apsimo 1.3.1__tar.gz → 1.3.2__tar.gz

{apsimo-1.3.1 → apsimo-1.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: apsimo
-Version: 1.3.1
+Version: 1.3.2
 Summary: Persistent identity, memory and autonomy for AI agents
 License-Expression: MIT
 Requires-Python: >=3.11
@@ -28,7 +28,7 @@ Requires-Dist: pandas>=2.0; extra == "vectors"
 Requires-Dist: pyarrow>=15.0; extra == "vectors"
 Requires-Dist: lancedb>=0.10; extra == "vectors"
 Provides-Extra: hermes
-Requires-Dist: apsimo-hermes<2,>=1.3.1; extra == "hermes"
+Requires-Dist: apsimo-hermes<2,>=1.3.2; extra == "hermes"
 Provides-Extra: mcp
 Requires-Dist: mcp[cli]<2,>=1.0; extra == "mcp"
 Provides-Extra: lancedb

{apsimo-1.3.1 → apsimo-1.3.2}/apsimo/api/routers/host.py

@@ -3109,10 +3109,14 @@ async def context_assemble(
                 _per = _comms_log.last_per_channel(contact_id)
                 if _per:
                     _chs = ", ".join(f"{ch} {str(v['ts'])[:10]}" for ch, v in _per.items())
-                    _bits.append(f"Channels used: {_chs}.")
+                    _bits.append(f"Recorded channels: {_chs}.")
                 _lo = _comms_log.last_outbound(contact_id)
                 if _lo:
-                    _bits.append(f"I last reached out via {_lo['channel']} on {str(_lo['ts'])[:10]}.")
+                    _bits.append(
+                        f"Last recorded outgoing message: {_lo['channel']} on {str(_lo['ts'])[:10]}. "
+                        "This summary includes conversation replies; it does not establish "
+                        "proactive outreach or delivery."
+                    )
                 if _commitment_store is not None:
                     try:
                         _cm = _commitment_store.list(person_id=contact_id, status=["pending", "overdue"], limit=5)

{apsimo-1.3.1 → apsimo-1.3.2}/apsimo/intelligence/graph/selection.py

@@ -70,6 +70,21 @@ class RecallSelector:
                 seen.add(key)
             unique.append(row)
         quotations = unique
+        def repeats_request(row):
+            # An exact repeat contributes no independent answer evidence. Keep
+            # it available as supplementary context, without classifying other
+            # questions, promoting speakers or changing its source bytes.
+            return bool(isinstance(query, str) and query.strip()
+                and row.get("kind") == "source_quote"
+                and row.get("epistemic_state") == "quotation"
+                and row.get("scope") == "person" and row.get("contact_id")
+                and row.get("role") in {"user", "assistant"}
+                and isinstance(row.get("content"), str)
+                and row["content"].strip() == query.strip()
+                and not any(row.get(key) for key in (
+                    "atomic_evidence", "validity_status", "procedure_context",
+                    "source_context", "_annotation_ids")))
+
         # Media is an independent candidate producer. Appending its first hit
         # after every text hit can exclude it from the bounded reranker before
         # relevance is assessed. Fuse its rank independently; it still shares
@@ -82,10 +97,17 @@ class RecallSelector:
         candidates = fuse_candidates(
             beliefs, text, limit=len(beliefs) + len(quotations),
             confidence_weighting=False, additional=(media,))
+        # Do this before bounded reranking, so repeated requests do not crowd
+        # out answer-bearing candidates before their relevance can be judged.
+        candidates.sort(key=repeats_request)
         ranked = await self.rerank(
             query, candidates, limit, confidence_weighting=False,
             candidate_limit=max(1, 4 * limit))
         ranked.sort(key=lambda row: row.get("relevance", 0), reverse=True)
+        # A semantic reranker can score the identical question highest. Preserve
+        # its abstention decisions and substantive order, then place only exact
+        # unqualified request repeats after the remaining evidence.
+        ranked.sort(key=repeats_request)
         return pack_memory_context(ranked, limit=limit, max_chars=max_chars)
 
     async def rerank(

{apsimo-1.3.1 → apsimo-1.3.2}/apsimo.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: apsimo
-Version: 1.3.1
+Version: 1.3.2
 Summary: Persistent identity, memory and autonomy for AI agents
 License-Expression: MIT
 Requires-Python: >=3.11
@@ -28,7 +28,7 @@ Requires-Dist: pandas>=2.0; extra == "vectors"
 Requires-Dist: pyarrow>=15.0; extra == "vectors"
 Requires-Dist: lancedb>=0.10; extra == "vectors"
 Provides-Extra: hermes
-Requires-Dist: apsimo-hermes<2,>=1.3.1; extra == "hermes"
+Requires-Dist: apsimo-hermes<2,>=1.3.2; extra == "hermes"
 Provides-Extra: mcp
 Requires-Dist: mcp[cli]<2,>=1.0; extra == "mcp"
 Provides-Extra: lancedb

{apsimo-1.3.1 → apsimo-1.3.2}/apsimo.egg-info/SOURCES.txt

@@ -833,6 +833,7 @@ tests/test_reachout_policy.py
 tests/test_recall_hybrid.py
 tests/test_recall_person_boost.py
 tests/test_recall_ranking.py
+tests/test_recall_request_repetition.py
 tests/test_recall_rerank.py
 tests/test_recall_source_presentation.py
 tests/test_recall_unified_context.py

{apsimo-1.3.1 → apsimo-1.3.2}/apsimo.egg-info/requires.txt

@@ -35,7 +35,7 @@ beautifulsoup4>=4.12
 neo4j>=5.18
 
 [hermes]
-apsimo-hermes<2,>=1.3.1
+apsimo-hermes<2,>=1.3.2
 
 [lancedb]
 lancedb>=0.6

{apsimo-1.3.1 → apsimo-1.3.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "apsimo"
-version = "1.3.1"
+version = "1.3.2"
 description = "Persistent identity, memory and autonomy for AI agents"
 readme = "README.md"
 license = "MIT"
@@ -29,7 +29,7 @@ dependencies = [
 graph = ["neo4j>=5.18"]
 neo4j = ["neo4j>=5.18"]
 vectors = ["sentence-transformers>=3.0", "pandas>=2.0", "pyarrow>=15.0", "lancedb>=0.10"]
-hermes = ["apsimo-hermes>=1.3.1,<2"]
+hermes = ["apsimo-hermes>=1.3.2,<2"]
 mcp = ["mcp[cli]>=1.0,<2"]
 lancedb = ["lancedb>=0.6", "pyarrow>=14.0", "pandas>=2.0"]
 postgres = ["asyncpg>=0.29"]

{apsimo-1.3.1 → apsimo-1.3.2}/tests/test_hermes_turn_outbox.py

@@ -1197,7 +1197,8 @@ def test_colony_client_uses_one_absolute_deadline_across_http_phases():
     assert elapsed < 0.20
 
 
-def test_colony_client_absolute_deadline_transport_accepts_exact_put():
+@pytest.mark.parametrize('channel_id', ['', 'email:thread-739'])
+def test_colony_client_absolute_deadline_transport_accepts_exact_put(channel_id):
     module = _load_client("colony_hermes_client_deadline_success_test")
     listener = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
     listener.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
@@ -1242,6 +1243,7 @@ def test_colony_client_absolute_deadline_transport_accepts_exact_put():
             session_id="session-1",
             contact_id="cid-owner",
             turn_id="turn/one",
+            channel_id=channel_id,
             timeout_seconds=0.20,
         ) is True
     finally:
@@ -1260,6 +1262,7 @@ def test_colony_client_absolute_deadline_transport_accepts_exact_put():
             "contact_id": "cid-owner",
             "session_id": "session-1",
             "turn_id": "turn/one",
+            **({"channel_id": channel_id} if channel_id else {}),
         },
         "identity": {"host_id": "hermes"},
     }

apsimo-1.3.2/tests/test_recall_request_repetition.py

@@ -0,0 +1,140 @@
+"""A previous copy of the request must not crowd out independent evidence."""
+from copy import deepcopy
+from datetime import datetime
+import json
+
+import pytest
+
+from apsimo.intelligence.graph.recall import source_candidates
+from apsimo.intelligence.graph.selection import RecallSelector
+from apsimo.beliefs.source_projection import SourceClaimProjection
+from apsimo.beliefs.source_time import interpret_time_query
+from apsimo.turns.idempotency import TurnIdempotencyLedger, source_message_hash
+from test_source_claim_projection import Model, claim
+
+
+def quotation(identifier, text, role="user", **extra):
+    return dict(source_candidates([{
+        "turn_id": identifier, "role": role, "content": text,
+        "source_message_hash": identifier + "-version",
+        "scope": "person", "contact_id": "person-a", "session_id": "earlier",
+    }])[0], **extra)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("mode", ["off", "shadow", "on", "failed"])
+async def test_exact_request_is_supplementary_across_rerank_paths(monkeypatch, mode):
+    query = "Where is the spare connector?"
+    echo = quotation("question", query)
+    correction = quotation("correction", "The spare connector is now in drawer six.")
+    observation = quotation("observation", "The connector inspection found bent pins.", "assistant")
+    rows = [echo, observation, correction]
+    original = deepcopy(rows)
+    monkeypatch.setenv("COLONY_RECALL_RERANK", "on" if mode == "failed" else mode)
+    monkeypatch.delenv("COLONY_RECALL_RERANK_MIN_SCORE", raising=False)
+
+    async def rank(query, documents, top_k):
+        assert documents == [observation["content"], correction["content"], echo["content"]]
+        if mode == "failed":
+            raise RuntimeError("controlled unavailable reranker")
+        # Exact lexical/semantic overlap is not independent answer evidence.
+        return [{"index": i, "score": 100 if text == query else 2 - i}
+                for i, text in enumerate(documents)]
+
+    selected, context = await RecallSelector(rank).select_context(query, [], rows, limit=2)
+    assert [row["source_turn_id"] for row in selected] == ["observation", "correction"]
+    assert rows == original
+    assert '"role": "assistant"' in context and '"role": "user"' in context
+    assert len(context) <= 6000
+    # The source is still available when the complete packet fits.
+    selected, _ = await RecallSelector().select_context(query, [], rows, limit=3)
+    assert selected[-1]["source_turn_id"] == "question"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("extra", [
+    {"atomic_evidence": True}, {"validity_status": "temporal_history"},
+    {"procedure_context": "complete"}, {"source_context": "complete"},
+    {"_annotation_ids": ["annotation"]}, {"scope": "session"},
+    {"contact_id": None}, {"epistemic_state": "derived_unverified"},
+])
+async def test_qualified_or_uncertain_source_is_not_deprioritized(monkeypatch, extra):
+    monkeypatch.setenv("COLONY_RECALL_RERANK", "off")
+    query = "Where was the connector yesterday?"
+    qualified = quotation("qualified", query, **extra)
+    rows = [qualified, quotation("other", "An unrelated observation.", "assistant")]
+    selected, _ = await RecallSelector().select_context(query, [], rows, limit=1)
+    assert selected[0]["source_turn_id"] == "qualified"
+
+
+@pytest.mark.asyncio
+async def test_history_and_assistant_observations_keep_exact_bytes(monkeypatch):
+    monkeypatch.setenv("COLONY_RECALL_RERANK", "off")
+    query = "What did we ask and observe about the connector yesterday?"
+    rows = [quotation("question", "Where is the spare connector?"),
+            quotation("observation", "I measured its resistance as 4 ohms.", "assistant"),
+            quotation("tool", "Continuity test: PASS", "tool")]
+    selected, _ = await RecallSelector().select_context(query, [], rows, limit=3)
+    assert [(r["source_turn_id"], r["content"], r["role"]) for r in selected] == [
+        (r["source_turn_id"], r["content"], r["role"]) for r in rows]
+
+
+@pytest.mark.asyncio
+async def test_no_semantic_or_case_normalization(monkeypatch):
+    monkeypatch.setenv("COLONY_RECALL_RERANK", "off")
+    rows = [quotation("case-differs", "Where is the Connector?"),
+            quotation("observation", "The connector is in drawer six.")]
+    selected, _ = await RecallSelector().select_context("Where is the connector?", [], rows, limit=1)
+    assert selected[0]["source_turn_id"] == "case-differs"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("admitted", [False, True])
+async def test_pending_and_admitted_correction_survive_repeated_recall(tmp_path, monkeypatch, admitted):
+    """Reconstructed ingestion/FTS/projection/packing case, not a model benchmark."""
+    monkeypatch.setenv("COLONY_RECALL_RERANK", "off")
+    query = "Which drawer currently holds the spare connector? Use retained evidence and answer briefly."
+    old = "Remember this workshop fact: the spare connector is in drawer four."
+    correction = "Correction: the spare connector is now in drawer six. The previous drawer is outdated. Remember the correction."
+    ledger = TurnIdempotencyLedger(tmp_path / "source.db")
+    projection = SourceClaimProjection(ledger)
+    model = Model({
+        old: claim("the spare connector is in drawer four.", "drawer four", subject="spare connector", predicate="location"),
+        correction: claim(correction, "drawer six", subject="spare connector", predicate="location", operation="correct", match_prior=True),
+    })
+
+    def record(identifier, user, assistant, *, derive=False):
+        ledger.record_source(identifier, contact_id="person-a", session_id="earlier",
+            messages=[{"role": "user", "content": user}, {"role": "assistant", "content": assistant}],
+            occurred_at="2026-03-01T12:00:00+00:00", derive_claims=derive)
+
+    record("formation", old, "", derive=True)
+    assert await projection.process_one(model)
+    record("question-old", query.replace("currently ", ""),
+        "Drawer four. Per retained memory, the spare connector is in drawer four.")
+    record("correction", correction, "", derive=True)
+    record("question-new", query, "Drawer six. Per corrected memory, the spare connector is in drawer six.")
+    record("question-repeat", query, "")
+    if admitted:
+        assert await projection.process_one(model)
+    scope = {"contact_id": "person-a", "session_id": "fresh"}
+    hits = ledger.search_sources(query, **scope, limit=10)
+    beliefs, quotes = projection.prepare_context([], hits, **scope,
+        time_query=interpret_time_query(query, now=datetime.fromisoformat("2026-03-02T12:00:00+00:00")))
+    selected, context = await RecallSelector().select_context(query, beliefs, quotes, limit=5, max_chars=6000)
+    # Test evidence membership and exact attribution, not just the new value
+    # appearing in a prior assistant's answer or a source-ID metadata list.
+    if admitted:
+        assertions = [a for r in selected if r.get("content_format") == "source_assertions_v1"
+                      for a in json.loads(r["content"])["assertions"]]
+        assert len(assertions) == 1
+        assert assertions[0]["source"] == "turn:correction"
+        assert assertions[0]["role"] == "user" and assertions[0]["quote"] == correction
+    else:
+        direct = [r for r in selected if r.get("source_turn_id") == "correction" and r.get("role") == "user"]
+        assert len(direct) == 1
+        assert direct[0]["content"] == correction
+        assert direct[0]["source_message_hash"] == source_message_hash("earlier", {"role": "user", "content": correction})
+        assert any(r.get("source_turn_id") == "formation" for r in selected)
+    assert any(r.get("role") == "assistant" for r in selected)
+    assert len(selected) <= 5 and len(context) <= 6000

{apsimo-1.3.1 → apsimo-1.3.2}/tests/test_source_vectors.py

@@ -102,17 +102,20 @@ async def test_repeated_semantic_questions_leave_room_for_evidence_and_survive_e
     monkeypatch.setenv('COLONY_RECALL_RERANK', 'on')
     monkeypatch.delenv('COLONY_RECALL_RERANK_MIN_SCORE', raising=False)
     selector = RecallSelector(rerank)
-    async def recall():
+    async def recall(limit=5):
         lexical = ledger.search_sources(query, contact_id='c', session_id='later', limit=10)
         semantic, _ = await projection.search(query, contact_id='c', session_id='later', limit=20)
         hits = merge_source_hits(lexical, semantic)
         _, rows = claims.prepare_context([], hits, contact_id='c', session_id='later', time_query=MemoryTimeQuery())
-        return rows, await selector.select_context(query, [], rows)
+        return rows, await selector.select_context(query, [], rows, limit=limit)
     rows, (selected, context) = await recall()
     duplicates = [row for row in rows if row['content'] == query]
     assert len(duplicates) == 5  # Fusion retains all occurrences for claim/time expansion.
     assert calls[0].count(query) == 1
-    assert sum(row['content'] in useful for row in selected) == 4
+    assert {row['content'] for row in selected} == set(useful)
+    # Exact repeated requests are supplementary. A larger packet still exposes
+    # their intact lineage, and erasure must advance to a surviving occurrence.
+    _, (selected, _) = await recall(limit=6)
     retained = next(row for row in selected if row['content'] == query)
     for key in ('id', 'source_uri', 'source_turn_id', 'source_message_hash', 'occurred_at', 'ingested_at'):
         assert retained.get(key) == duplicates[0].get(key)
@@ -121,6 +124,8 @@ async def test_repeated_semantic_questions_leave_room_for_evidence_and_survive_e
     rows, (selected, context) = await recall()
     assert erased not in {row['source_turn_id'] for row in rows + selected}
     assert calls[-1].count(query) == 1
+    assert {row['content'] for row in selected} == set(useful)
+    _, (selected, _) = await recall(limit=6)
     replacement = next(row for row in selected if row['content'] == query)
     assert replacement['source_turn_id'] != erased
     assert replacement['source_message_hash'] != retained['source_message_hash']

{apsimo-1.3.1 → apsimo-1.3.2}/tests/test_tom_p8_server_integration.py

@@ -680,6 +680,8 @@ async def test_p8_scoped_non_owner_queries_only_exact_person_commitments(
 
 @pytest.mark.asyncio
 async def test_p8_guest_comms_uses_neutral_owner_label(tmp_path, monkeypatch):
+    from colony_sidecar.contacts.comms import CommsLog
+
     monkeypatch.setenv("COLONY_RECIPIENT_SIMULATOR_MODE", "shadow")
     monkeypatch.setenv("COLONY_OWNER_CONTACT_ID", "owner")
     monkeypatch.setenv("COLONY_OWNER_NAME", "PRIVATE OWNER NAME")
@@ -689,6 +691,9 @@ async def test_p8_guest_comms_uses_neutral_owner_label(tmp_path, monkeypatch):
     _attach_p8_runtime(state_dir=tmp_path, facts_store=facts)
     spies = _PersonalContextSpies()
     spies.wire()
+    comms_log = CommsLog(str(tmp_path / "communications.db"))
+    comms_log.log("alice", channel="voice:thread", direction="out", summary="A routine reply.")
+    monkeypatch.setattr(host, "_comms_log", comms_log)
 
     response = await host.context_assemble(
         _context("alice"), request=_request(_authority("alice")))
@@ -698,6 +703,9 @@ async def test_p8_guest_comms_uses_neutral_owner_label(tmp_path, monkeypatch):
     )
     assert "PRIVATE OWNER NAME" not in comms.body
     assert "owner approval" in comms.body.lower()
+    assert "Last recorded outgoing message:" in comms.body
+    assert "does not establish proactive outreach or delivery" in comms.body
+    assert "I last reached out" not in comms.body
 
 
 @pytest.mark.asyncio