@skill-harness/cli 0.7.0 → 0.8.0

package/dist/cli.js

@@ -3,7 +3,7 @@ import { readFileSync, existsSync, mkdirSync, writeFileSync, mkdtempSync, rmSync
 import { load as yamlLoad } from "js-yaml";
 import { basename, dirname, join, resolve, relative } from "node:path";
 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, collectStability, boundaryCells, stabilityNote, PATH_LEGEND, resolveAdjudicationJudges, adjudicateRun, judgeResemblesSubject, computeCoverage, formatCoverage, selectAffected, formatAffected, gitDiff, exec, HARNESS_VERSION, defaultJudge, assertJudgeAllowed, assertNotDowngraded, downgradeWarning, } from "@skill-harness/core";
+import { discover, resolveSkill, loadSpec, parseSpec, appendScenario, parseModelRef, runSkillModel, formatScorecard, readResults, regradeRun, lintSkill, failsGate, renderTemplateSpec, isTemplateSpec, renderDraftSpec, buildSuggestPrompt, parseSuggestDraft, rescoreRun, regateRun, specPathForRunDir, collectLift, collectStability, boundaryCells, stabilityNote, PATH_LEGEND, restampSkill, resolveAdjudicationJudges, adjudicateRun, judgeResemblesSubject, computeCoverage, formatCoverage, selectAffected, formatAffected, gitDiff, exec, HARNESS_VERSION, defaultJudge, assertJudgeAllowed, assertNotDowngraded, downgradeWarning, } from "@skill-harness/core";
 import { getAdapter } from "@skill-harness/adapters";
 import { serveReview } from "./serve.js";
 const DEFAULT_MODEL = "fireworks:accounts/fireworks/models/deepseek-v4-pro";
@@ -358,6 +358,49 @@ export async function cmdRegate(args, adapterOverride) {
  * spec; it is a statement about how much one run of that cell is worth. Making it a
  * gate would turn "this needs more reps" into "your build is broken".
  */
+/**
+ * Upgrade committed runs to the model-visible skill digest — free, offline, no re-runs.
+ *
+ * Run it once on a board that lints clean. Every record whose SKILL.md (or agent file)
+ * still matches the bytes it measured gains a digest of that same file's model-visible
+ * text, after which editing frontmatter the model never receives — `allowed-tools:`, a
+ * tool ceiling — stops demanding a paid re-run. Records whose file has already moved are
+ * left alone: nothing in a one-way hash can say whether that edit touched the body, and
+ * inventing freshness is the one thing this gate must never do.
+ */
+async function cmdRestamp(args) {
+    const root = flagStr(args, "skills", process.cwd());
+    const target = args._[0] ?? "all";
+    const skills = target === "all" ? discover(root).filter((s) => s.hasSpec) : [resolveSkill(root, target)];
+    if (skills.length === 0)
+        throw new Error(`no skills with a spec under ${root}`);
+    let runs = 0;
+    let upgraded = 0;
+    let unprovable = 0;
+    let unchanged = 0;
+    let partial = 0;
+    for (const skill of skills) {
+        const r = restampSkill(skill.dir, { from: flagStr(args, "from") });
+        runs += r.runs;
+        upgraded += r.upgraded;
+        unprovable += r.unprovable;
+        unchanged += r.unchanged;
+        partial += r.partial;
+        console.log(`\n${skill.name}: ${r.upgraded} upgraded, ${r.unprovable} left alone, ${r.unchanged} already current (${r.runs} run(s))`);
+        for (const a of r.added)
+            console.log(`  + ${a}`);
+        if (r.unprovable > 0) {
+            console.log(`  ${r.unprovable} left alone — a document they measured has already moved, so no digest of it can be proven; \`lint\` names the remedy`);
+        }
+        if (r.partial > 0) {
+            console.log(`  ${r.partial} of the upgraded still carry a document that could not be proven`);
+        }
+    }
+    // The three buckets sum to the records examined, on purpose: a total that does not add
+    // up reads like the command skipped something rather than like there being nothing to do.
+    console.log(`\n${runs} run(s) examined: ${upgraded} upgraded, ${unprovable} left alone, ${unchanged} already current.` +
+        `${partial > 0 ? ` (${partial} upgraded only in part.)` : ""} No models were called.`);
+}
 async function cmdStability(args) {
     const root = flagStr(args, "skills", process.cwd());
     const target = args._[0] ?? "all";
@@ -713,6 +756,7 @@ export function help() {
                        ship-deciding). OFF by default; prints the exact MAX extra call count first.
   rescore <run-dir>...                          re-score saved reps vs current spec thresholds (free)
   regate <run-dir>...  [--judge prov:model]     re-evaluate diff needles against the saved diffs (free; judges only reps whose gate flipped)
+  restamp <skill|all> --skills <root> [--from <git-ref>]   record the model-visible skill digest on runs that still match (free, offline; one-time migration)
   stability <skill|all> --skills <root> [--window N] [--all]  run-over-run verdict flips per scenario (free, offline)
   review <skill>     --skills <root> [--port N] serve the interactive review UI
   add-test <skill>   --skills <root> --id ID --title T --turn ... --check ... [--critical] [--mode seeded --fixture path]
@@ -740,6 +784,7 @@ export async function main(argv) {
         case "grade": return cmdGrade(args);
         case "rescore": return cmdRescore(args);
         case "regate": return cmdRegate(args);
+        case "restamp": return cmdRestamp(args);
         case "stability": return cmdStability(args);
         case "review": return cmdReview(args);
         case "add-test": return cmdAddTest(args);

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@skill-harness/cli",
-    "version": "0.7.0",
+    "version": "0.8.0",
     "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": "rm -rf ./assets && mkdir -p ./assets && cp ../../assets/report.* ./assets/ && cp ../../LICENSE ./LICENSE"
     },
     "dependencies": {
-        "@skill-harness/core": "0.7.0",
-        "@skill-harness/adapters": "0.7.0"
+        "@skill-harness/core": "0.8.0",
+        "@skill-harness/adapters": "0.8.0"
     }
 }