agenr 2.0.1 → 3.0.0

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.
package/CHANGELOG.md CHANGED
@@ -2,6 +2,70 @@
2
2
 
3
3
  ## [Unreleased]
4
4
 
5
+ ## [3.0.0] - 2026-05-31
6
+
7
+ Skeln host integration, session and working memory, and goal-control major release.
8
+
9
+ ### Added
10
+
11
+ - **Skeln now has a publishable agenr plugin and runtime adapter.** Agenr ships `@agenr/skeln-plugin` with Skeln-native tool wiring, recall/store/update/work/goal surfaces, plugin packaging, setup docs, and host scope handling.
12
+ - **Session memory and working memory now support Skeln continuity.** The release adds session-memory repositories, trigger routing, continuity records, transient working-set projection, lifecycle checkpoints, external goal preparation, and durable work-event persistence.
13
+
14
+ ### Changed
15
+
16
+ - **OpenClaw and Skeln share more memory runtime infrastructure.** Entry tools, injection policy, memory-policy parsing, target resolution, claim support, and host-memory composition moved into shared adapters so future host integrations reuse the same behavior.
17
+ - **Session-start and before-turn memory policy is more explicit.** Plugin configuration now supports enabled and core-memory controls, JSON string policy parsing for Skeln, and tighter session-start and before-turn formatting.
18
+ - **Plugin dependencies now use the `@earendil-works` package namespace.** Runtime imports and lockfile metadata were updated for the current pi package names.
19
+
20
+ ### Fixed
21
+
22
+ - **Session-memory feature gating is quiet when disabled.** Session-memory trigger logging now treats feature-disabled intake as an intentional no-op instead of noisy runtime feedback.
23
+ - **Goal-generation documentation is clearer.** Working-memory goal seed text now documents the initial goal constant without changing behavior.
24
+
25
+ ### Validation
26
+
27
+ Changes since last push to `origin/master`:
28
+
29
+ - Add the publishable Skeln plugin package, Skeln plugin docs, and Skeln runtime adapter
30
+ - Add shared plugin-runtime composition for OpenClaw and Skeln host integrations
31
+ - Add session-memory schema, repositories, trigger routing, continuity, and post-compaction handling
32
+ - Add working-memory schema, repositories, projections, operation handlers, checkpoints, and goal continuation
33
+ - Add Skeln `agenr_work` and `agenr_goal` flows with regression coverage
34
+ - Add session episode ingestion, shutdown and goal-close episode support, and bounded episode services
35
+ - Update OpenClaw and Skeln memory-policy configuration, session-start injection, and before-turn formatting
36
+ - Move pi runtime dependencies to `@earendil-works` packages
37
+ - Fix disabled session-memory intake logging and refresh goal-generation documentation
38
+
39
+ ## [2.1.0] - 2026-04-22
40
+
41
+ Recall ranking overhaul, before-turn memory patching, and eval/debug tooling minor release.
42
+
43
+ ### Added
44
+
45
+ - **Before-turn memory patching and eval coverage ship as a first-class recall surface.** Agenr now supports hybrid session-start patch selection, dedicated before-turn patch selection and reranking, internal before-turn eval routes, richer diagnostics, and corpus-backed fixture seeding for replay coverage.
46
+ - **Cross-encoder reranking and OpenClaw debug artifacts are now available.** The release adds an OpenAI-compatible cross-encoder stage, opt-in replay debug artifacts, and an agenr-only JSONL debug sink for OpenClaw runtime investigation.
47
+
48
+ ### Changed
49
+
50
+ - **Recall ranking is now built around reciprocal rank fusion plus diversification.** Unified recall now uses RRF as the primary relevance combiner, typed neighborhood expansion, MMR diversification, lineage-aware shaping, and tighter ordering and abstention rules.
51
+ - **Documentation and procedures were refreshed for the new recall model.** The recall guide was rewritten as a reference, procedural docs were updated, and the Obsidian LLM Wiki procedure was added to the repo-authored procedure set.
52
+
53
+ ### Fixed
54
+
55
+ - **Before-turn selector edge cases are more stable.** The release fixes fallback selection, short-name family matching, cross-encoder wiring, and small-pool top-1 ranking regressions that showed up while hardening the new ranking stack.
56
+
57
+ ### Validation
58
+
59
+ Changes since last push to `origin/master`:
60
+
61
+ - Add Obsidian LLM Wiki procedure and refresh procedures documentation
62
+ - Add before-turn memory patch selection, reranking, diagnostics, and internal eval seam
63
+ - Adopt reciprocal rank fusion, typed neighborhood expansion, and MMR diversification across recall pipelines
64
+ - Add cross-encoder reranking and wire it through recall and before-turn eval seams
65
+ - Tune ordering, abstention, thresholds, and lineage shaping for snapshot and before-turn ranking quality
66
+ - Add precision-first entity attribute recall
67
+ - Refresh recall and eval documentation and land corpus-backed replay and debug tooling
68
+
5
69
  ## [2.0.1] - 2026-04-13
6
70
 
7
71
  OpenClaw startup-failure handling patch release.
@@ -1,5 +1,5 @@
1
1
  import * as openclaw_plugin_sdk_plugin_entry from 'openclaw/plugin-sdk/plugin-entry';
2
- import { T as TranscriptPort, P as ParsedTranscript } from '../../ports-D2NOK2gR.js';
2
+ import { T as TranscriptPort, P as ParsedTranscript } from '../../ports-CpzWESmZ.js';
3
3
 
4
4
  /**
5
5
  * Parses OpenClaw JSONL session exports into normalized agenr transcripts.