@lmzhen/dsh-evolution-replay 0.32.4 → 0.32.6

package/README.md

@@ -1,18 +1,13 @@
 # @lmzhen/dsh-evolution-replay
 
-Replay/A-B evaluation primitives for evolution plans: it reads the activity sidecar back,
-compares plan outcomes and provides the `evolutionReplay` service behind `evolution
-replay`.
+Scores and ranks recorded plan outcomes; it does not re-execute a plan.
 
-## Model surface
+`/evolution replay` compares live `evolution/plan-applied` records together with a one-time backfill from the activity sidecar. Scores use accepted/rejected operations, evidence, and estimated input cost. Execution failures and unread-skill skips remain visible in the report rather than being silently scored as clean zero-operation plans.
 
-- **Model-visible:** nothing of its own: no prompt section and no tool schema; `/evolution replay` is a human command; consumers own the model-visible effects.
-- **Prompt prefix / KV cache:** independent of request-prefix construction — it does not alter the assembled prompt or tool list; family rules: `packages/README.md` §"Model-visible prompt prefix and the KV cache".
-- **Mount it?** yes — the `evolution-replay` row, carried by the `evolution-host`, `evolution-all` and one-click `evolution-preset` bundles.
+`maxPlans` defaults to 50. The weight fields are `accepted`, `rejectedPenalty`, `evidence`, and `cost`; see [PARAMETERS.md](../PARAMETERS.md) for their deployment tier and ownership.
 
-## Known limitations
+## Data availability
 
-- No known durable consumer gaps at this time. Runtime contracts are covered by package and boundary tests.
-- **The activity-sidecar verdict is per read, and an unanswered read is its own state.** When `apply()` meets bytes this build cannot read as the current format, every `compare()` result carries `sourceCorrupt: true` and its report states the leaderboard is NOT the recorded history (an empty list does not mean nothing happened); a later successful read clears the qualification (a repaired sidecar plus an io reload (HMR / plugin restart)), while the corruption stays observable through the one warn `apply()` emits at mark time. A read that **THREW** (the io failed, or no state could be produced at all) is neither verdict: `compare()` then carries `sourceReadFailure: <message>` and its report says the read FAILED — never "nothing was recorded". Any ANSWERED read (clean or corrupt) retires that failure, so both qualifications describe the current read rather than latching.
+Default inject-mode review emits no `plan-applied`, so it produces no plan-based replay data. Branch review emits that event. An empty leaderboard therefore does not prove that no reviews ran.
 
-**Runtime invariant:** No companion is published. The platform auto-assembles nothing and the family mounts no `<pkg>/invariant` cordis row, so a companion here would never execute (v37 S2.1 / I-3).
+Unreadable activity bytes and failed backfill reads are carried as explicit report qualifications; neither is presented as a clean empty history.

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@lmzhen/dsh-evolution-replay",
   "description": "Replay/A-B evaluation primitives for evolution plans (community build)",
-  "version": "0.32.4",
+  "version": "0.32.6",
   "publishConfig": {
     "access": "public"
   },
@@ -27,15 +27,15 @@
   "license": "MIT",
   "dependencies": {
     "@deepseek-ai/schemastery": "^3.18.5",
-    "@lmzhen/dsh-evolution-activity": "^0.32.4",
-    "@lmzhen/dsh-evolution-core": "^0.32.4"
+    "@lmzhen/dsh-evolution-activity": "^0.32.6",
+    "@lmzhen/dsh-evolution-core": "^0.32.6"
   },
   "peerDependencies": {
     "@deepseek-ai/cordis": "^4.0.5"
   },
   "devDependencies": {
-    "@lmzhen/dsh-evolution-activity": "^0.32.4",
-    "@lmzhen/dsh-evolution-core": "^0.32.4"
+    "@lmzhen/dsh-evolution-activity": "^0.32.6",
+    "@lmzhen/dsh-evolution-core": "^0.32.6"
   },
   "engines": {
     "dsh": ">=0.2.1-rc.2 <0.2.2-0"