agenr 0.9.41 → 0.9.43

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/CHANGELOG.md CHANGED
@@ -1,5 +1,61 @@
1
1
  # Changelog
2
2
 
3
+ ## 0.9.43 (2026-03-02)
4
+
5
+ ### Features
6
+
7
+ - Embedding-based clustering fallback for orphan entries (#387)
8
+ - Assigns orphans to nearest existing cluster by embedding similarity
9
+ - Forms new clusters from orphans with similar embeddings
10
+ - Configurable similarity threshold (default 0.72), max cluster size (default 50)
11
+ - New `method` column on clusters table (`co_recall` | `embedding`)
12
+ - Reports fallback stats in detectAndStoreClusters result
13
+
14
+ ### Bug Fixes
15
+
16
+ - Replace sigmoid scoring compression with sqrt for better score differentiation (#390)
17
+ - Old formula squashed all scores into 0.22-0.26 band (0.033 spread)
18
+ - New sqrt compression gives 0.50-0.67 range (0.17 spread, 5x improvement)
19
+ - FTS bonus moved from additive post-compression (0.15) to multiplicative pre-compression (1.3x)
20
+ - Importance/recall blend (70/30) replaces MAX - recall history now always contributes
21
+ - Apply --limit flag to reflection task cluster count (#385)
22
+ - Tighten extraction importance calibration to reduce inflation (#391)
23
+ - Added stronger downward pressure in calibration text
24
+ - Rebalanced examples: two 8s lowered to 7, added importance-6 example
25
+ - Production showed 30.7% of entries at 8+ vs 20% target
26
+
27
+ ### Tests
28
+
29
+ - 8 new tests for embedding fallback clustering
30
+ - 3 new tests for scoring (spread, FTS proportionality, recall contribution)
31
+ - 1 new test for reflection limit enforcement
32
+
33
+ ## 0.9.42 (2026-03-02)
34
+
35
+ ### Features
36
+ - **Periodic reflection and synthesis** (#269): New `reflection` maintain task that
37
+ synthesizes clusters of related entries into higher-level reflections using
38
+ two-step LLM prompting (question generation + synthesis) with a verification pass.
39
+ Source entries are importance-demoted (not retired) so raw knowledge remains
40
+ accessible. Co-recall edges are transferred to reflection entries for graph
41
+ continuity.
42
+ - New entry type: `reflection` for synthesized knowledge
43
+ - New `reflections` table for process metadata
44
+ - Extended `entry_sources` with `action` column for unified provenance tracking
45
+ - New `--full` flag for full-brain reflection (ignores change detection)
46
+ - Configurable: reflection model, verification model, coverage threshold, demotion
47
+ amount, importance threshold, cluster size minimum
48
+ - Re-synthesis guards: minimum cumulative importance threshold for changed entries
49
+
50
+ ### Bug Fixes
51
+ - Fix parseSynthesisResponse bailing on first JSON candidate instead of trying alternatives
52
+ - Fix verification prompt missing explicit check instructions
53
+ - Fix demoteEntryImportance/retireReflectionEntry using wall clock instead of injected time
54
+ - Fix transferCoRecallEdges silently dropping stronger edges (now uses MAX weight)
55
+ - Fix verification ignoring the verified boolean (only checked coverage_score)
56
+ - Extract shared JSON parsing utilities to reduce duplication
57
+ - Add cluster churn detection for re-synthesis safety
58
+
3
59
  ## 0.9.41 - 2026-03-01
4
60
 
5
61
  ### Fixed