@pentatonic-ai/ai-agent-sdk 0.10.10 → 0.10.12

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.cjs CHANGED
@@ -878,7 +878,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
878
878
  }
879
879
 
880
880
  // src/telemetry.js
881
- var VERSION = "0.10.10";
881
+ var VERSION = "0.10.12";
882
882
  var TELEMETRY_URL = "https://sdk-telemetry.philip-134.workers.dev";
883
883
  function machineId() {
884
884
  const raw = typeof process !== "undefined" ? `${process.env?.USER || process.env?.USERNAME || "u"}:${process.platform || "x"}:${process.arch || "x"}` : "browser";
package/dist/index.js CHANGED
@@ -847,7 +847,7 @@ function fireAndForgetEmit(clientConfig, sessionOpts, messages, result, model) {
847
847
  }
848
848
 
849
849
  // src/telemetry.js
850
- var VERSION = "0.10.10";
850
+ var VERSION = "0.10.12";
851
851
  var TELEMETRY_URL = "https://sdk-telemetry.philip-134.workers.dev";
852
852
  function machineId() {
853
853
  const raw = typeof process !== "undefined" ? `${process.env?.USER || process.env?.USERNAME || "u"}:${process.platform || "x"}:${process.arch || "x"}` : "browser";
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@pentatonic-ai/ai-agent-sdk",
3
- "version": "0.10.10",
3
+ "version": "0.10.12",
4
4
  "description": "TES SDK — LLM observability and lifecycle tracking via Pentatonic Thing Event System. Track token usage, tool calls, and conversations. Manage things through event-sourced lifecycle stages with AI enrichment and vector search.",
5
5
  "type": "module",
6
6
  "main": "./dist/index.cjs",
@@ -56,12 +56,29 @@ def _distiller_post_fn(endpoint: str, model: str):
56
56
  """Build a post_fn(messages)->str hitting the distiller's OpenAI
57
57
  /v1/chat/completions (temperature 0, thinking off — same shape the worker
58
58
  uses). In-VPC: memory content never leaves the network. Raises on failure
59
- so the adjudicator fails closed (treats it as 'unsure', never merges)."""
59
+ so the adjudicator fails closed (treats it as 'unsure', never merges).
60
+
61
+ Granular timeout: a short connect deadline so a scaled-to-zero distiller
62
+ fails in seconds, not 60s, but a generous read deadline so a live-but-busy
63
+ model still gets time to generate. Plus a circuit breaker — once a connect
64
+ fails, every later call short-circuits immediately, so a down distiller
65
+ costs one timeout for the whole run instead of one per candidate pair (the
66
+ autoscaler scales the distiller on the distill queue, not on fusion need,
67
+ so it is routinely down at sweep time)."""
68
+ timeout = httpx.Timeout(connect=5.0, read=60.0, write=10.0, pool=5.0)
69
+ state = {"down": False}
70
+
60
71
  def post(messages):
61
- r = httpx.post(endpoint, timeout=60, json={
62
- "model": model, "messages": messages, "temperature": 0.0,
63
- "max_tokens": 120, "chat_template_kwargs": {"enable_thinking": False},
64
- })
72
+ if state["down"]:
73
+ raise RuntimeError("distiller circuit open (earlier connect failed)")
74
+ try:
75
+ r = httpx.post(endpoint, timeout=timeout, json={
76
+ "model": model, "messages": messages, "temperature": 0.0,
77
+ "max_tokens": 120, "chat_template_kwargs": {"enable_thinking": False},
78
+ })
79
+ except (httpx.ConnectError, httpx.ConnectTimeout, httpx.ReadTimeout) as e:
80
+ state["down"] = True
81
+ raise RuntimeError(f"distiller unreachable: {e}") from e
65
82
  r.raise_for_status()
66
83
  return r.json()["choices"][0]["message"]["content"]
67
84
  return post
@@ -72,8 +89,23 @@ def _norm(s: str) -> str:
72
89
 
73
90
 
74
91
  def _entity_dup_sets(cur, arena: str) -> list[list[dict]]:
75
- """Group same-(type) entities that are exact normalized-name dupes OR
76
- share a provenance event with a junk-leaning twin. Returns groups of >=2."""
92
+ """Tier 1, DETERMINISTIC auto-merge: same-(type) entities whose names are
93
+ exact normalized-name dupes (case/whitespace variants like 'PR'/'pr',
94
+ 'USDC'/'usdc', or two rows of the literally identical string). Returns
95
+ groups of >=2.
96
+
97
+ Co-occurrence (shared-provenance) is DELIBERATELY NOT here. A single event
98
+ routinely mentions a real entity AND a digit-heavy real entity that are
99
+ different things ("running Linux on the RTX 6000" — Linux and RTX 6000
100
+ share that event but are not the same node). looks_like_id flags any
101
+ digit-ratio>0.5 name as junk, which catches GPU models (RTX 6000, B2000),
102
+ protocol names (x402), and standards (EN18031) — all real. The legitimate
103
+ cross-run target ('1716801984' == 'Katie Cooper' across runs) is
104
+ INDISTINGUISHABLE from that false positive using only (type, shared-event,
105
+ digit-ratio), so co-occurrence merges CANNOT be made deterministically.
106
+ They are routed to the LLM-adjudicated tier (_cooccurrence_candidates) and
107
+ simply do not happen when the distiller is unreachable. (Was a Tier-1
108
+ auto-merge; the first prod dry-run merged RTX 6000→Linux, x402→USDC etc.)"""
77
109
  cur.execute(
78
110
  """SELECT id, entity_type, canonical_name, aliases, provenance_event_ids, disclosure_class
79
111
  FROM entities WHERE arena = %s AND disclosure_class <> 'restricted'""",
@@ -81,45 +113,10 @@ def _entity_dup_sets(cur, arena: str) -> list[list[dict]]:
81
113
  )
82
114
  ents = cur.fetchall()
83
115
  groups: dict[tuple, list[dict]] = {}
84
- # 1. exact normalized-name within (type)
85
116
  for e in ents:
86
117
  key = (e["entity_type"], _norm(e["canonical_name"]))
87
118
  groups.setdefault(key, []).append(e)
88
- exact = [g for g in groups.values() if len(g) > 1]
89
-
90
- # 2. cross-run shared-provenance: same type + same event in provenance,
91
- # where some members are junk-leaning (looks-like-id) — catches
92
- # name-divergent dupes like "1716801984" vs "Katie Cooper" that never
93
- # block on name.
94
- #
95
- # OVER-MERGE GUARD: a single event can legitimately mention several
96
- # distinct same-type entities (an email naming Alice, Bob, AND a
97
- # numeric-ID node). Merging the whole co-occurrence group would
98
- # conflate Alice and Bob. So the no-LLM tier ONLY proposes when the
99
- # group has EXACTLY ONE non-junk member: we fold the junk node(s) into
100
- # that unambiguous real master. Groups with 0 or >=2 non-junk members
101
- # are ambiguous and deferred to the LLM-adjudicated tier
102
- # (entity_resolution_v2.py) rather than auto-merged.
103
- by_event_type: dict[tuple, list[dict]] = {}
104
- for e in ents:
105
- for ev in (e["provenance_event_ids"] or []):
106
- by_event_type.setdefault((e["entity_type"], ev), []).append(e)
107
- cross = []
108
- seen_ids: set[str] = set()
109
- for members in by_event_type.values():
110
- if len(members) < 2:
111
- continue
112
- junk = [m for m in members if C.looks_like_id(m["canonical_name"])]
113
- non_junk = [m for m in members if not C.looks_like_id(m["canonical_name"])]
114
- if not junk or len(non_junk) != 1:
115
- continue # need junk to clean AND exactly one unambiguous master
116
- group = non_junk + junk
117
- ids = tuple(sorted(m["id"] for m in group))
118
- if ids in seen_ids:
119
- continue
120
- seen_ids.add(ids)
121
- cross.append(group)
122
- return exact + cross
119
+ return [g for g in groups.values() if len(g) > 1]
123
120
 
124
121
 
125
122
  def _authority_signals(cur, arena: str, entity_ids: list[str], current_model: str) -> dict:
@@ -216,22 +213,33 @@ def _execute_entity_plan(cur, plan) -> None:
216
213
  cur.execute(
217
214
  """INSERT INTO entity_merges (id, arena, canonical_id, deprecated_id,
218
215
  deprecated_canonical_name, deprecated_aliases, merge_signal,
219
- facts_repointed, rollback_payload)
220
- VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s::jsonb)""",
216
+ facts_repointed, relationships_repointed, merged_by, rollback_payload)
217
+ VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s::jsonb)""",
221
218
  ("em_" + uuid.uuid4().hex[:20], a["arena"], a["canonical_id"], a["deprecated_id"],
222
219
  a["deprecated_canonical_name"], a["deprecated_aliases"], a["merge_signal"],
223
220
  len(plan.fact_subject_repoints) + len(plan.fact_object_repoints),
221
+ len(plan.rel_endpoint_repoints), "fusion-drive",
224
222
  json.dumps(a["rollback_payload"], default=str)),
225
223
  )
226
224
  cur.execute("DELETE FROM entities WHERE id = ANY(%s)", (plan.deprecated_entity_ids,))
227
225
 
228
226
 
229
227
  def _dedup_master_facts(cur, arena: str, master_id: str) -> int:
230
- """After repointing facts onto the master, the master can hold several
231
- facts with the same (subject, predicate, object) but different statements
232
- (fact id is content_id(arena, statement), so they didn't collapse on
233
- insert). Fuse each such triple-group via build_fact_merge_plan: keep the
234
- best, union provenance, delete dups with a fact_merges receipt."""
228
+ """After repointing facts onto the master, collapse facts that are now
229
+ TRUE duplicates same (subject, predicate, object) AND the same normalized
230
+ statement. These exist because the fact id is content_id(arena, statement):
231
+ two rows with statements differing only in case/whitespace hash to distinct
232
+ ids and so survived insert-time dedup; once their subject/object entities
233
+ are unified they are genuinely the same assertion and fuse safely.
234
+
235
+ The statement is PART OF THE KEY on purpose. Grouping on the triple alone is
236
+ NOT identity: a NULL object with a generic predicate (e.g. subject "said"
237
+ NULL) buckets together unrelated assertions, and build_fact_merge_plan would
238
+ keep one and DELETE the rest — destroying distinct facts (it deleted 33% of
239
+ one arena's facts that way before this fix). Same-triple / different-meaning
240
+ facts are left untouched here; the LLM semantic tier (_semantic_fact_groups
241
+ + adjudicate_facts) is the only thing allowed to fuse facts whose statements
242
+ actually differ, and only on an affirmative same-assertion verdict."""
235
243
  cur.execute(
236
244
  """SELECT id, predicate, object_entity_id, statement, confidence, provenance_event_ids
237
245
  FROM facts
@@ -241,8 +249,12 @@ def _dedup_master_facts(cur, arena: str, master_id: str) -> int:
241
249
  rows = cur.fetchall()
242
250
  groups: dict[tuple, list[dict]] = {}
243
251
  for r in rows:
244
- # group key uses the master as the subject anchor + predicate + object
245
- groups.setdefault((master_id, r["predicate"], r["object_entity_id"]), []).append(r)
252
+ # key = master subject anchor + predicate + object + NORMALIZED STATEMENT.
253
+ # statement in the key => only byte-equal-after-normalization dupes fuse.
254
+ groups.setdefault(
255
+ (master_id, r["predicate"], r["object_entity_id"], _norm(r["statement"] or "")),
256
+ [],
257
+ ).append(r)
246
258
  deduped = 0
247
259
  for dup in groups.values():
248
260
  plan = build_fact_merge_plan(arena=arena, dup_facts=dup)
@@ -264,11 +276,19 @@ def _dedup_master_facts(cur, arena: str, master_id: str) -> int:
264
276
  return deduped
265
277
 
266
278
 
267
- def _ambiguous_cross_run(cur, arena: str) -> list[dict]:
268
- """Co-occurrence groups the deterministic tier PUNTED on: a junk-leaning
269
- node sharing an event with MULTIPLE non-junk candidates (so which real
270
- entity it belongs to is ambiguous). Returns [{junk, candidates:[...]}] for
271
- the LLM tier to adjudicate. Restricted disclosure excluded."""
279
+ def _cooccurrence_candidates(cur, arena: str) -> list[dict]:
280
+ """ALL co-occurrence merge candidates a junk-leaning node (looks_like_id)
281
+ sharing a provenance event with one OR more non-junk same-type nodes.
282
+ Returns [{junk, candidates:[...]}] for the LLM tier to adjudicate.
283
+
284
+ Co-occurrence is never sufficient evidence on its own (the deterministic
285
+ tier no longer auto-merges any of these — see _entity_dup_sets). Even the
286
+ single-candidate case ('1716801984' co-occurs with exactly one 'Katie
287
+ Cooper') must be confirmed by the adjudicator, because it's
288
+ indistinguishable from 'RTX 6000' co-occurring with exactly one 'Linux'.
289
+ The adjudicator merges only on an affirmative same-entity verdict and fails
290
+ closed (no merge) when the distiller is unreachable. Restricted disclosure
291
+ excluded."""
272
292
  cur.execute(
273
293
  """SELECT id, entity_type, canonical_name, aliases, provenance_event_ids
274
294
  FROM entities WHERE arena = %s AND disclosure_class <> 'restricted'""",
@@ -283,7 +303,7 @@ def _ambiguous_cross_run(cur, arena: str) -> list[dict]:
283
303
  for members in by_event_type.values():
284
304
  junk = [m for m in members if C.looks_like_id(m["canonical_name"])]
285
305
  non_junk = [m for m in members if not C.looks_like_id(m["canonical_name"])]
286
- if not junk or len(non_junk) < 2:
306
+ if not junk or len(non_junk) < 1:
287
307
  continue
288
308
  for j in junk:
289
309
  key = (j["id"], tuple(sorted(c["id"] for c in non_junk)))
@@ -366,14 +386,17 @@ def main() -> int:
366
386
  merged += len(loser_ids)
367
387
  return len(loser_ids)
368
388
 
369
- # Tier 1 — deterministic (exact-name + unambiguous cross-run)
389
+ # Tier 1 — deterministic: exact normalized-name dupes only
390
+ # (case/whitespace variants). Co-occurrence is NOT auto-merged.
370
391
  for group in _entity_dup_sets(cur, args.arena):
371
392
  do_merge(group)
372
393
 
373
- # Tier 2 — LLM adjudication via the in-VPC distiller (no egress)
394
+ # Tier 2 — LLM adjudication via the in-VPC distiller (no egress).
395
+ # ALL co-occurrence merges live here now — single- and multi-
396
+ # candidate alike — because co-occurrence never proves identity.
374
397
  if post_fn:
375
- # 2a. ambiguous cross-run: which real entity does the junk match?
376
- for amb in _ambiguous_cross_run(cur, args.arena):
398
+ # 2a. co-occurrence: does the junk node match a real entity?
399
+ for amb in _cooccurrence_candidates(cur, args.arena):
377
400
  j = amb["junk"]
378
401
  jctx = _entity_context(cur, args.arena, j["id"])
379
402
  for cand in amb["candidates"]:
@@ -0,0 +1,83 @@
1
+ """Guards _dedup_master_facts against the over-fusion that deleted 33% of an
2
+ arena's facts (2026-06-14): grouping post-merge facts by (subject, predicate,
3
+ object) alone treats a NULL object + generic predicate as identity and deletes
4
+ distinct assertions. The statement must be part of the dedup key so ONLY
5
+ byte-equal-after-normalization duplicates fuse; same-triple/different-meaning
6
+ facts are left for the LLM semantic tier."""
7
+
8
+ from __future__ import annotations
9
+
10
+ import importlib
11
+ import os
12
+ import sys
13
+ import types
14
+
15
+ HERE = os.path.dirname(__file__)
16
+
17
+
18
+ def _load_fuse(monkeypatch):
19
+ fake_psycopg = types.ModuleType("psycopg")
20
+ fake_rows = types.ModuleType("psycopg.rows")
21
+ fake_rows.dict_row = object()
22
+ fake_psycopg.rows = fake_rows
23
+ monkeypatch.setitem(sys.modules, "psycopg", fake_psycopg)
24
+ monkeypatch.setitem(sys.modules, "psycopg.rows", fake_rows)
25
+ monkeypatch.syspath_prepend(os.path.join(HERE, "..", "fusion_drive"))
26
+ monkeypatch.syspath_prepend(HERE)
27
+ sys.modules.pop("fusion_drive_fuse", None)
28
+ return importlib.import_module("fusion_drive_fuse")
29
+
30
+
31
+ class FakeCursor:
32
+ """Returns preset fact rows on the SELECT; records ids passed to DELETE."""
33
+
34
+ def __init__(self, rows):
35
+ self._rows = rows
36
+ self.deleted_ids = []
37
+
38
+ def execute(self, sql, params=None):
39
+ s = " ".join(sql.split())
40
+ if s.startswith("DELETE FROM facts WHERE id = ANY"):
41
+ # params is a 1-tuple holding the id list
42
+ self.deleted_ids.extend(params[0])
43
+ # SELECT / UPDATE / INSERT: no-op (fetchall serves the preset rows)
44
+
45
+ def fetchall(self):
46
+ return self._rows
47
+
48
+
49
+ def _fact(fid, predicate, obj, statement, conf):
50
+ return {
51
+ "id": fid,
52
+ "predicate": predicate,
53
+ "object_entity_id": obj,
54
+ "statement": statement,
55
+ "confidence": conf,
56
+ "provenance_event_ids": [f"ev_{fid}"],
57
+ }
58
+
59
+
60
+ def test_distinct_statements_same_triple_are_NOT_fused(monkeypatch):
61
+ fuse = _load_fuse(monkeypatch)
62
+ rows = [
63
+ _fact("f1", "said", None, "Standing by", 0.9),
64
+ _fact("f2", "said", None, "yeah ship it", 0.8), # distinct meaning
65
+ _fact("f3", "said", None, "modules/deep-memory is vestigial", 0.7),
66
+ ]
67
+ cur = FakeCursor(rows)
68
+ deleted = fuse._dedup_master_facts(cur, "arena", "m")
69
+ assert deleted == 0, "must not fuse same-triple facts with different statements"
70
+ assert cur.deleted_ids == []
71
+
72
+
73
+ def test_only_normalized_statement_duplicates_fuse(monkeypatch):
74
+ fuse = _load_fuse(monkeypatch)
75
+ rows = [
76
+ _fact("f1", "said", None, "Standing by", 0.9),
77
+ _fact("f2", "said", None, "standing by", 0.5), # same after _norm
78
+ _fact("f3", "said", None, "something else entirely", 0.7),
79
+ ]
80
+ cur = FakeCursor(rows)
81
+ deleted = fuse._dedup_master_facts(cur, "arena", "m")
82
+ assert deleted == 1, "the case/whitespace duplicate should fuse"
83
+ assert cur.deleted_ids == ["f2"], "lower-confidence true-dupe is the one deleted"
@@ -0,0 +1,70 @@
1
+ """Guards the distiller circuit breaker in fusion_drive_fuse._distiller_post_fn.
2
+
3
+ When the distiller is scaled to zero (the autoscaler scales it on the distill
4
+ queue, not on fusion need, so it is routinely down at sweep time) the first
5
+ connect must fail fast AND open the circuit so every later candidate pair
6
+ short-circuits instead of eating another connect timeout. A regression here
7
+ reintroduces the >20min dry-run grind that motivated the fix."""
8
+
9
+ from __future__ import annotations
10
+
11
+ import importlib
12
+ import os
13
+ import sys
14
+ import types
15
+
16
+ import httpx
17
+ import pytest
18
+
19
+ HERE = os.path.dirname(__file__)
20
+
21
+
22
+ def _load_fuse(monkeypatch):
23
+ # The script imports psycopg + adjudicate/canonical/merge at module level;
24
+ # stub psycopg (not installed in the unit env) and put the sibling package
25
+ # on the path so the real adjudicate/canonical/merge import cleanly.
26
+ fake_psycopg = types.ModuleType("psycopg")
27
+ fake_rows = types.ModuleType("psycopg.rows")
28
+ fake_rows.dict_row = object()
29
+ fake_psycopg.rows = fake_rows
30
+ monkeypatch.setitem(sys.modules, "psycopg", fake_psycopg)
31
+ monkeypatch.setitem(sys.modules, "psycopg.rows", fake_rows)
32
+ monkeypatch.syspath_prepend(os.path.join(HERE, "..", "fusion_drive"))
33
+ monkeypatch.syspath_prepend(HERE)
34
+ sys.modules.pop("fusion_drive_fuse", None)
35
+ return importlib.import_module("fusion_drive_fuse")
36
+
37
+
38
+ def test_circuit_opens_after_first_connect_failure(monkeypatch):
39
+ fuse = _load_fuse(monkeypatch)
40
+ calls = {"n": 0}
41
+
42
+ def boom(*a, **k):
43
+ calls["n"] += 1
44
+ raise httpx.ConnectError("connection refused")
45
+
46
+ monkeypatch.setattr(httpx, "post", boom)
47
+ post = fuse._distiller_post_fn("http://10.0.0.1:8005/v1/chat/completions", "m")
48
+
49
+ with pytest.raises(RuntimeError, match="distiller unreachable"):
50
+ post([{"role": "user", "content": "a"}])
51
+ # second call must NOT hit the network again — circuit is open
52
+ with pytest.raises(RuntimeError, match="circuit open"):
53
+ post([{"role": "user", "content": "b"}])
54
+
55
+ assert calls["n"] == 1, "circuit breaker should make exactly one network attempt"
56
+
57
+
58
+ def test_happy_path_returns_content(monkeypatch):
59
+ fuse = _load_fuse(monkeypatch)
60
+
61
+ class FakeResp:
62
+ def raise_for_status(self):
63
+ pass
64
+
65
+ def json(self):
66
+ return {"choices": [{"message": {"content": "yes"}}]}
67
+
68
+ monkeypatch.setattr(httpx, "post", lambda *a, **k: FakeResp())
69
+ post = fuse._distiller_post_fn("http://10.0.0.1:8005/v1/chat/completions", "m")
70
+ assert post([{"role": "user", "content": "a"}]) == "yes"