linkedparticles-core 1.192.2__tar.gz → 1.192.10__tar.gz

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: linkedparticles-core
-Version: 1.192.2
+Version: 1.192.10
 Summary: Store-free client library for Particles: define, validate, serialize, and extract provenance-tracked belief particles
 Project-URL: Homepage, https://linkedparticles.org
 Project-URL: Documentation, https://docs.linkedparticles.org
@@ -29,7 +29,7 @@ Classifier: Topic :: Software Development :: Libraries
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Classifier: Typing :: Typed
 Requires-Python: >=3.11
-Requires-Dist: anthropic>=0.28
+Requires-Dist: anthropic>=1.0
 Requires-Dist: html2text>=2024.2
 Requires-Dist: httpx>=0.27
 Requires-Dist: jsonschema>=4.22
@@ -37,17 +37,17 @@ Requires-Dist: lxml-stubs>=0.5.1
 Requires-Dist: numpy>=1.26
 Requires-Dist: opentelemetry-api>=1.27
 Requires-Dist: pydantic>=2.7
-Requires-Dist: pyld<4,>=2.0
-Requires-Dist: pypdf>=6.15.0
-Requires-Dist: pyshacl<0.32,>=0.26
+Requires-Dist: pyld<4,>=3.0
+Requires-Dist: pypdf>=6.19.0
+Requires-Dist: pyshacl<0.41,>=0.40
 Requires-Dist: pyyaml>=6.0
 Requires-Dist: rdflib<8,>=7.0
 Requires-Dist: scipy>=1.13
-Requires-Dist: sentence-transformers>=2.7
-Requires-Dist: sqlalchemy[asyncio]<2.2,>=2.0
+Requires-Dist: sentence-transformers>=5.6.0
+Requires-Dist: sqlalchemy[asyncio]<2.2,>=2.1
 Requires-Dist: torch
 Provides-Extra: vision
-Requires-Dist: pypdfium2>=4.30; extra == 'vision'
+Requires-Dist: pypdfium2>=5.0; extra == 'vision'
 Description-Content-Type: text/markdown
 
 # linkedparticles-core

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/constraints.txt

@@ -10,7 +10,7 @@ annotated-doc==0.0.5
     #   typer
 annotated-types==0.8.0
     # via pydantic
-anthropic==1.11.0
+anthropic==1.13.0
     # via linkedparticles
 anyio==4.14.2 ; python_full_version < '3.15'
     # via
@@ -67,9 +67,9 @@ distro==1.9.0
     # via langsmith
 docstring-parser==0.18.0
     # via anthropic
-fastapi==0.142.2
+fastapi==0.143.0
     # via linkedparticles
-filelock==4.0.12
+filelock==4.1.1
     # via
     #   huggingface-hub
     #   linkedparticles
@@ -148,7 +148,7 @@ jsonschema==4.26.0
     #   mcp
 jsonschema-specifications==2025.9.1
     # via jsonschema
-langchain-core==1.6.7
+langchain-core==1.6.10
     # via linkedparticles
 langchain-protocol==0.0.19
     # via langchain-core
@@ -266,7 +266,7 @@ protobuf==7.36.2
     #   opentelemetry-proto
 pycparser==3.0 ; implementation_name != 'PyPy' and platform_python_implementation != 'PyPy'
     # via cffi
-pydantic==2.13.5
+pydantic==2.14.1
     # via
     #   anthropic
     #   fastapi
@@ -275,7 +275,7 @@ pydantic==2.13.5
     #   linkedparticles
     #   mcp
     #   pydantic-settings
-pydantic-core==2.46.5
+pydantic-core==2.50.1
     # via pydantic
 pydantic-settings==2.15.0
     # via mcp
@@ -287,7 +287,7 @@ pyld==3.3.0
     # via linkedparticles
 pyparsing==3.3.3
     # via rdflib
-pypdf==6.19.0
+pypdf==6.20.1
     # via linkedparticles
 pypdfium2==5.14.0
     # via linkedparticles
@@ -356,7 +356,7 @@ sniffio==1.3.1
     # via
     #   anthropic
     #   langsmith
-sqlalchemy==2.1.3
+sqlalchemy==2.1.4
     # via
     #   alembic
     #   linkedparticles

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/config.py

@@ -956,6 +956,12 @@ class ExtractionConfig(BaseModel):
     # The largest delta chunk sent in one call. Each chunk after the first
     # takes the end of the chunk before it as its context.
     append_chunk_chars: int = Field(default=7500, ge=1000)
+    # the agent extractor's prompt (the Stop hook's ask and its API
+    # proxy) tells the model how many claims text of its size usually yields,
+    # as a thoroughness guide, never a quota. 1.9 is the median claims per
+    # 1000 characters of the general extractor's transcript deltas in the
+    # maintainers' store (262 reads, measured 2026-10-16).
+    agent_expected_claims_per_kchar: float = Field(default=1.9, gt=0.0)
     # a pair the §6.6 contradiction probe confirms is read a second
     # time, before the write loop, with each claim's source kind, name, time
     # and passage (on ``llm.verification``). A pair the reading does not
@@ -2801,12 +2807,21 @@ class AgentExtractionConfig(BaseModel):
     # An ask unanswered after this many seconds is expired.
     ask_ttl_seconds: int = Field(default=900, ge=60)
     # A submission covers its span only with at least this many accepted claims
-    # per 1000 characters. 2.7 is the 10th percentile of the general
-    # extractor's own transcript deltas in the maintainers' store.
-    min_claims_per_kchar: float = Field(default=2.7, ge=0.0)
+    # per 1000 characters. 1.27 is the 10th percentile of the general
+    # extractor's own transcript deltas in the maintainers' store (262 delta
+    # reads, measured 2026-10-16; it was 2.7 on 2026-10-09, before the
+    # session-state and tool-turn filters thinned transcript extraction).
+    min_claims_per_kchar: float = Field(default=1.27, ge=0.0)
     # ...and with no more than this share of candidates dropped for evidence
     # not found in the span. Placeholder until the shadow week measures it.
     max_evidence_drop_ratio: float = Field(default=0.5, ge=0.0, le=1.0)
+    # In shadow mode, keep each submission's accepted candidates (claim text,
+    # stated confidence, evidence) in a local JSONL file, so a later labelling
+    # pass can fit a calibration map on real session claims. The
+    # operator event log keeps counts only. Off writes nothing.
+    keep_shadow_candidates: bool = True
+    # None ⇒ ``<claude_code.state_dir>/agent-extraction-shadow.jsonl``.
+    shadow_candidates_path: str | None = None
 
     @field_validator("mode", mode="before")
     @classmethod

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/core/contradiction_disclosure.py

@@ -76,8 +76,7 @@ DISCLOSURE_ACTOR = "memory-consolidate"
 READING_STANDING = "standing/1"
 #: The instruction for a pair with a claim from a time-stamped record (a
 #: session transcript, another append-only log): a value observed at a later
-#: time may be a change, and only the same moment or a fixed fact confirms
-#:.
+#: time may be a change, and only the same moment or a fixed fact confirms.
 READING_OBSERVED = "observed/1"
 
 _EXCERPT_CHARS = 120

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/core/held_release.py

@@ -31,9 +31,9 @@ from particles.core.observer_scope import PairPrecondition
 from particles.core.schema import Particle
 from particles.core.status import Status, StatusReason
 
-#: The retirements an autonomous pass writes: rung 2.5 and its sweep
-#:, the re-anchor pass, and the document sweep
-#:. A claim a person retired is not among them, and a record
+#: The retirements an autonomous pass writes: rung 2.5 and its sweep,
+#: the re-anchor pass, and the document sweep.
+#: A claim a person retired is not among them, and a record
 #: behind one stays with review.
 AUTONOMOUS_RETIREMENTS: frozenset[StatusReason] = frozenset(
     {

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/extraction/agent.py

@@ -55,7 +55,7 @@ log = logging.getLogger(__name__)
 EXTRACTOR_ID = "claude-code-agent"
 #: The agent prompt's version. A calibration map applies only to the version
 #: it was fitted under, so bump this on any prompt change.
-EXTRACTOR_VERSION = "0.1.0"
+EXTRACTOR_VERSION = "0.2.0"
 
 #: The MCP tool the agent submits through.
 SUBMIT_TOOL = "extraction_submit"
@@ -78,16 +78,54 @@ _JSON_ONLY = "Return ONLY a JSON array. No prose before or after. Each element:"
 _AGENT_SCHEMA_FIELD = ',\n  "evidence": "<verbatim passage from the source text>"'
 
 _SUBMIT_INSTRUCTION = """\
-Particles memory: before you finish, extract this session's durable claims.
+Particles memory: before you finish, extract the claims in the session text below.
 
-Read the session text below. It is the part of this conversation not yet
-recorded. Apply the rules, then call the `{tool}` tool exactly once with
-`ask_id` set to "{ask_id}" and `candidates` set to the JSON array described.
-If nothing in the text is worth recording, call it with an empty array.
-Do not summarise what you recorded afterwards; reply with one short line.
+The text is the part of this conversation not yet recorded. When you have
+extracted it, call the `{tool}` tool exactly once with `ask_id` set to
+"{ask_id}" and `candidates` set to the JSON array described. If the text
+holds no claim the rules admit, call it with an empty array. Do not summarise
+what you recorded afterwards; reply with one short line.
+
+{thoroughness}
 
 """
 
+#: Shared by the ask and the API proxy, so the proxy's benchmark and
+#: calibration measure the prompt the agent is given. The shadow week found the
+#: 0.1.0 ask yielding about 0.4x the general extractor's claims on the same
+#: transcripts: the agent read "durable claims ... worth recording" as an
+#: invitation to curate ("keeping things modest"), took a whole span in one
+#: pass where the general extractor reads 7,500-character chunks, and wrote
+#: every schema field of every candidate, nulls included.
+_THOROUGHNESS = """\
+Extract exactly as a background extractor reading this text would: every
+individual claim the rules admit, not a summary and not a selection of
+highlights. Do not leave a claim out because it seems minor, obvious, or true
+only for now; what is kept is decided later, not here. Work through the text
+in order, a few paragraphs at a time, and finish each part before moving on.
+
+The text is about {kchars} thousand characters long. Text like this usually
+yields about {expected} claims. That is a guide to how thorough to be, never a
+quota: do not invent, split, or repeat claims to reach it, and record fewer
+when the text holds fewer.
+
+Keep each candidate compact. Always give content, subjects, confidence_value
+and evidence; leave out every other field whose value would be null or the
+default (uncertainty_nature EPISTEMIC, scope WORLD, assertion_modality
+FALSIFIABLE, polarity ASSERTED)."""
+
+
+def expected_claims(span_chars: int) -> int:
+    """How many claims a span this size usually yields: the ask's thoroughness guide."""
+    rate = get_config().extraction.agent_expected_claims_per_kchar
+    return max(1, round(rate * span_chars / 1000))
+
+
+def thoroughness(span_text: str) -> str:
+    """The completeness and compactness instructions for a span of this size."""
+    chars = len(span_text)
+    return _THOROUGHNESS.format(kchars=max(1, round(chars / 1000)), expected=expected_claims(chars))
+
 
 def agent_rules(reference_date: str) -> str:
     """The agent extractor's rules: the general extractor's, plus the evidence field.
@@ -115,7 +153,9 @@ def build_ask(ask_id: str, span_text: str, *, reference_date: str | None = None)
     date = reference_date or datetime.now(UTC).date().isoformat()
     rules = agent_rules(date).replace(_JSON_ONLY, "The `candidates` array. Each element:")
     return (
-        _SUBMIT_INSTRUCTION.format(tool=SUBMIT_TOOL, ask_id=ask_id)
+        _SUBMIT_INSTRUCTION.format(
+            tool=SUBMIT_TOOL, ask_id=ask_id, thoroughness=thoroughness(span_text)
+        )
         + rules
         + "\n\n"
         + data_fence_instruction(nonce)
@@ -328,7 +368,9 @@ class AgentExtractor:
         text = content_to_text(content)
         nonce = make_nonce()
         date = datetime.now(UTC).date().isoformat()
-        system = agent_rules(date) + "\n\n" + data_fence_instruction(nonce)
+        system = (
+            thoroughness(text) + "\n\n" + agent_rules(date) + "\n\n" + data_fence_instruction(nonce)
+        )
         try:
             raw, provider_model = await complete_with_provider_model(
                 "extraction",

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/extraction/general.py

@@ -555,8 +555,8 @@ _EXTRACT_SCHEMA_TAIL = """
 }"""
 
 
-#: The rule a delta read carries when its prompt holds a context section
-#:. In the prompt proper, never inlined in the chunk text, so the
+#: The rule a delta read carries when its prompt holds a context section.
+#: In the prompt proper, never inlined in the chunk text, so the
 #: instruction is a trusted rule rather than words inside the fenced data.
 APPEND_CONTEXT_RULE = """
 - The user message holds two fenced sections. The one labelled "context" is

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/extraction/mcp_memory.py

@@ -113,8 +113,8 @@ OBSERVATION_TAG = "memory-compat:observation"
 RELATION_TAG = "memory-compat:relation"
 #: Marks a Subject as deleted through the façade.
 TOMBSTONE_TAG = "memory-compat:deleted"
-#: Marks a particle as arriving by migration rather than by an agent write
-#:. Distinct from ``CalibrationSource.IMPORTED`` on purpose: the
+#: Marks a particle as arriving by migration rather than by an agent write.
+#: Distinct from ``CalibrationSource.IMPORTED`` on purpose: the
 #: calibration source says *how the number was produced*, this says *which run
 #: put the record here*, so an operator can find or retract one import.
 IMPORT_TAG = "migration:mcp-memory"

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/extraction/subject_scope.py

@@ -40,8 +40,8 @@ from particles.core.schema import ParticleType
 from particles.extraction.polarity import is_non_asserted
 from particles.extraction.scope import is_excluded_document_meta
 
-#: Key on the ``properties`` dict carrying the subject-scope record. Prefixed
-#:; the ``extraction:`` namespace is registered in §5.
+#: Key on the ``properties`` dict carrying the subject-scope record. Prefixed;
+#: the ``extraction:`` namespace is registered in §5.
 SUBJECT_SCOPE_KEY = "extraction:subject_scope"
 
 #: The one value that marks a claim as author-scoped. Absence of the key means

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/particles/render/markdown.py

@@ -568,8 +568,8 @@ def is_within_directory(directory: Path, path: Path) -> bool:
         return False
 
 
-#: Ownership ledger a one-file-per-subject exporter keeps at its export root
-#:: the root-relative paths of every ``.md`` file its last run wrote.
+#: Ownership ledger a one-file-per-subject exporter keeps at its export root:
+#: the root-relative paths of every ``.md`` file its last run wrote.
 EXPORT_MANIFEST_NAME = ".particles-export.json"
 
 #: Append-only journal of files a run has written that its prior manifest did

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/pyproject.toml

@@ -19,7 +19,7 @@
 # version and advance together.
 [project]
 name = "linkedparticles-core"
-version = "1.192.2"
+version = "1.192.10"
 description = "Store-free client library for Particles: define, validate, serialize, and extract provenance-tracked belief particles"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -63,14 +63,33 @@ classifiers = [
 # by a module in this distribution. It is a subset of the engine's, never a
 # superset: installing the engine installs this, so a package needed by both
 # is declared in both and resolved once.
+#
+# These ranges are the published compatibility claim, and they are allowed to
+# be wider than the development upstream's own floors, which a dependency bot
+# raises to each new release. Three rules bound them, and a test in the
+# upstream checks all three against its own pyproject and lockfile:
+#
+#   1. Every range admits the version the upstream's suite runs against.
+#   2. A next-breaking upper bound is the upstream's bound, moved when it is,
+#      and the floor beneath it sits on the release line just below it: the
+#      bound declares that line the only one vetted.
+#   3. No floor reaches back into an older major than the upstream's floor,
+#      since the code is ported to and tested on that major alone.
+#
+# Where a published advisory affects a path this layer exercises, the floor
+# is the first release that fixes it, and the advisory is named beside it.
 dependencies = [
     "pydantic>=2.7",
     "pyyaml>=6.0",
     "httpx>=0.27",
-    "anthropic>=0.28",
+    # The 1.x SDK surface: the `omit` sentinel, and sampling parameters passed
+    # through `extra_body` since 1.0 dropped the typed keyword arguments.
+    "anthropic>=1.0",
     "numpy>=1.26",
     "scipy>=1.13",
-    "sentence-transformers>=2.7",
+    # GHSA-jhr6-gm9c-rqjv (fixed in 5.6.0): the encoder model is loaded at
+    # runtime, so the fix is a floor here.
+    "sentence-transformers>=5.6.0",
     # Declared explicitly *only* so `[tool.uv.sources]` below can route it to
     # the CPU-only wheel index — uv sources apply to a project's direct
     # dependencies, and torch is otherwise transitive via
@@ -81,22 +100,25 @@ dependencies = [
     # async session; the import is executed at module load even though no
     # Client module opens a connection. Declared so a store-free install still
     # imports, and with `[asyncio]` because that import needs greenlet, which
-    # SQLAlchemy stopped installing by default in 2.1. Bounded below the next
-    # minor: 2.1 changed ORM and typing behaviour this code depends on.
-    "sqlalchemy[asyncio]>=2.0,<2.2",
+    # SQLAlchemy stopped installing by default in 2.1. Floored at 2.1 and
+    # bounded below the next minor: 2.1 changed ORM and typing behaviour this
+    # code depends on.
+    "sqlalchemy[asyncio]>=2.1,<2.2",
     # JSON-LD / RDF / SHACL parsing stack — these handle attacker-influenced
     # input (interchange import, conformance validation), so they carry tested
     # next-breaking upper bounds in addition to the floor: a future breaking
     # major of any of them must be vetted before adoption rather than pulled in
     # silently by a downstream `pip install`.
-    "pyld>=2.0,<4",
-    "pyshacl>=0.26,<0.32",
+    "pyld>=3.0,<4",
+    "pyshacl>=0.40,<0.41",
     "rdflib>=7.0,<8",
     "jsonschema>=4.22",
     # Deposited PDFs are attacker-influenced input on the runtime path, so this
-    # floor (CVE-2026-71852 / CVE-2026-71870, fixed in 6.15.0) is a direct
-    # dependency floor rather than an environment constraint.
-    "pypdf>=6.15.0",
+    # is a direct dependency floor rather than an environment constraint. 6.19.0
+    # is the first release with every published advisory fixed, the latest
+    # being GHSA-php9-fj8v-98fj, GHSA-v247-6f48-mgcj and GHSA-w23x-9jrw-r45c
+    # (CVE-2026-71852 / CVE-2026-71870 were fixed in 6.15.0).
+    "pypdf>=6.19.0",
     "html2text>=2024.2",
     "lxml-stubs>=0.5.1",
     # Observability API only: a small, pure-Python, no-op-by-default package.
@@ -113,7 +135,7 @@ dependencies = [
 # extractor imports it lazily and raises an actionable ImportError when vision
 # extraction is enabled without it, so the base install stays lean.
 vision = [
-    "pypdfium2>=4.30",
+    "pypdfium2>=5.0",
 ]
 
 # PyPI renders these beside the description and maps the recognized names
@@ -221,11 +243,20 @@ explicit = true
 [tool.uv.sources]
 torch = [{ index = "pytorch-cpu", marker = "sys_platform != 'darwin'" }]
 
-# Security floors on packages that are not direct dependencies — they arrive
-# transitively, so a floor here raises the minimum without adding a spurious
-# direct dependency. Revisit each line once the parent's own floor moves past
-# the fixed version.
+# `exclude-newer` is the seven-day supply-chain cooldown: a resolve here never
+# picks a release published in the last week, which is the window in which a
+# compromised upload is usually caught and yanked. It is a relative span, so the
+# committed `uv.lock` (cut by the export from this file) records the span and
+# not a timestamp, and `uv sync --locked` stays satisfied as the days pass. The
+# lock and this line travel together: drop the line and `uv lock --check`
+# reports the lock stale.
+#
+# The `constraint-dependencies` are security floors on packages that are not
+# direct dependencies. They arrive transitively, so a floor here raises the
+# minimum without adding a spurious direct dependency. Revisit each line once
+# the parent's own floor moves past the fixed version.
 [tool.uv]
+exclude-newer = "7 days"
 constraint-dependencies = [
     "idna>=3.15",         # CVE-2026-45409, runtime hostname path
     "setuptools>=83.0.0", # PYSEC-2026-3447, build path (pulled in by torch)

{linkedparticles_core-1.192.2 → linkedparticles_core-1.192.10}/tests/test_extraction_agent.py

@@ -94,6 +94,18 @@ class TestVerifySubmission:
         )
         assert verified.accepted[0].uncertainty_nature.value == "ALEATORY"
 
+    def test_a_compact_candidate_takes_the_defaults(self) -> None:
+        # The 0.2.0 ask tells the agent to leave out null and default fields.
+        compact = {
+            "content": "Staging runs Postgres 17.",
+            "subjects": ["staging database"],
+            "confidence_value": 0.9,
+            "evidence": "Postgres 17",
+        }
+        verified = agent.verify_submission([compact], SPAN)
+        assert len(verified.accepted) == 1 and not verified.malformed
+        assert verified.accepted[0].uncertainty_nature.value == "EPISTEMIC"
+
 
 class TestCoverage:
     cfg = AgentExtractionConfig(min_claims_per_kchar=2.0, max_evidence_drop_ratio=0.5)
@@ -203,6 +215,35 @@ class TestPrompt:
         assert '<session nonce="' in ask and "hello" in ask
         assert "{reference_date}" not in ask
 
+    def test_ask_asks_for_every_claim_and_names_the_expected_count(self) -> None:
+        span = "## User\n\n" + "x" * 9990
+        ask = agent.build_ask("ask-123", span, reference_date="2026-10-09")
+        assert "durable" not in ask.split("Rules:")[0]
+        assert "not a selection of" in ask
+        # 10,000 characters at the default 1.9 claims per 1000.
+        assert "about 10 thousand characters" in ask and "about 19 claims" in ask
+        assert "never a\nquota" in ask
+
+    def test_expected_count_follows_config(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        cfg = agent.get_config().extraction
+        monkeypatch.setattr(cfg, "agent_expected_claims_per_kchar", 3.0)
+        assert agent.expected_claims(20_000) == 60
+        assert agent.expected_claims(10) == 1
+
+    async def test_the_proxy_is_given_the_same_guide(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        seen: dict[str, str] = {}
+
+        async def fake(purpose: str, user: str, **kwargs: Any) -> tuple[str, str]:
+            seen["system"] = kwargs["system"]
+            return "[]", "anthropic:test"
+
+        monkeypatch.setattr("particles.llm.complete_with_provider_model", fake)
+        await agent.AgentExtractor().extract(
+            None,  # type: ignore[arg-type]
+            SPAN.encode(),
+        )
+        assert agent.thoroughness(SPAN) in seen["system"]
+
     def test_rules_carry_the_general_extractor_rules(self) -> None:
         rules = agent.agent_rules("2026-10-09")
         assert "One particle = one falsifiable assertion" in rules