pi-sleep-agent 0.4.0 → 0.5.0

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.
Files changed (2) hide show
  1. package/package.json +1 -1
  2. package/sleepkit/store.py +47 -0
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "pi-sleep-agent",
3
- "version": "0.4.0",
3
+ "version": "0.5.0",
4
4
  "description": "A reviewable sleep and dream cycle for Pi agent memory.",
5
5
  "type": "module",
6
6
  "license": "MIT",
package/sleepkit/store.py CHANGED
@@ -146,6 +146,15 @@ def buffer_counts() -> dict:
146
146
 
147
147
 
148
148
  # ------------------------------------------------------- similarity/dedup ---
149
+
150
+ # A proposal that paraphrases an already-refused candidate must not reborn as a
151
+ # fresh provisional idea the reviewer then judges all over again. Token-set
152
+ # Jaccard (the general duplicate bar) falls off too fast to catch even a close
153
+ # restate, so against a *refused* idea we compare by character bigrams, which
154
+ # tolerate synonym swaps and reordering, and block on a lower bar. The idea was
155
+ # already judged and rejected, so colliding with it should refuse the write.
156
+ REFUSED_DUP_THRESHOLD = 0.42
157
+
149
158
  _WORD = re.compile(r"[a-z0-9_]+")
150
159
 
151
160
 
@@ -153,6 +162,14 @@ def _tokens(text: str) -> set:
153
162
  return set(_WORD.findall((text or "").lower()))
154
163
 
155
164
 
165
+ def _bigrams(text: str) -> set:
166
+ """Character bigrams of the lowercased text, for paraphrase-resistant
167
+ similarity. A pair of observations that differ only in wording shares most
168
+ bigrams, where token Jaccard already diverged."""
169
+ t = (text or "").lower()
170
+ return set(t[i:i + 2] for i in range(len(t) - 1))
171
+
172
+
156
173
  def similarity(a: str, b: str) -> float:
157
174
  ta, tb = _tokens(a), _tokens(b)
158
175
  if not ta or not tb:
@@ -169,6 +186,32 @@ def nearest(text: str, pool: list[tuple[str, str]], threshold: float = 0.6) -> t
169
186
  return (best, score) if score >= threshold else ("", score)
170
187
 
171
188
 
189
+ def ngram_similarity(a: str, b: str) -> float:
190
+ ga, gb = _bigrams(a), _bigrams(b)
191
+ if not ga or not gb:
192
+ return 0.0
193
+ return len(ga & gb) / float(len(ga | gb))
194
+
195
+
196
+ def refused_collision(text: str, threshold: float = REFUSED_DUP_THRESHOLD) -> tuple[str, float]:
197
+ """Best matching already-refused candidate body, or ("", 0.0) under threshold.
198
+
199
+ Refusals are the one class we hard-block aggressively: the idea was judged and
200
+ rejected, so coming back paraphrased must not become a new provisional review.
201
+ Bigram Jaccard separates a restate of the same fact from a different fact about
202
+ the same subject far better than token Jaccard on short observations, which is
203
+ why this gate uses it instead of similarity().
204
+ """
205
+ best, score = "", 0.0
206
+ for m in list_candidates(with_body=True):
207
+ if m.get("status") != "refused":
208
+ continue
209
+ s = ngram_similarity(text, m.get("_body", ""))
210
+ if s > score:
211
+ best, score = m["id"], s
212
+ return (best, score) if score >= threshold else ("", score)
213
+
214
+
172
215
  # ------------------------------------------------------------- candidates ---
173
216
  def add_candidate(text: str, kind: str = "fact", source: str = "agent",
174
217
  trust: str = "observed", session: str = "", cwd: str = "",
@@ -196,6 +239,10 @@ def add_candidate(text: str, kind: str = "fact", source: str = "agent",
196
239
  if dup:
197
240
  return False, "duplicate of %s (%.2f); use `sleepctl supersede %s %s`" % (
198
241
  dup, score, dup, "new-text")
242
+ ref_id, ref_score = refused_collision(red)
243
+ if ref_id:
244
+ return False, "duplicate of earlier refusal %s (%.2f); this idea was judged and rejected" % (
245
+ ref_id, ref_score)
199
246
 
200
247
  rec = {
201
248
  "id": paths.new_id("c"),