arbiter-engine 0.2.3__tar.gz → 0.2.4__tar.gz

{arbiter_engine-0.2.3 → arbiter_engine-0.2.4}/CHANGELOG.md

@@ -24,14 +24,40 @@ Nothing yet.
 
 ---
 
+## [0.2.4] — 2026-09-18
+
+### Fixed
+
+- **The MCP server raised at construction whenever the SDK was installed.**
+  `rollout` and `plan` were added to `TOOL_SPECS` and to `_HANDLERS` in 0.2.3
+  and not to the wrapper table, so `build_server()` hit its own guard —
+  *TOOL_SPECS declares ['plan', 'rollout'] with no wrapper to register* — and
+  the `mcp` extra was unusable for the whole of 0.2.3. `dispatch` served both
+  verbs correctly the entire time; only the transport could not.
+
+  **The guard was right and fired too late.** Constructing a server needs the
+  optional SDK, so the only lane that could reach it was the one installing the
+  extra: every other lane skipped the transport test and reported green. 0.2.3
+  was verified locally at 1660 passed / 14 skipped and was red in CI at 1668
+  passed / 5 skipped. The skip COUNT was the only available signal and nothing
+  compares it across lanes.
+
+  `test_every_declared_tool_has_a_wrapper.py` now reads the wrapper table with
+  the AST instead of importing it, so a declared tool with no wrapper fails on
+  EVERY lane without the SDK present. Installing the extra everywhere would
+  have made an optional dependency required in all but name, which is the
+  measured two-dependency claim this package keeps.
+
+---
+
 ## [0.2.3] — 2026-09-18
 
 **The first MINOR, and it is not numbered 0.2.0.** `arbiter-engine` stayed on
 `0.1.x` by decision through eighteen patch releases; `rollout` and `plan` are
 new verbs on `api`, and COMPATIBILITY.md has always said that is a minor.
 
-**0.2.0 AND 0.2.1 were both skipped because both are unavailable**, not because
-anything was wrong with either — see
+**0.2.0, 0.2.1, 0.2.2 and 0.3.0 were all skipped because all four are
+unavailable**, not because anything was wrong with any of them — see
 [Version numbers that do not exist](#version-numbers-that-do-not-exist). PyPI
 permanently reserves any filename it has ever served and had deleted, including
 under an earlier owner of the name, so an upload returns `400 This filename was
@@ -40,10 +66,12 @@ sequence runs 0.1.1 to 0.1.4.
 
 That section predicted this release and named the wrong number: it said *the
 version that first breaks compatibility cannot BE 0.2.0. so it will be
-0.2.1.* 0.2.1 was an INFERENCE from three measured numbers, and it was wrong.
-An upload is the only oracle — a 404 does not mean a filename is free — so the
-prediction could not have been checked when it was written, and it is corrected
-here by having been run.
+0.2.1.* 0.2.1 was an INFERENCE from three measured numbers, and it was wrong —
+as were 0.2.2 and 0.3.0 after it. Three uploads were refused before one landed,
+and the reserved set turned out NOT to be contiguous: 0.3.0 is taken and 0.2.3
+was free. An upload is the only oracle — a 404 does not mean a filename is
+free — so no amount of reading could have ordered these, and nothing shorter
+than trying them would have found it.
 
 Every consumer pinning `<0.2` must move its ceiling to `<0.3` to see this —
 that is the ceiling doing its job, not breaking.
@@ -2229,16 +2257,17 @@ topology traversal, an MCP transport, a YAML domain-model loader; `numpy` and
 
 ## Version numbers that do not exist
 
-**0.1.2, 0.1.3, 0.2.0 and 0.2.1 are permanently unavailable** and were never
-published by this project. PyPI reserves any filename that has ever been used and
+**0.1.2, 0.1.3, 0.2.0, 0.2.1, 0.2.2 and 0.3.0 are permanently unavailable** and
+were never published by this project. The set is NOT a contiguous range and
+cannot be extrapolated: 0.2.3 was free while 0.3.0 was not. PyPI reserves any filename that has ever been used and
 deleted, including from an earlier owner of the name, so an upload under those
 numbers returns `400 This filename was previously used by a file that has since
-been deleted`. The sequence runs 0.1.1 to 0.1.4, and 0.1.18 to 0.2.2, for that
+been deleted`. The sequence runs 0.1.1 to 0.1.4, and 0.1.18 to 0.2.3, for that
 reason and no other.
 
-**0.2.1 was added to this list by trying it**, and that is the point worth
-keeping. Until 2026-09-18 this paragraph named three numbers and then predicted a
-fourth: *the version that first breaks compatibility cannot BE 0.2.0 — the number
+**0.2.1, 0.2.2 and 0.3.0 were added to this list by trying them**, and that is
+the point worth keeping. Until 2026-09-18 this paragraph named three numbers and
+then predicted a fourth: *the version that first breaks compatibility cannot BE 0.2.0 — the number
 is unavailable — so it will be 0.2.1.* The three were measured; the fourth was an
 inference from them, and the release that needed it found it false at the upload.
 
@@ -2246,7 +2275,7 @@ inference from them, and the release that needed it found it false at the upload
 free, the JSON and `/simple/` surfaces list only what is live, and a deleted
 release leaves no trace either serves. So a number in this list was measured and a
 number predicted from it was a guess wearing the same sentence — which is why the
-prediction is gone rather than re-pointed at 0.2.2.
+prediction is gone rather than re-pointed at the next number.
 
 **This matters to anyone pinning `<0.2`.** That ceiling stops at 0.1.18 and does
 not resolve the 0.2 series at all. Move it to `<0.3` deliberately; a tool or a

{arbiter_engine-0.2.3 → arbiter_engine-0.2.4}/CITATION.cff

@@ -9,7 +9,7 @@ authors:
     alias: "james-sheen"
   - name: "The Arbiter Project Authors"
 license: Apache-2.0
-version: "0.2.3"
+version: "0.2.4"
 date-released: "2026-09-18"
 repository-code: "https://github.com/james-sheen/arbiter"
 url: "https://github.com/james-sheen/arbiter"

{arbiter_engine-0.2.3 → arbiter_engine-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: arbiter-engine
-Version: 0.2.3
+Version: 0.2.4
 Summary: A detection engine that reports what it did not check.
 Project-URL: Homepage, https://github.com/james-sheen/arbiter
 Project-URL: Source, https://github.com/james-sheen/arbiter

{arbiter_engine-0.2.3 → arbiter_engine-0.2.4}/arbiter_engine/mcp/server.py

@@ -534,6 +534,29 @@ def build_server(session: EngineSession | None = None):
             "overrides": overrides,
         })
 
+    async def rollout_tool(
+        actions: List[Dict[str, Any]] | None = None,
+        horizon_s: float = 3600.0,
+        step_s: float = 300.0,
+        seed_mode: str = "current",
+        max_transitions: int = 100_000,
+    ) -> str:
+        return _emit("rollout", {
+            "actions": actions, "horizon_s": horizon_s, "step_s": step_s,
+            "seed_mode": seed_mode, "max_transitions": max_transitions,
+        })
+
+    async def plan_tool(
+        candidates: List[Dict[str, Any]] | None = None,
+        horizon_s: float = 3600.0,
+        step_s: float = 300.0,
+        max_transitions: int = 100_000,
+    ) -> str:
+        return _emit("plan", {
+            "candidates": candidates, "horizon_s": horizon_s,
+            "step_s": step_s, "max_transitions": max_transitions,
+        })
+
     async def gaps_tool(start_node: str | None = None) -> str:
         return _emit("gaps", {"start_node": start_node})
 
@@ -610,6 +633,15 @@ def build_server(session: EngineSession | None = None):
         "load_model": load_model_tool,
         "add_entity": add_entity_tool,
         "add_observations": add_observations_tool,
+        # The SAME omission the comment above records, made again by
+        # the world-model arc and caught by the same guard -- at construction,
+        # on the one lane that installs the `mcp` extra, AFTER 0.2.3 shipped.
+        # Every local lane skips the transport test because the SDK is absent,
+        # so 1660 local passes and 14 skips read as green while CI read 1668
+        # passes and 5 skips. The skip count is the tell, and nothing compares
+        # the two.
+        "rollout": rollout_tool,
+        "plan": plan_tool,
     }
     missing = {spec["name"] for spec in TOOL_SPECS} - set(wrappers)
     if missing:

{arbiter_engine-0.2.3 → arbiter_engine-0.2.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "arbiter-engine"
-version = "0.2.3"
+version = "0.2.4"
 description = "A detection engine that reports what it did not check."
 requires-python = ">=3.10"
 license = "Apache-2.0"

{arbiter_engine-0.2.3 → arbiter_engine-0.2.4}/tests/test_a_declared_floor_is_checked_on_every_path.py

@@ -34,11 +34,12 @@ def _examples_dir() -> pathlib.Path:
     """Whichever copy this tree has: the built package ships `examples/` at its
     root, the source tree keeps them under the publication docs.
 
-    - this was a path assembled from the SOURCE repository's own name,
-    which is two defects at once. It named a private tree in a file that ships,
-    and it could only ever resolve in the tree it was written in, so the shipped
-    copy of this test would have failed at import on a reader's machine. The
-    resolver here is the one `test_every_shipped_example_actually_runs.py`
+    - this was a hard-coded path into a directory that does not exist
+    in the published package, so the shipped copy of this test could only ever
+    fail at import on a reader's machine. It was written where it happened to
+    resolve and nothing between there and the wheel disagreed.
+
+    The resolver here is the one `test_every_shipped_example_actually_runs.py`
     already uses; copying its shape rather than inventing a second one is the
     point, since two resolvers for one fact are what drift.
     """

arbiter_engine-0.2.4/tests/test_every_declared_tool_has_a_wrapper.py

@@ -0,0 +1,110 @@
+"""Every name in ``TOOL_SPECS`` has a wrapper, checked without the SDK.
+
+THE GUARD THAT EXISTS ALREADY RUNS TOO LATE AND ON TOO FEW LANES.
+`build_server()` raises `RuntimeError: TOOL_SPECS declares [...] with no wrapper
+to register` at construction, which is the right behaviour and the right
+message. But constructing the server needs the `mcp` SDK, so the only lane that
+can reach it is the one installing the extra -- and every other lane SKIPS the
+transport test and reports green.
+
+That is not hypothetical. The four disciplines went in without wrappers once,
+and `server.py` carries a comment about it. The world-model arc then did the
+same thing with `rollout` and `plan`, and it shipped: 0.2.3 went to the index
+with an MCP server that raised on construction for anyone who had the SDK
+installed. Locally the suite was 1660 passed / 14 skipped; in CI it was 1668
+passed / 5 skipped and red. The skip COUNT was the only local signal, and
+nothing compares it across lanes.
+
+SO THIS READS THE SOURCE INSTEAD OF IMPORTING IT. The wrapper table is a dict
+literal inside `build_server`, so its keys are visible to the AST without the
+function ever running and without the SDK being present. A name in `TOOL_SPECS`
+and not in that dict fails here on EVERY lane, at collection speed.
+
+WHY NOT JUST INSTALL THE SDK IN EVERY LANE. Because the dependency is optional
+by decision -- the README's two-dependency claim is measured, and the transport
+is behind an extra. A guard that requires the optional thing in order to check
+the required thing has made it required in all but name.
+"""
+
+from __future__ import annotations
+
+import ast
+import pathlib
+
+import pytest
+
+# DOTTED, and that is not a style choice. `build-engine.sh` rewrites absolute
+# imports with a regex anchored on `arbiter_engine\.` -- a DOT is required --
+# so `from arbiter_engine import arbiter_mcp` passes through unrewritten and
+# the scrub then refuses the built tree for naming the source package. Measured:
+# that exact line refused the 0.2.4 build.
+from arbiter_engine.mcp import server as _server
+from arbiter_engine.mcp.server import TOOL_SPECS
+
+SERVER = pathlib.Path(_server.__file__)
+
+#: The dict whose keys are the wrappers. Named here so a rename of the local
+#: variable fails loudly rather than silently emptying this test.
+WRAPPER_TABLE = "wrappers"
+
+
+def _wrapper_names() -> set:
+    """Keys of the `wrappers = {...}` literal, read with the AST.
+
+    Assigned inside `build_server`, so the walk is over the whole module rather
+    than the module body -- and the assignment is found by TARGET NAME, not by
+    position, because a literal found by position is one an edit above it
+    silently redefines.
+    """
+    tree = ast.parse(SERVER.read_text(encoding="utf-8"))
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Assign) or not isinstance(node.value, ast.Dict):
+            continue
+        if not any(isinstance(t, ast.Name) and t.id == WRAPPER_TABLE
+                   for t in node.targets):
+            continue
+        return {k.value for k in node.value.keys
+                if isinstance(k, ast.Constant) and isinstance(k.value, str)}
+    raise AssertionError(
+        f"no `{WRAPPER_TABLE} = {{...}}` dict literal in {SERVER.name}; this "
+        f"test reads that table and cannot see a renamed or restructured one")
+
+
+def test_the_wrapper_table_is_readable():
+    """Guards the guard. A resolver that returned an empty set would make the
+    comparison below vacuously pass for a table that lost every entry."""
+    names = _wrapper_names()
+    assert len(names) >= 12, sorted(names)
+
+
+def test_there_are_specs_to_check():
+    assert len(TOOL_SPECS) >= 12
+
+
+def test_every_declared_tool_has_a_wrapper():
+    """The one that would have caught `rollout` and `plan` before the release."""
+    declared = {spec["name"] for spec in TOOL_SPECS}
+    missing = sorted(declared - _wrapper_names())
+    assert not missing, (
+        f"{missing} are in TOOL_SPECS with no entry in the wrapper table, so "
+        f"the server would advertise a tool it cannot serve and `build_server` "
+        f"would raise at construction -- but only on a lane that installs the "
+        f"`mcp` extra")
+
+
+def test_no_wrapper_is_registered_for_an_undeclared_tool():
+    """The other direction. A wrapper with no spec is dead code that reads as
+    a served tool, and `build_server` iterates TOOL_SPECS so it would never be
+    registered at all."""
+    stray = sorted(_wrapper_names() - {spec["name"] for spec in TOOL_SPECS})
+    assert not stray, f"{stray} have wrappers and no TOOL_SPECS entry"
+
+
+def test_the_two_new_verbs_are_among_them():
+    """Named explicitly, because a set-comparison test stays green if BOTH
+    sides lose a member together -- which is exactly how the pair went missing
+    from the dict while remaining in `_HANDLERS`."""
+    declared = {spec["name"] for spec in TOOL_SPECS}
+    for verb in ("rollout", "plan"):
+        assert verb in declared, f"{verb} left TOOL_SPECS"
+        assert verb in _wrapper_names(), f"{verb} has no wrapper"