@skill-harness/cli 0.28.0 → 0.30.0

package/dist/serve.js

@@ -3,7 +3,7 @@ import { readFileSync, existsSync } from "node:fs";
 import { join, dirname } from "node:path";
 import { fileURLToPath } from "node:url";
 import { spawn } from "node:child_process";
-import { collectReport, renderReport, collectTrends, readResults, writeResults, applyOverride, preserveTranscript, findTranscriptFiles, ensureResultsGitignore, appendJournal, loadSpec, regradeScenario, refreshRubricHashes, findJudgeRawFiles, effectiveThreshold, scoreContextFor, isScoredMode, rebuildScenarioResult, mergeScenarioMetrics, carryRepObjectives, envFlag, recordedJudgeOrDefault, assertJudgeAllowed, } from "@skill-harness/core";
+import { collectReport, renderReport, collectTrends, readResults, writeResults, applyOverride, preserveTranscript, findTranscriptFiles, ensureResultsGitignore, appendJournal, loadSpec, regradeScenario, refreshRubricHashes, findJudgeRawFiles, effectiveThreshold, scoreContextFor, isScoredMode, rebuildScenarioResult, mergeScenarioMetrics, carryRepObjectives, envFlag, recordedJudgeOrDefault, assertJudgeAllowed, parseModelRef, resolveJudgeReferences, } from "@skill-harness/core";
 import { getAdapter } from "@skill-harness/adapters";
 const __dirname = dirname(fileURLToPath(import.meta.url));
 /** Locate assets/report.template.html relative to dist/ or src/. */
@@ -145,10 +145,16 @@ export async function serveReview(opts) {
                 }
                 const threshold = effectiveThreshold(prev, scenario);
                 try {
+                    // The prompt and refreshed rubric must name the same selected bytes,
+                    // even when a reference is edited while a multi-repetition judge runs.
+                    const specDir = dirname(specPath);
+                    const references = resolveJudgeReferences(scenario, specDir);
                     const rr = await regradeScenario({
                         runDir: column.runDir, spec, scenario, adapter, judge: resolvedRecordedJudge.judge,
-                        specDir: dirname(specPath), threshold, mode: results.mode,
+                        specDir, threshold, mode: results.mode, references,
                         expectedReps: prev.reps ?? 1,
+                        prior: prev, piVersion: results.harness_cli_version,
+                        ...(results.model.includes(":") ? { subject: parseModelRef(results.model) } : {}),
                     });
                     const merged = results.scenarios.map((s) => {
                         if (s.id !== body.scenarioId)
@@ -174,7 +180,7 @@ export async function serveReview(opts) {
                         // retired the staleness gate for any run re-judged from the UI. Carried,
                         // with the one `rubric:` key this re-judge actually applied refreshed —
                         // the same doctrine `grade` follows (see refreshRubricHashes).
-                        source_hashes: refreshRubricHashes(results.source_hashes, spec, [body.scenarioId]),
+                        source_hashes: refreshRubricHashes(results.source_hashes, spec, [body.scenarioId], specDir, { [body.scenarioId]: references }),
                         source_hash_roots: results.source_hash_roots,
                     }, scoreContextFor(results, spec));
                     ensureResultsGitignore(join(opts.skillDir, "tests", "results"));

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@skill-harness/cli",
-    "version": "0.28.0",
+    "version": "0.30.0",
     "description": "skill-harness CLI — run, grade, review, and lint agent-skill scenarios on the pi harness",
     "type": "module",
     "license": "MIT",
@@ -44,7 +44,7 @@
         "prepack": "node ../../scripts/release-pack.mjs --guard"
     },
     "dependencies": {
-        "@skill-harness/core": "0.28.0",
-        "@skill-harness/adapters": "0.28.0"
+        "@skill-harness/core": "0.30.0",
+        "@skill-harness/adapters": "0.30.0"
     }
 }