@skill-harness/cli 0.10.0 → 0.11.0

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, restampSkill, resolveAdjudicationJudges, adjudicateRun, judgeResemblesSubject, computeCoverage, formatCoverage, selectAffected, formatAffected, gitDiff, exec, HARNESS_VERSION, defaultJudge, assertJudgeAllowed, assertNotDowngraded, downgradeWarning, isScoredMode, runTrajectoryMutationSelfTest, } 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, isScoredMode, runTrajectoryMutationSelfTest, resolveArm, } from "@skill-harness/core";
 import { getAdapter } from "@skill-harness/adapters";
 import { serveReview } from "./serve.js";
 import { runCompareCommand } from "./compare.js";
@@ -151,10 +151,15 @@ export async function cmdRun(args) {
         throw new Error(`harness \`${harnessName}\` is not on PATH`);
     const mode = flagStr(args, "mode", "green") || "green";
     const canary = flagBool(args, "canary");
+    const structured = flagBool(args, "structured");
     const label = flagStr(args, "label") || null;
     const parallel = Math.max(1, Number(flagStr(args, "parallel", "1")) || 1);
     const { reps, passThreshold } = parseRunTuning(args);
     const modelTokens = resolveModels(args);
+    // Resolved ONCE, before the skill loop: a typo'd arm name fails before any
+    // token is spent on any skill.
+    const armName = flagStr(args, "arm") || null;
+    const arm = resolveArm(root, armName);
     const skills = target === "all"
         ? discover(root).filter((s) => s.hasSpec)
         : [resolveSkill(root, target)];
@@ -201,6 +206,9 @@ export async function cmdRun(args) {
                 passThreshold,
                 only,
                 canary,
+                structured,
+                arm,
+                skillsRoot: root,
                 onProgress: (m) => console.log(m),
             });
             summaries.push(summary);
@@ -858,6 +866,9 @@ export function help() {
                      [--affected --base <git-ref>]  run only the scenarios a change could touch (partial; never SHIPs)
                      [--mode red|green|force] [--judge prov:model] [--harness pi] [--label name] [--parallel N] [--reps N] [--pass-threshold T]
                      [--canary]  green only: spend ONE probe proving the skill reached the model, and abort the run if it did not
+                     [--structured]  record subject tokens, cost and wall time (needs pi --mode json)
+                     [--arm <name>]  measure under a named arm from <skills-root>/tests/arms.yaml
+                                     (loads its extensions, seeds pi-daddy definitions, tags the run dir)
   compare <skill|all> --reference <git-ref-or-skills-root> --candidate <skills-root>
                      [--model prov:model ...] [--mode green|force] [--judge prov:model] [--reps N] [--only IDs | --affected --base ref]
                      [--output dir] [--max-subject-token-increase R] [--max-wall-time-increase R] [--max-tool-call-increase R]

package/dist/serve.js

@@ -144,6 +144,13 @@ export async function serveReview(opts) {
                         partial: results.partial,
                         // Provenance survives a UI re-judge, same as it does through `grade`.
                         harness_cli_version: results.harness_cli_version, delivery_canary: results.delivery_canary,
+                        // The arm is provenance of the MEASUREMENT, not of this rewrite, and it is the
+                        // only record that a `+<arm>` run actually delegated: rebuilding the draft
+                        // field-by-field without it silently deleted `definitions`/`ledger_events`
+                        // from any arm run that was ever re-graded, leaving a record
+                        // indistinguishable from a vacuous arm. Same reason `harness_cli_version`,
+                        // `delivery_canary` and `source_hashes` are carried here.
+                        arm: results.arm,
                         // Recorded hashes were being dropped here entirely, which silently
                         // retired the staleness gate for any run re-judged from the UI. Carried,
                         // with the one `rubric:` key this re-judge actually applied refreshed —

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@skill-harness/cli",
-    "version": "0.10.0",
+    "version": "0.11.0",
     "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": "0.10.0",
-        "@skill-harness/adapters": "0.10.0"
+        "@skill-harness/core": "0.11.0",
+        "@skill-harness/adapters": "0.11.0"
     }
 }