@dzhechkov/harness-cli 0.3.204 → 0.3.205

package/README.md

@@ -1248,6 +1248,48 @@ Bridge: with the agentdb backend enabled, `dz recall` itself is the bridge — i
 
 `dz recall` is **hybrid** when the vector tier is available and **exactly the old lexical command** when it is not — enabling it never changes behavior for projects that skip it.
 
+#### The apply leg — learned lessons injected into the prompt that needs them
+
+A feedback loop has three legs: **collect → rank → apply**. dz collected automatically and never read
+back: learned lessons reached an agent only when `feature-adr` asked, or when a human typed
+`dz recall`. A store nobody reads is a write-only log.
+
+Two hooks close the loop:
+
+| hook | what it does |
+|---|---|
+| `.claude/helpers/dz-embed-daemon.mjs` (SessionStart) | holds the embedder resident and answers recall over a unix socket |
+| `.claude/helpers/recall-hook.cjs` (UserPromptSubmit) | scores your prompt against the store and injects only what clears the relevance floor |
+
+**Silence is the feature.** When nothing is relevant the hook prints nothing at all. The floors are
+MEASURED, not chosen: on a 32-probe labeled set (16 RU + 16 EN, half irrelevant) a max-cosine floor of
+`0.38` for Cyrillic prompts and `0.31` for Latin ones classified every probe correctly. `"спасибо"`
+scores `0.259` and `"какой статус?"` `0.318` — both stay silent.
+A z-score `(max−mean)/sd` was tried and **failed at 69 % accuracy** (MEASURED — reproducer `npx vitest run test/recall-hook-policy.test.ts`).
+Why it fails: an off-topic query has a flat similarity profile, so its best hit
+stands out relative to its own mean. A prompt shorter than `10` characters or carrying fewer than two
+content tokens is skipped outright — a bare `"json"` really is close to technical lessons, but carries
+no intent.
+
+**Cost, stated honestly.** Loading the embedder takes ~1478 ms; a warm embed takes 6–8 ms. Without a
+resident model the hook would add ~2 s to every prompt and you would switch it off within a day. With
+the daemon a turn costs **50–83 ms**, and **37 ms** when the daemon is absent (it then stays silent
+rather than blocking). The daemon holds the model in memory — **~427 MB RSS** — and exits after 30
+minutes idle. It refuses to start twice and opens the store read-only, so it can never be the writer
+that tears a database file for a concurrent reader.
+
+Nothing about it can stall a turn: no daemon, no deps, a malformed payload, a slow socket — all exit 0
+in silence.
+
+```bash
+# ask in plain language; the hook has already surfaced the relevant lessons
+"почему codex барьер даёт ложный grade D"      # → injects the matching lesson (0.63)
+"спасибо"                                       # → nothing injected
+```
+
+Disable it by removing the two hook entries from `.claude/settings.json`; stop the daemon with
+`printf '{"op":"stop"}\n' | nc -U .dz/embed.sock`.
+
 #### Cross-lingual recall, and changing the embedding model
 
 The default embedder (`Xenova/paraphrase-multilingual-MiniLM-L12-v2`) is **multilingual**: a Russian

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dzhechkov/harness-cli",
-  "version": "0.3.204",
+  "version": "0.3.205",
   "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",
@@ -54,7 +54,7 @@
     "@dzhechkov/skills-reverse-engineering": "^0.1.0",
     "@dzhechkov/skills-presentation-storyteller": "^0.1.0",
     "@dzhechkov/skills-website-cloner": "^0.1.0",
-    "@dzhechkov/harness-core": "0.3.103"
+    "@dzhechkov/harness-core": "0.3.104"
   },
   "devDependencies": {
     "@types/node": "^25.6.0",