pi-jev-lens 0.5.0 → 0.5.2

package/README.md

@@ -99,7 +99,7 @@ For development, clone the repository and load it directly:
 ```sh
 git clone https://github.com/dizk/jev-lens.git && cd jev-lens && npm install
 echo 'TYPESAFE_API_KEY=...' > .env
-pi -e ./packages/pi/index.ts
+pi -e ./packages/pi-jev-lens/index.ts
 ```
 
 ## How it works
@@ -221,8 +221,10 @@ whether the agent needed the text. The benchmark uses real OpenHands trajectorie
 - edit-miss: the agent later edited a line that the view had dropped.
 - quote-miss: the agent quoted text that the view had dropped.
 - ref-miss: the agent used an identifier that only existed in the dropped part.
+- recall-miss: in a recorded session with the extension, the agent recalled the result, and the replayed view would
+  still not have answered that recall (the whole output, a line range or a pattern). Only real sessions have recalls.
 
-An edit-miss counts five times in the objective, and edit-misses are rare. So you must score every change that
+The objective is saved% − 5·edit-miss% − 2·quote-miss% − ref-miss% − 2·recall-miss%. An edit-miss counts five times, and edit-misses are rare. So you must score every change that
 touches code views on the 500-trajectory slice, not only on the 100-trajectory holdout.
 
 | slice | large results | saved | edit-miss | quote-miss | ref-miss |
@@ -271,10 +273,10 @@ as a library; its README shows the API. What is where:
 | the jev questions, the state shape, the decision rule, block expansion | `packages/core/src/presend.ts` | `@typesafe-ai/sdk` |
 | the parser for bash display commands | `packages/core/src/shell-display.ts` | nothing |
 | the recall tool's slicing | `packages/core/src/recall.ts` | nothing |
-| the wiring for pi (tool_result, the recall tool, the UI) | `packages/pi/index.ts`, `packages/pi/src/ui.ts` | pi |
+| the wiring for pi (tool_result, the recall tool, the UI) | `packages/pi-jev-lens/index.ts`, `packages/pi-jev-lens/src/ui.ts` | pi |
 | the wiring for Claude Code (PostToolUse hook, MCP server) | `packages/claude-code/src/` | Claude Code |
 | the benchmark and the metrics on real trajectories | `eval/presend-score.ts`, `eval/bench/` | run `eval/bench/fetch.sh` first |
-| post-send decisions and the frozen ledger (pi only) | `packages/core/src/classifier.ts`, `packages/pi/src/policy.ts`, `packages/pi/src/ledger.ts` | `@typesafe-ai/sdk` |
+| post-send decisions and the frozen ledger (pi only) | `packages/core/src/classifier.ts`, `packages/pi-jev-lens/src/policy.ts`, `packages/pi-jev-lens/src/ledger.ts` | `@typesafe-ai/sdk` |
 
 The sequence is this. When a large tool result arrives, code builds the views from the text. jev says which view is
 enough and whether the exact text is needed. Code applies the selection policy. If needed, a second request expands

package/package.json

@@ -1,13 +1,13 @@
 {
   "name": "pi-jev-lens",
-  "version": "0.5.0",
+  "version": "0.5.2",
   "description": "pi extension that compresses large tool results before they reach the model: jev picks the view (outline, relevant blocks, sections, signals, testlog), full text stays recallable",
   "author": "Didrik Rognstad",
   "license": "MIT",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/dizk/jev-lens.git",
-    "directory": "packages/pi"
+    "directory": "packages/pi-jev-lens"
   },
   "homepage": "https://github.com/dizk/jev-lens#readme",
   "bugs": {