feat(ms-ai-architect): Spor D Steg B — skill-quality lifecycle-gate + score-cache (TDD)

B1: gateSkill(report, name) — pure post-mutation verdict (blocked/provisional/
improvements), wired into apply-skill-op.mjs to re-score the affected skill from
fresh disk after create/merge/sanitize-apply. K10 floor enforced immediately;
unjudged → provisional + nudge to re-run the judge pass. retire → skipped.

B2: buildScoreCache(result) — compact, deterministic cache shape + score-skill.mjs
--write [path] → data/skill-score-report.json (always whole-corpus). Gitignored
(derived/regenerable; avoids churn in the public repo). Consumed by Steg C surfacing.

8 new tests (tests/kb-eval/test-skill-score-gate.test.mjs). kb-eval 150 pass,
validate 239 pass. Live score unchanged: advisor 91, eng/gov/infra/sec 96, 0 < 90.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Kjell Tore Guttormsen 2026-06-23 18:26:29 +02:00
commit 92cd93771b
5 changed files with 294 additions and 3 deletions

4
.gitignore vendored
View file

@ -30,6 +30,10 @@ scripts/kb-update/data/*
# Generated skill-lifecycle detection report (Spor B / B1) — regenerated on demand,
# like the kb-update reports above. The detector script + curated inputs are tracked.
scripts/kb-eval/data/skill-lifecycle-report.json
# Generated skill-quality score cache (Spor D / STEG B) — derived from the corpus,
# regenerated by `score-skill.mjs --write` and the apply-skill-op gate. Consumed by
# the STEG C SessionStart surfacing; not committed (avoids churn in the public repo).
scripts/kb-eval/data/skill-score-report.json
.kb-backup/
.rollback-in-progress

View file

@ -24,7 +24,8 @@
import { readFileSync, readdirSync, existsSync } from 'node:fs';
import { dirname, join } from 'node:path';
import { fileURLToPath } from 'node:url';
import { splitFrontmatter, extractDescription } from './eval.mjs';
import { splitFrontmatter, extractDescription, buildReport } from './eval.mjs';
import { gateSkill } from './lib/skill-score.mjs';
import { loadTaxonomy } from '../kb-update/lib/taxonomy.mjs';
import { loadDecisions, listActions, actionKey } from '../kb-update/lib/decisions-io.mjs';
import {
@ -50,6 +51,8 @@ const OP_BY_VERB = {
create: CREATE_OPERATION,
};
const QUALITY_TARGET = 90; // corpus invariant: every skill stays >=90 (Spor D)
/** Read the SKILL.md descriptions from disk (read-only) — create's K10 siblings. */
function loadDescriptions() {
const out = {};
@ -137,6 +140,54 @@ function printResult(res, verb, label, apply) {
console.log(`\nStatus: APPLIED. Verifiser med 'git status' og re-kjør validate/kb-integrity/kb-eval.\n`);
}
/**
* The mutated skill that should be re-scored after an apply. create/sanitize keep
* a scorable skill; merge leaves the absorber; retire removes the skill entirely
* (nothing to score -> null, gate skipped). Spor D STEG B lifecycle-gate.
*/
function affectedSkill(op, targets) {
if (op === CREATE_OPERATION) return targets.name;
if (op === MERGE_OPERATION) return targets.absorber;
if (op === SANITIZE_OPERATION) return targets.skill;
return null; // RETIRE: skill removed
}
/**
* Re-score the affected skill from fresh disk after a successful apply and surface
* a quality verdict. The deterministic floor (K10) is enforced immediately as a
* loud warning; an unjudged skill is flagged provisional with a nudge to re-run
* the judge pass (the K1 floor cannot be enforced from the cache alone).
*/
function printQualityGate(op, targets) {
const affected = affectedSkill(op, targets);
if (!affected) {
console.log(`\nKvalitetsgate: ingen gjenværende skill å re-score (retire). Hopper over.`);
return;
}
const v = gateSkill(buildReport(), affected, { target: QUALITY_TARGET });
if (!v.found) {
console.log(`\nKvalitetsgate: fant ikke '${affected}' på disk for re-scoring (hoppet over).`);
return;
}
const tags = [];
if (v.floored) tags.push('floored');
if (v.provisional) tags.push('provisional');
const tagStr = tags.length ? ` [${tags.join(', ')}]` : '';
if (v.blocked) {
console.log(`\n⚠ KVALITETSGATE: '${affected}' scorer ${v.score}/100 — UNDER MÅL ${QUALITY_TARGET} %${tagStr}.`);
for (const imp of (v.improvements || []).slice(0, 3)) {
const floor = imp.floor ? ' ⚑gulv' : '';
console.log(`${imp.key} ${imp.label}${floor} (${imp.detail}) [-${imp.pointsLost.toFixed(2)}] → ${imp.fix}`);
}
console.log(` Rett før commit, eller dokumenter avviket eksplisitt.`);
} else {
console.log(`\nKvalitetsgate: '${affected}' scorer ${v.score}/100 — OK (≥${QUALITY_TARGET} %)${tagStr}.`);
}
if (v.provisional) {
console.log(` judge-cache mangler/utdatert for '${affected}' → K1-gulv ikke håndhevet. Kjør judge-passet (scripts/kb-eval/judge-prompt.md) og re-merge data/judge-results.json, deretter re-score.`);
}
}
function main() {
const args = process.argv.slice(2);
const verb = args[0];
@ -209,6 +260,11 @@ function main() {
promptSet: isCreate ? loadPromptSet() : undefined,
});
printResult(res, verb, label, apply);
// STEG B: re-score the affected skill from fresh disk after a real mutation.
// Preview never mutates -> nothing to re-score (create's projected skill isn't
// on disk), so the gate runs only on applied mutations.
if (res.applied) printQualityGate(op, targets);
}
if (process.argv[1] && fileURLToPath(import.meta.url) === process.argv[1]) {

View file

@ -250,6 +250,67 @@ export function scoreReport(report, opts = {}) {
return { target, scored, below };
}
/**
* Re-score ONE affected skill out of a freshly built corpus report and return a
* lifecycle-gate verdict (Spor D, STEG B). Used by apply-skill-op.mjs after a
* create/merge/sanitize mutation: the whole corpus is needed for K10 sibling
* overlap, but only the affected skill is judged. The deterministic floor (K10)
* is enforced immediately via `blocked`; an unjudged skill is `provisional`
* (the K1 floor cannot be enforced -> the caller nudges to re-run the judge).
* @param {{skills: object[]}} report buildReport() output (fresh disk)
* @param {string} skillName the mutated skill to re-score
* @param {{target?: number, floorCap?: number}} [opts]
* @returns {object} { name, found, score?, meetsTarget?, blocked?, provisional?, judged?, floored?, improvements? }
*/
export function gateSkill(report, skillName, opts = {}) {
const target = opts.target ?? TARGET;
const { scored } = scoreReport(report, { ...opts, target });
const s = scored.find((x) => x.name === skillName);
if (!s) return { name: skillName, found: false };
return {
name: skillName,
found: true,
score: s.score,
meetsTarget: s.meetsTarget,
blocked: !s.meetsTarget,
provisional: s.provisional,
judged: s.judged,
floored: s.floored,
improvements: s.improvements,
};
}
/**
* Project a scoreReport() result into the compact, PURE cache shape persisted as
* data/skill-score-report.json and consumed by the STEG C SessionStart surfacing
* (which must read the cache, never score live). Drops the verbose criteria[] and
* keeps only what the one-liner needs; `below` is a list of names. Deterministic
* given an injected `generatedAt` (no Date.now in the pure layer).
* @param {{target: number, scored: object[]}} result scoreReport() output
* @param {{generatedAt?: string|null}} [opts]
* @returns {{generatedAt: string|null, target: number, scored: object[], below: string[]}}
*/
export function buildScoreCache(result, opts = {}) {
const target = result?.target ?? TARGET;
const scored = (result?.scored || []).map((s) => ({
name: s.name,
score: s.score,
meetsTarget: s.meetsTarget,
provisional: s.provisional,
floored: s.floored,
judged: s.judged,
topFix: s.improvements?.[0]
? { key: s.improvements[0].key, label: s.improvements[0].label, fix: s.improvements[0].fix }
: null,
}));
return {
generatedAt: opts.generatedAt ?? null,
target,
scored,
below: scored.filter((s) => !s.meetsTarget).map((s) => s.name),
};
}
/**
* Render a human-readable quality report (pure). Each skill shows its score and
* status; below-target skills list their sorted improvements (biggest weighted

View file

@ -13,16 +13,25 @@
// node scripts/kb-eval/score-skill.mjs --skill <name> # scope output to one skill
// # (still scores the full corpus
// # so K10 sibling-overlap is correct)
// node scripts/kb-eval/score-skill.mjs --write [path] # persist the whole-corpus cache
// # (default: data/skill-score-report.json;
// # consumed by the STEG C surfacing)
//
// K10 requires every sibling description, so even single-skill (incremental)
// scoring evaluates the whole corpus and then filters the reported set — this is
// the building block for the lifecycle-gate in apply-skill-op.mjs.
// the building block for the lifecycle-gate in apply-skill-op.mjs. The --write
// cache is ALWAYS whole-corpus (surfacing needs every skill), independent of --skill.
//
// Zero dependencies. Pure scoring lives in lib/skill-score.mjs.
import { writeFileSync } from 'node:fs';
import { dirname, join } from 'node:path';
import { fileURLToPath } from 'node:url';
import { buildReport } from './eval.mjs';
import { scoreReport, formatScoreReport, TARGET } from './lib/skill-score.mjs';
import { scoreReport, formatScoreReport, buildScoreCache, TARGET } from './lib/skill-score.mjs';
const __dirname = dirname(fileURLToPath(import.meta.url));
const DEFAULT_CACHE = join(__dirname, 'data', 'skill-score-report.json');
function argValue(args, flag) {
const i = args.indexOf(flag);
@ -36,9 +45,24 @@ function main() {
const gate = gateRaw !== null ? Number(gateRaw) : null;
const skillFilter = argValue(args, '--skill');
// --write [path]: bare flag uses the default cache path; an explicit non-flag
// argument overrides it (the tests write to a temp path to avoid clobbering).
const writeIdx = args.indexOf('--write');
const doWrite = writeIdx !== -1;
const writePath = doWrite && args[writeIdx + 1] && !args[writeIdx + 1].startsWith('--')
? args[writeIdx + 1]
: DEFAULT_CACHE;
const target = Number.isFinite(gate) ? gate : TARGET;
const full = scoreReport(buildReport(), { target });
// The cache is whole-corpus regardless of --skill (surfacing needs every skill).
if (doWrite) {
const cache = buildScoreCache(full, { generatedAt: new Date().toISOString().slice(0, 10) });
writeFileSync(writePath, JSON.stringify(cache, null, 2) + '\n');
process.stdout.write(`✓ Skrev skill-score-cache → ${writePath} (${cache.scored.length} skills, ${cache.below.length} under mål).\n`);
}
// Scope the reported/gated set to one skill if requested (scoring already ran
// over the whole corpus, so K10 is computed correctly).
let scored = full.scored;

View file

@ -0,0 +1,146 @@
// tests/kb-eval/test-skill-score-gate.test.mjs
// Tests for the Spor D, STEG B lifecycle-gate building blocks:
// - lib/skill-score.mjs gateSkill(report, name): the PURE post-mutation verdict
// used by apply-skill-op.mjs to re-score the affected skill incrementally.
// The deterministic floor (K10) is always enforced; an unjudged skill is
// flagged `provisional` (the K1 floor cannot be enforced -> nudge to re-judge).
// - lib/skill-score.mjs buildScoreCache(result): the compact, PURE cache shape
// persisted as data/skill-score-report.json and consumed by STEG C surfacing.
// - the score-skill.mjs CLI --write flag: persists the whole-corpus cache.
import { test } from 'node:test';
import assert from 'node:assert/strict';
import { execFileSync } from 'node:child_process';
import { readFileSync, rmSync } from 'node:fs';
import { fileURLToPath } from 'node:url';
import { dirname, join } from 'node:path';
import { tmpdir } from 'node:os';
import { gateSkill, buildScoreCache, scoreReport } from '../../scripts/kb-eval/lib/skill-score.mjs';
const __dirname = dirname(fileURLToPath(import.meta.url));
const CLI = join(__dirname, '..', '..', 'scripts', 'kb-eval', 'score-skill.mjs');
// A synthetic skill whose deterministic criteria all pass; K10 + judge are
// parametrizable so a single helper drives the meets/blocked/provisional cases.
function skill(name, { k10Pass = true, k10Combined = 1, worst = 'other', judge = null } = {}) {
return {
name,
deterministic: {
K2_descriptionFormat: { pass: true },
K3_bodyLength: { bodyLines: 100, pass: true },
K5_progressiveDisclosure: { namedRatio: 0.5, pass: true },
K6_routingTable: { pass: true },
refCountConsistency: { consistent: true },
K10_siblingScopeOverlap: { maxCombined: k10Combined, pass: k10Pass, threshold: 7, worstSibling: worst },
N1_nameValidity: { pass: true },
N2_descriptionLength: { pass: true },
N3_refNestingDepth: { pass: true },
N4_refToc: { ratio: 1, pass: true, withToc: 1, largeFiles: 1 },
N5_forwardSlashPaths: { pass: true },
},
judge,
};
}
const FULL_JUDGE = {
K1_triggerPrecision: { precision: 1, pass: true },
K4_noDuplication: { score: 5, pass: true },
K9_noTimeSensitive: { pass: true, findings: [] },
K7_imperativeStyle: { ratio: 1, pass: true },
K8_sourceCitation: { ratio: 1, pass: true },
};
test('gateSkill — affected skill that meets the target is not blocked', () => {
const report = { skills: [skill('alpha', { judge: FULL_JUDGE }), skill('beta', { judge: FULL_JUDGE })] };
const v = gateSkill(report, 'alpha', { target: 90 });
assert.equal(v.found, true);
assert.equal(v.meetsTarget, true);
assert.equal(v.blocked, false);
assert.equal(v.provisional, false);
});
test('gateSkill — a failing K10 floor blocks the affected skill (deterministic, enforced immediately)', () => {
const report = {
skills: [skill('good', { judge: FULL_JUDGE }), skill('bad', { k10Pass: false, k10Combined: 9, worst: 'good', judge: FULL_JUDGE })],
};
const v = gateSkill(report, 'bad', { target: 90 });
assert.equal(v.found, true);
assert.equal(v.meetsTarget, false);
assert.equal(v.blocked, true);
assert.ok(v.score <= 89, `K10 floor must cap below 90, got ${v.score}`);
assert.ok(v.improvements.some((i) => i.key === 'K10'), 'K10 must surface as an improvement');
});
test('gateSkill — an unjudged skill is provisional but the deterministic floor still blocks it', () => {
// judge:null -> K1 floor cannot be enforced (provisional), but K10 (deterministic) still blocks.
const report = { skills: [skill('good'), skill('bad', { k10Pass: false, k10Combined: 9, worst: 'good' })] };
const blocked = gateSkill(report, 'bad', { target: 90 });
assert.equal(blocked.provisional, true, 'missing judge -> provisional');
assert.equal(blocked.blocked, true, 'deterministic K10 floor still blocks when provisional');
const ok = gateSkill(report, 'good', { target: 90 });
assert.equal(ok.provisional, true, 'missing judge -> provisional even when passing');
assert.equal(ok.meetsTarget, true, 'deterministic-only score still passes');
});
test('gateSkill — an unknown affected skill returns found:false (no crash)', () => {
const report = { skills: [skill('alpha', { judge: FULL_JUDGE })] };
const v = gateSkill(report, 'does-not-exist', { target: 90 });
assert.equal(v.found, false);
assert.equal(v.name, 'does-not-exist');
});
test('buildScoreCache — compact whole-corpus shape with below[] as names', () => {
const report = {
skills: [skill('good', { judge: FULL_JUDGE }), skill('bad', { k10Pass: false, k10Combined: 9, worst: 'good', judge: FULL_JUDGE })],
};
const result = scoreReport(report, { target: 90 });
const cache = buildScoreCache(result, { generatedAt: '2026-06-23' });
assert.equal(cache.target, 90);
assert.equal(cache.generatedAt, '2026-06-23');
assert.equal(cache.scored.length, 2);
// compact per-skill entry: the fields STEG C surfacing needs, not the verbose criteria[]
const good = cache.scored.find((s) => s.name === 'good');
assert.equal(typeof good.score, 'number');
assert.equal(typeof good.meetsTarget, 'boolean');
assert.ok(!('criteria' in good), 'cache must stay compact (no criteria[])');
// below[] is a list of NAMES (cheap for the SessionStart one-liner)
assert.deepEqual(cache.below, ['bad']);
});
test('buildScoreCache — deterministic given an injected generatedAt (no Date.now)', () => {
const report = { skills: [skill('alpha', { judge: FULL_JUDGE })] };
const result = scoreReport(report, { target: 90 });
const a = buildScoreCache(result, { generatedAt: '2026-01-01' });
const b = buildScoreCache(result, { generatedAt: '2026-01-01' });
assert.deepEqual(a, b);
});
test('CLI --write <path> — persists the whole-corpus cache (all 5 skills) to disk', () => {
const out = join(tmpdir(), `skill-score-cache-${process.pid}.json`);
try {
execFileSync('node', [CLI, '--write', out], { encoding: 'utf8' });
const cache = JSON.parse(readFileSync(out, 'utf8'));
assert.equal(cache.target, 90);
assert.equal(cache.scored.length, 5, 'cache must cover the whole corpus, not a filtered view');
assert.ok(Array.isArray(cache.below), 'below[] present');
for (const s of cache.scored) {
assert.equal(typeof s.name, 'string');
assert.equal(typeof s.score, 'number');
assert.equal(typeof s.meetsTarget, 'boolean');
}
} finally {
rmSync(out, { force: true });
}
});
test('CLI --write with --skill still writes the whole corpus (cache is corpus-wide)', () => {
const out = join(tmpdir(), `skill-score-cache-corpus-${process.pid}.json`);
try {
execFileSync('node', [CLI, '--skill', 'ms-ai-advisor', '--write', out], { encoding: 'utf8' });
const cache = JSON.parse(readFileSync(out, 'utf8'));
assert.equal(cache.scored.length, 5, '--skill scopes stdout, not the persisted cache');
} finally {
rmSync(out, { force: true });
}
});