topos-node 1.5.1__tar.gz → 1.5.2__tar.gz

{topos_node-1.5.1/topos_node.egg-info → topos_node-1.5.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: topos-node
-Version: 1.5.1
+Version: 1.5.2
 Summary: Topos personal AI engine (FastAPI): local data, sync, control plane WebSocket client
 License-Expression: Apache-2.0
 Project-URL: Homepage, https://topos.dialogues.ai

{topos_node-1.5.1 → topos_node-1.5.2}/pyproject.toml

@@ -9,7 +9,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "topos-node"
-version = "1.5.1"
+version = "1.5.2"
 description = "Topos personal AI engine (FastAPI): local data, sync, control plane WebSocket client"
 readme = "README.md"
 license = "Apache-2.0"

{topos_node-1.5.1 → topos_node-1.5.2}/topos/__version__.py

@@ -1,6 +1,6 @@
 """Topos Control Plane Version Information."""
 
-__version__ = "1.5.1"
-__version_info__ = (1, 5, 1)
+__version__ = "1.5.2"
+__version_info__ = (1, 5, 2)
 
 __all__ = ["__version__", "__version_info__"]

{topos_node-1.5.1 → topos_node-1.5.2}/topos/permissions_v2/answer_checks.py

@@ -39,6 +39,18 @@ FORM_WORDS = frozenset({"owner", "owners", "herself", "himself", "theirs", "them
     "cite", "evidence", "support", "supporting", "supported", "describe", "described", "mention", "mentioned",
     "discuss", "discussed", "written", "question", "questions",
     "lately", "recent", "recently"})
+# A2A-4 amendment 5 (AR152 class I, 1.5.2): the catalog templates' own instruction words, a second closed list read
+# exactly like FORM_WORDS by the anchor rule (`_SCAFFOLD_FOLDS`) and the subject check
+# (answer_generation._GENERIC_QUESTION_TERMS), after the fold. They say how to answer ("Separate stated facts from
+# inference", "Only report what the messages state", "Do not connect unrelated messages or infer chronology", "Name
+# topics only, never pages", "going by their browsing"), never what an item is about; a supported sentence for every
+# fact, goal, relationship and interest template, and the goals, home and family questions, was dropped for want of
+# one. It only stops requiring words no item can carry: names still bind, an unknown word still echoes, every other
+# subject word still binds. Not in it, by WS0's ruling: "relationships", "people", "spend", "goal", "topic"
+# (singular). Pinned by an exact-set test; adding a word is an amendment.
+INSTRUCTION_WORDS = frozenset({"stated", "state", "states", "report", "guess", "inference", "infer", "separate",
+    "statement", "statements", "explicit", "explicitly", "connect", "unrelated", "chronology", "establish",
+    "establishes", "around", "pursue", "going", "browsing", "browse", "topics", "never", "pages", "page"})
 
 
 def _redact_phone(match: re.Match[str]) -> str:
@@ -204,7 +216,7 @@ def _stem(word: str) -> str:
 
 
 # A word is scaffold when its fold is the fold of a scaffold word ("explained" is "explain", "happening" "happen").
-_SCAFFOLD_FOLDS = frozenset(_stem(word) for word in _QUESTION_SCAFFOLD | FORM_WORDS)
+_SCAFFOLD_FOLDS = frozenset(_stem(word) for word in _QUESTION_SCAFFOLD | FORM_WORDS | INSTRUCTION_WORDS)
 
 
 def question_anchors(question: str) -> frozenset[str]:

{topos_node-1.5.1 → topos_node-1.5.2}/topos/permissions_v2/answer_generation.py

@@ -12,7 +12,7 @@ import sqlite3
 import unicodedata
 from dataclasses import dataclass
 
-from .answer_checks import (_CITATION, _TOKEN, FORM_WORDS, TEMPLATE_VERSION, citation_numbers, copied_sentence,
+from .answer_checks import (_CITATION, _TOKEN, FORM_WORDS, INSTRUCTION_WORDS, TEMPLATE_VERSION, citation_numbers, copied_sentence,
     echoes_protected_word, echoes_question_only_word, post_check_citations, protected_question_words,
     question_anchors, question_only_anchors, scrub_sentences, split_sentences, _stem, _tokens)
 from .answer_protocol import AnswerOnly, AnswerWithSources, NoAnswer, VERSION
@@ -329,7 +329,7 @@ _GENERIC_QUESTION_TERMS = frozenset({"about", "after", "again", "before", "could
     "short", "permitted", "material",
     # Every item this check reads was released by the share, so "shared" in a question to a share names the act of
     # sharing, not a subject an item must carry ("What plans were shared?" asks about plans).
-    "share", "shared", "shares", "sharing"}) | FORM_WORDS
+    "share", "shared", "shares", "sharing"}) | FORM_WORDS | INSTRUCTION_WORDS   # amendment 5: instruction words
 
 
 # A word is a question word when its fold is the fold of a listed word ("plans" is "plan", "updating" is "updat").

{topos_node-1.5.1 → topos_node-1.5.2}/topos/permissions_v2/answer_release.py

@@ -29,6 +29,27 @@ MAX_WAIT_SECONDS = 60
 MAX_END_SECONDS = 110
 UNFETCHED_SECONDS = 600
 
+# The closed vocabulary of the private receipt's `reason` (A2A-4 §7.1, `_receipt`): counts-only words, never
+# content. The first group is the contract's list with amendment 2's `question_not_supported`. The second (BL-156,
+# 1.5.2) is what the retrieval inside a job (`retrieve_for_answer`, its index load and its boundary) refuses with:
+# carried as the refusal's own code, so a dark share reads as the index it lacks and not as a move of authority.
+# `authority_moved` is kept for the one case it names: the permitted set changed between retrieval and the body.
+# A refusal with any other code is recorded as `refused`: the receipt never carries a word outside this set.
+RECEIPT_REASONS = frozenset({
+    "answered", "nothing_matched", "question_protected", "question_not_supported", "all_sentences_dropped",
+    "answer_protected", "authority_moved", "body_invalid", "model_unavailable", "model_error", "queue_deadline",
+    "deadline",
+    "search_index_missing", "search_index_stale", "search_index_over_cap", "search_index_integrity",
+    "search_index_unavailable", "search_index_binding", "search_verification_closed", "review_database_binding",
+    "grant_inactive", "policy_time", "authority_stale", "entity_protection_lineage_unavailable", "refused",
+})
+REASON_OTHER = "refused"
+
+
+def receipt_reason(code: str) -> str:
+    """The receipt word for a refusal's code: the code itself when it is in the closed vocabulary, else `refused`."""
+    return code if code in RECEIPT_REASONS else REASON_OTHER
+
 
 def answer_jobs_active() -> bool:
     """A job is queued or running: the background assessments wait (A2A-4 Q4, `answer_gate`)."""
@@ -291,8 +312,10 @@ class AnswerService:
             if digest(still.model_dump()) != digest(output.model_dump()):
                 return None, "authority_moved"
             return checked, checked.reason
-        except PolicyError:
-            return None, "authority_moved"
+        except PolicyError as exc:
+            # BL-156: the refusal's own code (the index missing, stale or not ready; the grant inactive; the
+            # authority stale), never `authority_moved` for all of them. `receipt_reason` keeps it counts-only.
+            return None, receipt_reason(exc.code)
         except Exception:
             return None, "model_error"
 

{topos_node-1.5.1 → topos_node-1.5.2}/topos/permissions_v2/entity_boundary.py

@@ -65,6 +65,19 @@ from .english_short_words import WORDS_2_3, WORDS_4, WORDS_ENDING_S_3_4
 # part's possessive with no apostrophe opening a sentence is a genitive, and a name part in a script written without
 # spaces is a part from two characters and found anywhere in a run. Matching only widens.
 VERSION = "node-observed-entity-boundary/v9"
+# The context REVISION a share index seals for each member (`check`; search_index.py `entity_dependencies` and
+# `entity_context_revision`) is content-based (BL-155, 1.5.2): it leaves out the columns the message store
+# rewrites on every sync batch that touches a conversation (conversations_tables.py `_upsert_conversation_row`
+# and the participants upsert set `updated_at = datetime('now')`), so a batch that adds nothing a share decides
+# on no longer stales every index holding a message of that conversation. These columns record when a row was
+# touched, never who is in it or what was said. The VETO (`matched`) still reads every column of every row on
+# every check, so a protected name or id written anywhere, these columns included, withholds at once whatever
+# the revision says. Every index sealed before this contract fails its dependency check once and rebuilds once.
+CONTEXT_REVISION_CONTRACT = "context-bookkeeping-free/v1"
+CONTEXT_BOOKKEEPING = {
+    "conversations": frozenset({"created_at", "updated_at"}),
+    "conversation_participants": frozenset({"created_at", "updated_at"}),
+}
 # Evidence leaves with no conversational context (evidence_families, IF-5).
 CONTEXTLESS_TABLES = frozenset({"journal_entries"})
 # The family whose rows first matched each part of a protected name (v3, module docstring). Since v8 every row and
@@ -1022,6 +1035,21 @@ def rows_revision(groups):
         for index, group in enumerate(groups))})
 
 
+def context_bookkeeping(table: str) -> frozenset:
+    """The columns of `table` the context revision leaves out (CONTEXT_REVISION_CONTRACT): the store's
+    bookkeeping timestamps on a conversation and its roster, and for the other context tables the columns the
+    owner's review surface already leaves out (`evidence.REVIEW_SURFACE_EXCLUSIONS`: batch ids, ingest and row
+    timestamps, the disclosure sweep's columns). The veto never uses this: it reads every column."""
+    from .evidence import REVIEW_SURFACE_EXCLUSIONS
+    return CONTEXT_BOOKKEEPING.get(table, frozenset()) | REVIEW_SURFACE_EXCLUSIONS.get(table, frozenset())
+
+
+def _revision_rows(table: str, rows: list) -> list:
+    """`rows` without `table`'s bookkeeping columns, for the context revision only."""
+    excluded = context_bookkeeping(table)
+    return [{key: value for key, value in row.items() if key not in excluded} for row in rows]
+
+
 def _spelled(values) -> list:
     """Every spelling of each name (name_spellings), in order."""
     return [spelling for value in values for spelling in name_spellings(value)]
@@ -1455,6 +1483,7 @@ class EntityBoundary:
         if len(parent) != 1 or len(siblings) > MAX_CONTEXT_ROWS:
             raise PolicyError(UNAVAILABLE)
         rows = [*parent, *roster, *siblings]
+        # The veto reads every column of every row (the bookkeeping columns included).
         matched = any(item.get("contact_id") in self.contacts or item.get("sender_id") in self.contacts
                       or item.get("sender_id") in self.ids or self._hits(item, True, False) for item in rows)
         # Parent/ancestor record protection is a direct veto even without names.
@@ -1462,7 +1491,11 @@ class EntityBoundary:
         matched |= self._linked(conversation, parent_table, source_id)
         matched |= self.conn.execute("SELECT 1 FROM owner_only_records WHERE canonical_table=? AND record_id=?",
                                      (parent_table, conversation)).fetchone() is not None
-        revision = rows_revision([rows])
+        # The revision is content-based (CONTEXT_REVISION_CONTRACT): the parent row and the roster without the
+        # store's bookkeeping columns, and the distinct senders. A sync batch that only re-touches the
+        # conversation leaves it unchanged; a new participant, a new sender or a changed column moves it.
+        revision = digest({"contract": CONTEXT_REVISION_CONTRACT, "rows": rows_revision([[
+            *_revision_rows(parent_table, parent), *_revision_rows("conversation_participants", roster), *siblings]])})
         self._context_cache[key] = (matched, revision)
         return matched, revision
 
@@ -1515,8 +1548,11 @@ class EntityBoundary:
                 context_matched, context_revision = self._context(table, row, source_id, dataset_id)
                 matched |= context_matched
                 replies = self._reply_context(table, row, source_id, dataset_id)
-                matched |= any(self._hits(parent, True, False) for parent in replies)
-                context_revision = digest({"context": context_revision, "replies": rows_revision([replies])})
+                matched |= any(self._hits(parent, True, False) for parent in replies)   # every column
+                # A reply ancestor's revision is its reviewed surface (`_revision_rows`: the same exclusions
+                # `evidence._row_revision` applies to the member itself), not its ingest bookkeeping.
+                context_revision = digest({"context": context_revision,
+                                           "replies": rows_revision([_revision_rows(table, replies)])})
             return matched, digest({"boundary": self.revision, "context": context_revision})
         except (sqlite3.Error, TypeError, ValueError, RecursionError):
             raise PolicyError(UNAVAILABLE) from None

{topos_node-1.5.1 → topos_node-1.5.2}/topos/upgrades/manifests.json

@@ -976,6 +976,14 @@
         "No schema change: the database stays at 81. The pre-1.5.0 backup stays the way back to 1.4.4."
       ]
     },
+    {
+      "version": "1.5.2",
+      "summary": "Release 1.5.2: a share index survives sync bookkeeping (BL-155), receipts name the real refusal (BL-156), template instruction words are request words (A2A-4 amendment 5).",
+      "steps": [],
+      "notes": [
+        "No schema change. The share index context digest moves to context-bookkeeping-free/v1, so every existing share index rebuilds once at the first sweep after start."
+      ]
+    },
     {
       "version": "unreleased",
       "summary": "Staging entry for in-flight PRs. cut_release.py stamps this with the next version at release cut. Never executed by steps_between.",

{topos_node-1.5.1 → topos_node-1.5.2/topos_node.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: topos-node
-Version: 1.5.1
+Version: 1.5.2
 Summary: Topos personal AI engine (FastAPI): local data, sync, control plane WebSocket client
 License-Expression: Apache-2.0
 Project-URL: Homepage, https://topos.dialogues.ai