@dzhechkov/harness-cli 0.3.214 → 0.3.215

package/README.md

@@ -1260,6 +1260,27 @@ When using `--memory agentdb`, the following algorithms automatically tune searc
 
 The bandit automatically selects the best algorithm for your usage pattern — no manual tuning needed.
 
+### SAFLA delta — rank lessons by payoff *slope*, not just level
+
+Ported from rUv's SAFLA (`safla/core/delta_evaluation.py`). Normal memory ranks a learned lesson by a
+**level** — how often it's been used, how recently, its reward. That can't tell a lesson that was hammered
+early and is now dead from one quietly proving itself lately. SAFLA delta measures the **slope**: the
+*change* in a lesson's payoff over time, computed from its reinforce history in `.dz/sessions.jsonl` (a
+weighted sum of four deltas — performance, efficiency, stability, capability).
+
+Two surfaces:
+
+- **`dz consolidate`** prints a SAFLA-delta section — which lessons are **rising** (`↑`, still paying off)
+  and which are **stale** (`↓`, slope ≤ 0). Stale lessons are shown as *prune candidates for review*, never
+  auto-deleted (a stale-but-valid lesson is not noise). Works out of the box.
+- **Opt-in recall re-rank.** Set `deltaRerank` in `.dz/config.json` to let the slope nudge recall ranking —
+  rising lessons up, stale ones down — as a **bounded** ± term (it never overrides lexical/vector
+  relevance). **Off by default** (byte-identical to the reinforce-only re-rank until you enable it):
+
+  ```json
+  { "memory": { "learning": { "deltaRerank": true } } }
+  ```
+
 ### How to enable AgentDB
 
 ```bash

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dzhechkov/harness-cli",
-  "version": "0.3.214",
+  "version": "0.3.215",
   "description": "The dz CLI — install AI skills for Claude Code, Codex, OpenCode, Hermes, OpenClaude, GitHub Copilot. 35 commands, 13 presets, 6 platform targets.",
   "type": "module",
   "license": "MIT",