatalogRow) -> list[str]:
+    """The python dependency list for THIS row's serve venv.
+
+    MODEL-level selection: an explicit ``row.engine_deps`` is the COMPLETE
+    list (requirement strings — including pinned wheel URLs — pass through
+    unchanged); ``None`` falls back to the ENGINE-level source of truth
+    (:func:`serve_deps_for_engine`, ``strict=True`` — an engine the map does
+    not know fails the deploy, never a silently empty venv).
+    """
+    # The engine itself is validated strictly on BOTH paths: an unknown engine
+    # with an override would otherwise reach shim generation and the
+    # function-engines sidecar, moving a clear pre-deploy failure into the pod.
+    default = serve_deps_for_engine(row.engine, strict=True)
+    if row.engine_deps is not None:
+        return list(row.engine_deps)
+    return default
+
+
 @dataclass(frozen=True)
 class CatalogRow:
     """One flattened ``(model_id, variant, gpu_spec)`` catalog placement.
@@ -764,6 +794,10 @@ class CatalogRow:
     default_gpu: str | None = None
     min_vram_gb: int | None = None
     min_host_ram_gb: int | None = None
+    # MODEL-level python dependency override (additive — older catalogs predate
+    # it): null = the ENGINE-level dep group, verbatim; an explicit list is the
+    # COMPLETE replacement (never merged with the engine's group).
+    engine_deps: list[str] | None = None
 
     @property
     def ref(self) -> str:
@@ -790,6 +824,40 @@ class CatalogRow:
         return self.gpu_count == 1 and not self.disagg
 
 
+# PEP 508 requirement SHAPE (stdlib-only; this module deliberately carries no
+# `packaging` dependency): ``name[extras] (@ url | version clauses)? (; marker)?``
+# with a real name and a version after every operator — enough to refuse the
+# garbage a hand-typed catalog row can carry (``vllm===``, ``==``, ``torch 2``)
+# at the gate instead of minutes later inside the venv build.
+_REQ_NAME = r"[A-Za-z0-9](?:[A-Za-z0-9._-]*[A-Za-z0-9])?"
+_REQ_VERSION_CLAUSE = r"(?:===|==|!=|<=|>=|~=|<|>)\s*[A-Za-z0-9][A-Za-z0-9.*+!_-]*"
+_REQUIREMENT_SHAPE_RE = re.compile(
+    rf"^{_REQ_NAME}\s*(?:\[[A-Za-z0-9._,\s-]*\])?\s*"
+    rf"(?:@\s*\S+|{_REQ_VERSION_CLAUSE}(?:\s*,\s*{_REQ_VERSION_CLAUSE})*)?"
+    rf"\s*(?:;.+)?$"
+)
+
+
+def _is_requirement_shaped(dep: str) -> bool:
+    """Whether ``dep`` has the shape of a PEP 508 requirement string."""
+    return _REQUIREMENT_SHAPE_RE.match(dep.strip()) is not None
+
+
+def _requirement_url_carries_credentials(dep: str) -> bool:
+    """Whether a direct-URL requirement (``name @ url``) embeds userinfo.
+
+    The catalog is public reference data (anon-readable) and the resolved list
+    is echoed into deploy artifacts, so a ``https://token@host/...`` URL would
+    publish a credential. Private indexes go through a secret-managed auth
+    path, never through the row.
+    """
+    _name, sep, url = dep.partition("@")
+    if not sep:
+        return False
+    parts = urllib.parse.urlsplit(url.strip())
+    return bool(parts.username or parts.password)
+
+
 def parse_row(raw: dict[str, Any]) -> CatalogRow:
     """Build a :class:`CatalogRow` from one PostgREST JSON object.
 
@@ -825,6 +893,36 @@ def parse_row(raw: dict[str, Any]) -> CatalogRow:
         value = raw.get(key)
         return value if isinstance(value, int) and not isinstance(value, bool) else None
 
+    # Additive MODEL-level dep override: ABSENT (older catalogs) and SQL NULL
+    # both mean "the engine's own dep group". An explicit value REPLACES that
+    # group verbatim, so anything malformed fails LOUDLY here — never a silent
+    # fall-back to the default (old) engine stack.
+    engine_deps = raw.get("engine_deps")
+    if engine_deps is not None:
+        if (
+            not isinstance(engine_deps, list)
+            or not engine_deps
+            or not all(isinstance(dep, str) and dep.strip() for dep in engine_deps)
+        ):
+            raise UrunError(
+                "catalog row 'engine_deps' must be null or a non-empty list of "
+                f"non-blank requirement strings, got {engine_deps!r}"
+            )
+        bad = [dep for dep in engine_deps if not _is_requirement_shaped(dep)]
+        if bad:
+            raise UrunError(
+                "catalog row 'engine_deps' entries must be PEP 508 requirement "
+                f"strings (name[extras] @ url | version clauses ; marker), got {bad!r}"
+            )
+        for dep in engine_deps:
+            if _requirement_url_carries_credentials(dep):
+                # Name only: never echo the credential-bearing URL.
+                raise UrunError(
+                    "catalog row 'engine_deps' entry for "
+                    f"{dep.partition('@')[0].strip()!r} embeds credentials in its URL; "
+                    "the catalog is public reference data — use a secret-managed index"
+                )
+
     return CatalogRow(
         model_id=req_str("model_id"),
         variant=req_str("variant"),
@@ -842,6 +940,7 @@ def parse_row(raw: dict[str, Any]) -> CatalogRow:
         default_gpu=default_gpu,
         min_vram_gb=opt_int("min_vram_gb"),
         min_host_ram_gb=opt_int("min_host_ram_gb"),
+        engine_deps=list(engine_deps) if engine_deps is not None else None,
     )