@skill-harness/cli 0.24.2 → 0.25.0

package/README.md

@@ -41,3 +41,9 @@ npm i -g @skill-harness/cli
 
 - Repo + full docs: https://github.com/mojomanyana/skill-harness
 - Step-by-step usage: [`docs/USAGE.md`](https://github.com/mojomanyana/skill-harness/blob/main/docs/USAGE.md)
+
+## Optional decision evaluation and learning data
+
+`skill-harness decision --help` exposes the npm-installed offline fixture, comparison, selected-session import and reviewed LoRA export workflow. `/skill-harness jev enable` asks for a fresh per-session storage choice every time; declining still permits explicitly confirmed JEV calls. Storage retains selected unlabeled inputs only and never authorizes training.
+
+The exact Pi1.0.4 OAuth subscription comparison is separate from optional metered JEV. Reports preserve grouped splits, independent labels, abstention/error coverage and model/prompt identities. Synthetic fixtures qualify plumbing only. OpenAI Decisions remains excluded. See [the npm acceptance walkthrough](https://github.com/mojomanyana/skill-harness/blob/main/docs/DECISION-LEARNING.md) and the [data contracts](https://github.com/mojomanyana/skill-harness/blob/main/experiments/decision-shadow/LEARNING-WORKFLOW.md).

package/dist/cli.js

@@ -5,6 +5,7 @@ import { tmpdir } from "node:os";
 import { discover, resolveSkill, loadSpec, parseSpec, appendScenario, parseModelRef, runSkillModel, formatScorecard, readResults, regradeRun, lintSkill, failsGate, renderTemplateSpec, isTemplateSpec, renderDraftSpec, buildSuggestPrompt, parseSuggestDraft, rescoreRun, regateRun, specPathForRunDir, collectLift, collectReport, collectStability, boundaryCells, stabilityNote, PATH_LEGEND, restampSkill, computeCoverage, formatCoverage, HARNESS_VERSION, defaultJudge, recordedJudgeOrDefault, assertJudgeAllowed, assertNotDowngraded, downgradeWarning, isScoredMode, resolveArm, isFreeOfflineCommand, } from "@skill-harness/core";
 import { getAdapter } from "@skill-harness/adapters";
 import { serveReview } from "./serve.js";
+import { cmdDecision } from "./decision.js";
 const DEFAULT_MODEL = "fireworks:accounts/fireworks/models/deepseek-v4-pro";
 // The judge default lives in core (`defaultJudge()`), which resolves
 // SKILL_HARNESS_JUDGE over a baked value — it was duplicated in three places
@@ -754,6 +755,8 @@ export function help() {
   Every command with --skills also accepts --specs <root> (default: SKILL_HARNESS_SPECS).
   Skill text comes from --skills; test assets and results come from --specs.
 
+  decision <command>                          decision evaluation, session consent and learning export; decision --help
+
   version  print ${HARNESS_VERSION} and exit (also --version / -v)
 
 defaults: model=${DEFAULT_MODEL}  judge=${defaultJudge()}  mode=green  harness=pi
@@ -765,6 +768,8 @@ defaults: model=${DEFAULT_MODEL}  judge=${defaultJudge()}  mode=green  harness=p
 }
 export async function main(argv) {
     const cmd = argv[0];
+    if (cmd === "decision")
+        return cmdDecision(argv.slice(1));
     const args = parseArgs(argv.slice(1));
     assertNoRetiredFlags(cmd, args);
     switch (cmd) {

package/dist/decision.d.ts

@@ -0,0 +1 @@
+export declare function cmdDecision(argv: string[]): Promise<void>;