ms-ai-architect/tests/kb-eval/test-skill-ops-merge-apply.test.mjs
Kjell Tore Guttormsen de3a9a483b feat(ms-ai-architect): Sesjon 19 — B3 merge-apply (cross-skill + taksonomi-persist) [skip-docs]
Merge-apply = den tyngste destruktive op-en: eksekver en operatør-godkjent
merge_skills-entry → faktisk cross-skill skills/-mutasjon. Tre mekanismer
som sanitize/retire ikke trengte: kuratert relokasjon (ikke arkiv),
taksonomi-persistering, retire-av-absorbert.

- taxonomy.mjs (ny persist): saveTaxonomy (atomisk .tmp+rename, speiler
  saveDecisions) + ren applyCategoryReassignments (repointer KUN der eier
  === from; absent/allerede-flyttet = no-op). Lag-0 er normalt read-only;
  ENESTE writer = gated merge-apply.
- applyApprovedAction merge-gren (skill-ops.mjs): re-plan fersk → revalidér
  (kollisjon → guardrail.ok=false → abort) → flytt absorbed-refs → absorber
  (moveFile; flyttingen ER bevaringen set-equality beviser) → persister
  taksonomi-repoint fra FERSK plan (drift-sikker) → arkivér absorbed-SKILL.md
  FØR rmdir → flipp ledger approved→applied. Description-forsoning forblir
  MANUELL (planner-kontrakt: apply rører aldri absorber-SKILL.md).
  archiveMove→moveFile (generisk: arkiv + kuratert flytt).
- CLI apply-skill-op.mjs merge <absorber> <absorbed> — lookup via actionKey
  (merge_skills:<sortert par>); default PREVIEW, --apply = dobbel-gate.

TDD, tmpdir-fixtures: 0 ekte skills/-mutasjon. Ny test-skill-ops-merge-apply
(7) + 2 taksonomi-tester; fjernet foreldet merge-throws fra S18-fila. Tester:
kb-eval 78→84, kb-update 137→139; validate 239 · kb-integrity 192/192 uendret.
eval --json deterministisk uendret (K10 eng+infra FAIL, øvrige PASS).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-20 20:54:13 +02:00

225 lines
11 KiB
JavaScript

// tests/kb-eval/test-skill-ops-merge-apply.test.mjs
// Spor B / B3 (Sesjon 19): merge-apply — the HEAVIEST destructive op. An
// operator-approved (status:approved) merge_skills entry is executed into a real
// skills/ mutation: the absorbed skill's curated references MOVE under the
// absorber, the taxonomy's category_skill ownership is REPOINTED + persisted,
// and the now-empty absorbed shell is archived (SKILL.md) + its dir removed.
//
// Binding contract proven here:
// 1. Curated preservation: every absorbed reference survives under the absorber
// (the guardrail's set-equality made real on disk — a MOVE, not an archive).
// 2. Taxonomy persistence: categories owned by the absorbed skill are repointed
// to the absorber and written back to domain-taxonomy.json (drift-safe: the
// reassignments are recomputed from a FRESH plan at apply time).
// 3. Archive-before-delete for the absorbed shell: SKILL.md lands under archive/
// before the skills/ dir is removed; absorber + siblings untouched.
// 4. Refuses (zero mutation, ledger untouched) on: collision guardrail-fail,
// disk drift since approval, or a non-approved entry.
//
// Every scenario runs against a tmpdir fixture — NEVER the real skills/ tree.
import { test } from 'node:test';
import assert from 'node:assert/strict';
import { mkdtempSync, rmSync, mkdirSync, writeFileSync, existsSync } from 'node:fs';
import { tmpdir } from 'node:os';
import { join } from 'node:path';
import {
applyApprovedAction,
runMergePlan,
} from '../../scripts/kb-eval/lib/skill-ops.mjs';
import { loadTaxonomy, getCategorySkill } from '../../scripts/kb-update/lib/taxonomy.mjs';
import {
loadDecisions,
saveDecisions,
recordAction,
actionKey,
} from '../../scripts/kb-update/lib/decisions-io.mjs';
// ---------------------------------------------------------------------------
// Fixture: absorber (skill-a) + absorbed (skill-b) under one root, plus a
// domain-taxonomy.json in dataDir whose category_skill assigns cat3 to the
// absorbed skill (so the merge must repoint + persist it). archive/ lands at the
// same root so rename stays on one filesystem. `extraBRef` lets a test inject a
// colliding reference-relative path (present under BOTH skills).
// ---------------------------------------------------------------------------
function withFixture(fn, { extraBRef } = {}) {
const root = mkdtempSync(join(tmpdir(), 'mergeapply-'));
const skillsDir = join(root, 'skills');
const dataDir = join(root, 'data');
const mkRef = (skill, cat, file, body) => {
const dir = join(skillsDir, skill, 'references', cat);
mkdirSync(dir, { recursive: true });
writeFileSync(join(dir, file), body);
};
const mkSkillMd = (skill, body) => {
mkdirSync(join(skillsDir, skill), { recursive: true });
writeFileSync(join(skillsDir, skill, 'SKILL.md'), body);
};
// absorber: 2 curated refs under cat1
mkRef('skill-a', 'cat1', 'a1.md', '# A1\n');
mkRef('skill-a', 'cat1', 'a2.md', '# A2\n');
mkSkillMd('skill-a', '# skill-a (absorber)\n');
// absorbed: 1 curated ref under cat3
mkRef('skill-b', 'cat3', 'b1.md', '# B1\n');
mkSkillMd('skill-b', '# skill-b (absorbed)\n');
if (extraBRef) mkRef('skill-b', extraBRef.cat, extraBRef.file, extraBRef.body ?? '# collide\n');
// taxonomy: cat1 owned by absorber, cat3 owned by absorbed
mkdirSync(dataDir, { recursive: true });
writeFileSync(
join(dataDir, 'domain-taxonomy.json'),
JSON.stringify({ version: 1, category_skill: { cat1: 'skill-a', cat3: 'skill-b' } }, null, 2) + '\n',
);
try {
return fn({ root, skillsDir, dataDir });
} finally {
rmSync(root, { recursive: true, force: true });
}
}
const CATEGORY_SKILL = { cat1: 'skill-a', cat3: 'skill-b' };
// Seed an operator-approved merge entry by snapshotting a dry-run's guardrail —
// exactly what the operator gate would write after reviewing the dry-run.
function approveMerge(absorber, absorbed, { skillsDir, dataDir, status = 'approved' }) {
const { entry } = runMergePlan(absorber, absorbed, {
skillsDir,
dataDir,
write: false,
decided_at: '2026-06-20',
categorySkill: CATEGORY_SKILL,
});
const seeded = { ...entry, status };
saveDecisions(recordAction(loadDecisions(dataDir), seeded), dataDir);
return seeded;
}
const applyOpts = (root, skillsDir, dataDir, apply) => ({
pluginRoot: root,
skillsDir,
dataDir,
taxonomyDataDir: dataDir,
categorySkill: CATEGORY_SKILL,
apply,
decided_at: apply ? '2026-06-21' : null,
});
// ===========================================================================
// (A) Happy path — refs moved, taxonomy persisted, absorbed archived+removed
// ===========================================================================
test('merge-apply — absorbed refs move under absorber; absorber + siblings untouched', () => {
withFixture(({ root, skillsDir, dataDir }) => {
const approved = approveMerge('skill-a', 'skill-b', { skillsDir, dataDir });
const res = applyApprovedAction(approved, applyOpts(root, skillsDir, dataDir, true));
assert.equal(res.applied, true);
assert.equal(res.revalidate.ok, true);
// Absorbed ref relocated under the absorber (preservation = a MOVE).
assert.ok(existsSync(join(skillsDir, 'skill-a/references/cat3/b1.md')), 'absorbed ref now under absorber');
// Absorber's own refs untouched.
assert.ok(existsSync(join(skillsDir, 'skill-a/references/cat1/a1.md')), 'absorber ref a1 preserved');
assert.ok(existsSync(join(skillsDir, 'skill-a/references/cat1/a2.md')), 'absorber ref a2 preserved');
});
});
test('merge-apply — taxonomy ownership repointed to absorber and persisted to disk', () => {
withFixture(({ root, skillsDir, dataDir }) => {
const approved = approveMerge('skill-a', 'skill-b', { skillsDir, dataDir });
const res = applyApprovedAction(approved, applyOpts(root, skillsDir, dataDir, true));
assert.equal(res.applied, true);
assert.equal(res.report.taxonomyReassignments, 1, 'one category (cat3) repointed');
const tax = loadTaxonomy(dataDir);
assert.equal(getCategorySkill(tax, 'cat3'), 'skill-a', 'absorbed category now owned by absorber');
assert.equal(getCategorySkill(tax, 'cat1'), 'skill-a', 'absorber category unchanged');
});
});
test('merge-apply — absorbed shell archived (SKILL.md), then its dir removed; ledger applied', () => {
withFixture(({ root, skillsDir, dataDir }) => {
const approved = approveMerge('skill-a', 'skill-b', { skillsDir, dataDir });
const res = applyApprovedAction(approved, applyOpts(root, skillsDir, dataDir, true));
assert.equal(res.applied, true);
// Archive-before-delete: the absorbed SKILL.md is recoverable under archive/.
assert.ok(existsSync(join(root, 'archive/skills/skill-b/SKILL.md')), 'absorbed SKILL.md archived');
// The absorbed skill dir is gone entirely.
assert.equal(existsSync(join(skillsDir, 'skill-b')), false, 'absorbed skill dir removed');
// Ledger flipped approved -> applied.
assert.equal(loadDecisions(dataDir).actions[actionKey(approved)].status, 'applied');
});
});
// ===========================================================================
// (B) Safety — collision, drift, preview, non-approved
// ===========================================================================
test('merge-apply — a colliding reference path fails the guardrail and applies nothing', () => {
withFixture(
({ root, skillsDir, dataDir }) => {
// skill-b now also has cat1/a1.md — the same references-relative path as the
// absorber. The merge guardrail is ok=false (would clobber on the target).
const approved = approveMerge('skill-a', 'skill-b', { skillsDir, dataDir });
assert.equal(approved.guardrail.ok, false, 'precondition: collision makes the dry-run guardrail fail');
const res = applyApprovedAction(approved, applyOpts(root, skillsDir, dataDir, true));
assert.equal(res.applied, false, 'a failing guardrail must never reach the filesystem');
assert.equal(res.revalidate.freshOk, false);
assert.ok(existsSync(join(skillsDir, 'skill-b')), 'absorbed skill untouched');
assert.equal(existsSync(join(root, 'archive')), false, 'nothing archived');
assert.equal(loadDecisions(dataDir).actions[actionKey(approved)].status, 'approved', 'ledger left at approved');
},
{ extraBRef: { cat: 'cat1', file: 'a1.md' } },
);
});
test('merge-apply — DRIFT after approval aborts with zero mutation and an untouched ledger', () => {
withFixture(({ root, skillsDir, dataDir }) => {
const approved = approveMerge('skill-a', 'skill-b', { skillsDir, dataDir });
// Disk drifts after approval: a new ref appears under the absorbed skill, so a
// fresh plan's guardrail counts differ from the approved snapshot.
mkdirSync(join(skillsDir, 'skill-b/references/cat3'), { recursive: true });
writeFileSync(join(skillsDir, 'skill-b/references/cat3/b2.md'), '# B2 (new)\n');
const res = applyApprovedAction(approved, applyOpts(root, skillsDir, dataDir, true));
assert.equal(res.applied, false, 'drift must abort the apply');
assert.equal(res.revalidate.drift, true);
assert.match(res.revalidate.reason, /drift/i);
assert.ok(existsSync(join(skillsDir, 'skill-b')), 'absorbed skill untouched');
assert.equal(existsSync(join(root, 'archive')), false, 'nothing archived on a drift abort');
assert.equal(loadDecisions(dataDir).actions[actionKey(approved)].status, 'approved');
// Taxonomy untouched too.
assert.equal(getCategorySkill(loadTaxonomy(dataDir), 'cat3'), 'skill-b');
});
});
test('merge-apply — preview (apply:false) mutates nothing; ledger + taxonomy unchanged', () => {
withFixture(({ root, skillsDir, dataDir }) => {
const approved = approveMerge('skill-a', 'skill-b', { skillsDir, dataDir });
const res = applyApprovedAction(approved, applyOpts(root, skillsDir, dataDir, false));
assert.equal(res.applied, false);
assert.equal(res.preview, true);
assert.equal(res.revalidate.ok, true, 'disk unchanged -> revalidation would pass');
assert.ok(existsSync(join(skillsDir, 'skill-b/references/cat3/b1.md')), 'absorbed ref still in place (preview only)');
assert.equal(existsSync(join(root, 'archive')), false, 'preview archives nothing');
assert.equal(getCategorySkill(loadTaxonomy(dataDir), 'cat3'), 'skill-b', 'taxonomy unchanged in preview');
assert.equal(loadDecisions(dataDir).actions[actionKey(approved)].status, 'approved');
});
});
test('merge-apply — refuses a non-approved (pending) entry with zero mutation', () => {
withFixture(({ root, skillsDir, dataDir }) => {
const pending = approveMerge('skill-a', 'skill-b', { skillsDir, dataDir, status: 'pending' });
const res = applyApprovedAction(pending, applyOpts(root, skillsDir, dataDir, true));
assert.equal(res.applied, false);
assert.match(res.revalidate.reason, /approved/i);
assert.ok(existsSync(join(skillsDir, 'skill-b')), 'no mutation');
assert.equal(existsSync(join(root, 'archive')), false);
assert.equal(getCategorySkill(loadTaxonomy(dataDir), 'cat3'), 'skill-b');
});
});