aria-code 0.108.0__tar.gz → 0.110.0__tar.gz

{aria_code-0.108.0/src/aria_code.egg-info → aria_code-0.110.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aria-code
-Version: 0.108.0
+Version: 0.110.0
 Summary: Open-source AI agent for coding, financial analysis and logistics operations: numbers you can check, client data kept apart, actions a person approves
 License: Apache-2.0
 Project-URL: Homepage, https://github.com/artheras/aria-code

{aria_code-0.108.0 → aria_code-0.110.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "aria-code"
-version = "0.108.0"
+version = "0.110.0"
 description = "Open-source AI agent for coding, financial analysis and logistics operations: numbers you can check, client data kept apart, actions a person approves"
 readme = "README.md"
 license = { text = "Apache-2.0" }

{aria_code-0.108.0 → aria_code-0.110.0}/src/aria_code/_version.py

@@ -17,4 +17,4 @@ from __future__ import annotations
 
 __all__ = ["__version__"]
 
-__version__ = "0.108.0"
+__version__ = "0.110.0"

{aria_code-0.108.0 → aria_code-0.110.0}/src/aria_code/apps/cli/prompts/select.py

@@ -102,7 +102,12 @@ OUTPUT_DISCIPLINE = (
     "Before you finish, confirm it exists by reading it back or listing the folder. Describing the "
     "result in your reply does not create the file.\n"
     "- Compute totals, counts, averages and other figures from the data with code (run a short "
-    "script), not by hand. Every number you report or write must come from a tool's output."
+    "script), not by hand. Every number you report or write must come from a tool's output.\n"
+    "- Something written for one client, customer or shipper contains only their data. Data "
+    "files often hold several clients: filter to theirs, and leave out every other client's "
+    "name, IDs, SKUs and totals, including any that appear inside this client's own rows "
+    "(notes, comments, references). Copy free-text fields into a client document only after "
+    "checking them for that."
 )
 
 

{aria_code-0.108.0 → aria_code-0.110.0}/src/aria_code/apps/cli/update_check.py

@@ -139,7 +139,7 @@ def _worker(current: str, lang: str, channel: str = "native") -> None:
         import urllib.request
         req = urllib.request.Request(
             source_url,
-            headers={"Accept": "application/vnd.github+json", "User-Agent": "aria-code-update-check"},
+            headers={"Accept": "application/json", "User-Agent": "aria-code-update-check"},
         )
         with urllib.request.urlopen(req, timeout=_FETCH_TIMEOUT) as resp:
             data   = json.loads(resp.read())

{aria_code-0.108.0 → aria_code-0.110.0}/src/aria_code/apps/cli/updater.py

@@ -21,8 +21,8 @@ from aria_code.apps.cli.update_check import (
 )
 
 
-def _request(url: str) -> urllib.request.Request:
-    return urllib.request.Request(url, headers={"User-Agent": "aria-code-updater", "Accept": "application/vnd.github+json"})
+def _request(url: str, *, accept: str = "application/json") -> urllib.request.Request:
+    return urllib.request.Request(url, headers={"User-Agent": "aria-code-updater", "Accept": accept})
 
 
 def _fetch_json(url: str) -> dict:
@@ -51,7 +51,7 @@ def check_update(channel: str) -> tuple[str, dict]:
 
 
 def _download(url: str, target: Path) -> None:
-    with urllib.request.urlopen(_request(url), timeout=60) as response, target.open("wb") as output:
+    with urllib.request.urlopen(_request(url, accept="*/*"), timeout=60) as response, target.open("wb") as output:
         shutil.copyfileobj(response, output)
 
 

{aria_code-0.108.0 → aria_code-0.110.0}/src/aria_code/evals/harness.py

@@ -132,6 +132,13 @@ class TaskSpec:
     # point is that the agent must not *break* it. Opting out is explicit so
     # that it is a decision someone made, not a fixture that quietly rotted.
     allow_green_start: bool = False
+    # Files the agent never sees: present for the pre-flight, removed (with
+    # the pre-flight's caches) for the agent's turn, put back from the fixture
+    # for the score. A grader that holds the expected answers — EXPECTED =
+    # {"F-201": 5000.00, ...} — or spells out the trap in its docstrings is
+    # an answer key when it sits in the workspace. gemini-3.5-flash passed 23
+    # of the first 24 tasks with every grader in view.
+    hidden: tuple[str, ...] = ()
 
     @classmethod
     def from_dict(cls, data: dict) -> "TaskSpec":
@@ -153,6 +160,7 @@ class TaskSpec:
                 if "protect" in data else cls.protect
             ),
             allow_green_start=bool(data.get("allow_green_start", False)),
+            hidden=tuple(str(g) for g in (data.get("hidden") or ())),
         )
 
 
@@ -366,18 +374,47 @@ def _missing_modules(names: Iterable[str]) -> list[str]:
     return missing
 
 
-def _violates_protection(changed: Iterable[str], patterns: Iterable[str]) -> tuple[str, ...]:
-    """Protected files the agent modified."""
+def _matches(rel: str, globs: Iterable[str]) -> bool:
     import fnmatch
 
+    name = rel.replace("\\", "/")
+    base = name.rsplit("/", 1)[-1]
+    return any(fnmatch.fnmatch(name, g) or fnmatch.fnmatch(base, g) for g in globs)
+
+
+def _violates_protection(changed: Iterable[str], patterns: Iterable[str]) -> tuple[str, ...]:
+    """Protected files the agent modified."""
     globs = tuple(patterns)
-    hits: list[str] = []
-    for rel in changed:
-        name = rel.replace("\\", "/")
-        base = name.rsplit("/", 1)[-1]
-        if any(fnmatch.fnmatch(name, g) or fnmatch.fnmatch(base, g) for g in globs):
-            hits.append(rel)
-    return tuple(hits)
+    return tuple(rel for rel in changed if _matches(rel, globs))
+
+
+def _hide(workspace: Path, globs: Sequence[str]) -> tuple[str, ...]:
+    """Remove the hidden files, and every cache the pre-flight left, from *workspace*.
+
+    The caches matter as much as the files: .pytest_cache records the failing
+    tests by name and __pycache__ holds the grader compiled, docstrings and
+    expected values included.
+    """
+    if not globs:
+        return ()
+    hidden = tuple(
+        str(path.relative_to(workspace)) for path in sorted(workspace.rglob("*"))
+        if path.is_file() and _matches(str(path.relative_to(workspace)), globs)
+    )
+    for rel in hidden:
+        (workspace / rel).unlink()
+    for cache in sorted(workspace.rglob("*"), reverse=True):
+        if cache.is_dir() and cache.name in ("__pycache__", ".pytest_cache"):
+            shutil.rmtree(cache, ignore_errors=True)
+    return hidden
+
+
+def _restore(hidden: Sequence[str], source: Path, workspace: Path) -> None:
+    """Put the hidden files back from the fixture, over anything the agent wrote there."""
+    for rel in hidden:
+        target = workspace / rel
+        target.parent.mkdir(parents=True, exist_ok=True)
+        shutil.copy2(source / rel, target)
 
 
 def _snapshot(root: Path) -> dict[str, str]:
@@ -551,6 +588,7 @@ def run_task(
             )
 
         # ── the agent's turn ──────────────────────────────────────────────
+        hidden = _hide(workspace, task.hidden)
         before = _snapshot(workspace)
         try:
             outcome = solver(task.prompt, workspace)
@@ -565,12 +603,18 @@ def run_task(
         # edited the file correctly and then exited 1 on an empty final message
         # is a PASS: the work is on disk and the tests are green.
         changed = _changed_paths(before, _snapshot(workspace))
+        if hidden:
+            _restore(hidden, fixtures / task.fixture, workspace)
 
         # Checked before the verdict, and it overrides a green check. Observed
         # in a real run: an agent edited test_settlement.py rather than the
         # module under test. The suite would have called that a PASS and the
         # number would have been a lie.
-        tampered = _violates_protection(changed, task.protect)
+        #
+        # Only files that were there for the agent to tamper with: a test file
+        # it writes for itself is ordinary work, and with hidden graders the
+        # tests it is told not to edit are not there at all.
+        tampered = _violates_protection([rel for rel in changed if rel in before], task.protect)
         if tampered:
             return _result(
                 FAIL,

{aria_code-0.108.0 → aria_code-0.110.0/src/aria_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aria-code
-Version: 0.108.0
+Version: 0.110.0
 Summary: Open-source AI agent for coding, financial analysis and logistics operations: numbers you can check, client data kept apart, actions a person approves
 License: Apache-2.0
 Project-URL: Homepage, https://github.com/artheras/aria-code

{aria_code-0.108.0 → aria_code-0.110.0}/tests/test_cli_updater.py

@@ -5,6 +5,7 @@ from pathlib import Path
 import tarfile
 import time
 from types import SimpleNamespace
+from urllib.error import HTTPError
 
 import pytest
 
@@ -32,6 +33,45 @@ def test_npm_launcher_channel_wins_over_frozen_executable(monkeypatch):
     assert update_check._install_channel() == "npm"
 
 
+@pytest.mark.parametrize("background", [False, True])
+def test_npm_updates_survive_registry_content_negotiation(monkeypatch, tmp_path, background):
+    """npm rejects the GitHub-only media type with 406, including /latest."""
+    monkeypatch.setattr(update_check, "_CACHE_FILE", tmp_path / "update.json")
+    monkeypatch.setattr(update_check, "_notice", None)
+    requests = []
+
+    def registry(request, **kwargs):
+        requests.append(request)
+        if request.get_header("Accept") != "application/json":
+            raise HTTPError(request.full_url, 406, "Not Acceptable", {}, None)
+        return io.BytesIO(b'{"version":"0.108.1"}')
+
+    monkeypatch.setattr(updater.urllib.request, "urlopen", registry)
+    if background:
+        update_check._worker("0.108.0", "en", "npm")
+        assert "v0.108.1" in update_check.get_update_notice()
+    else:
+        latest, _ = updater.check_update("npm")
+        assert latest == "0.108.1"
+    assert len(requests) == 1
+    assert requests[0].full_url == update_check._NPM_URL
+    assert update_check._read_cache()["latest"] == "0.108.1"
+
+
+def test_update_downloads_request_the_raw_asset(monkeypatch, tmp_path):
+    payload = b"release checksum or archive bytes"
+
+    def asset(request, **kwargs):
+        if request.get_header("Accept") != "*/*":
+            raise HTTPError(request.full_url, 406, "Not Acceptable", {}, None)
+        return io.BytesIO(payload)
+
+    monkeypatch.setattr(updater.urllib.request, "urlopen", asset)
+    target = tmp_path / "SHA256SUMS"
+    updater._download("https://github.com/artheras/aria-code/releases/download/v0.108.1/SHA256SUMS", target)
+    assert target.read_bytes() == payload
+
+
 def test_pip_update_selects_github_version_and_checks_the_pinned_package(monkeypatch):
     calls = []
     def fetch(url):