Spor B / B3 start. Operatør-gated dry-run for skill-livssyklus, merge først.
- decisions-io.mjs: additivt skill-nøklet `actions`-lag ved siden av urørt
URL-nøklet `decisions` (14 URL-tester urørt = regresjonsbevis). version=1,
loadDecisions backfiller manglende actions-nøkkel. Nye rene fn: actionKey
(merge symmetrisk på paret), isActionDecided, recordAction (no-mutate),
listActions. Policy A gjelder også actions.
- scripts/kb-eval/lib/skill-ops.mjs: ren planMergeSkills → {entry,diff,guardrail},
leser ingen disk, anvender INGENTING. Guardrail = ingen kuratert verdi tapt
(count-invariant + set-equality + kollisjons-deteksjon; ok=false ved klobring).
Impur skall loadSkillRefs/runMergePlan + CLI plan-skill-op.mjs (gated --write).
- Ekte eng↔infra dry-run: guardrail OK (34+153=187, 0 kollisjoner), 153 fil-flytt
+ 7 taksonomi-reassign + retire eng; 0 skriving til skills/, ledger uendret
(retning er operatør/judge-valg, ikke pushet gjennom gaten).
- TDD: test-decisions-actions (10) + test-skill-ops-merge (10, inkl. byte-for-byte
skills/-snapshot som no-write-bevis). kb-eval 40→50, kb-update 122→132.
Regresjon grønn: validate 239, kb-integrity 192/192.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
174 lines
8.1 KiB
JavaScript
174 lines
8.1 KiB
JavaScript
// tests/kb-eval/test-skill-ops-merge.test.mjs
|
|
// Spor B / B3 (Sesjon 16): merge_skills as a GATED DRY-RUN PLANNER.
|
|
//
|
|
// Two layers:
|
|
// (A) pure core — planMergeSkills(absorber, absorbed, ctx): mutates nothing,
|
|
// reads no disk, returns { entry, diff, guardrail }. The guardrail proves
|
|
// no curated value is lost (count-invariant + set-equality) and refuses
|
|
// (ok=false) when two ref files would clobber on the same target path.
|
|
// (B) impure shell — runMergePlan(...) + loadSkillRefs(...): read-only disk
|
|
// load, optional gated `--write` of a PENDING entry to decisions.json.
|
|
// The destructive proof: running the plan makes ZERO changes under skills/.
|
|
|
|
import { test } from 'node:test';
|
|
import assert from 'node:assert/strict';
|
|
import { mkdtempSync, rmSync, mkdirSync, writeFileSync, readFileSync, existsSync, readdirSync, statSync } from 'node:fs';
|
|
import { tmpdir } from 'node:os';
|
|
import { join } from 'node:path';
|
|
import { planMergeSkills, loadSkillRefs, runMergePlan, MERGE_OPERATION } from '../../scripts/kb-eval/lib/skill-ops.mjs';
|
|
import { loadDecisions, isActionDecided, actionKey } from '../../scripts/kb-update/lib/decisions-io.mjs';
|
|
|
|
// ===========================================================================
|
|
// (A) Pure core — planMergeSkills
|
|
// ===========================================================================
|
|
|
|
const CTX = {
|
|
skillRefs: {
|
|
'skill-a': ['cat1/a1.md', 'cat1/a2.md', 'cat2/a3.md'], // absorber, 3
|
|
'skill-b': ['cat3/b1.md', 'cat3/b2.md'], // absorbed, 2
|
|
},
|
|
descriptions: { 'skill-a': 'desc A', 'skill-b': 'desc B' },
|
|
categorySkill: { cat1: 'skill-a', cat2: 'skill-a', cat3: 'skill-b' },
|
|
decided_at: '2026-06-20',
|
|
};
|
|
|
|
test('planMergeSkills — guardrail OK for a clean (no-collision) union', () => {
|
|
const { guardrail } = planMergeSkills('skill-a', 'skill-b', CTX);
|
|
assert.equal(guardrail.ok, true);
|
|
assert.equal(guardrail.preCountAbsorber, 3);
|
|
assert.equal(guardrail.preCountAbsorbed, 2);
|
|
assert.equal(guardrail.expectedPostCount, 5, 'no loss: union of both ref sets');
|
|
assert.equal(guardrail.postCount, 5);
|
|
assert.equal(guardrail.countInvariant, true);
|
|
assert.equal(guardrail.setEquality, true);
|
|
assert.deepEqual(guardrail.collisions, []);
|
|
});
|
|
|
|
test('planMergeSkills — set-equality: EVERY ref from both skills survives in the plan', () => {
|
|
const { diff, guardrail } = planMergeSkills('skill-a', 'skill-b', CTX);
|
|
// absorbed files appear as moves; absorber files stay in place. Their union of
|
|
// references-relative identities must equal the pre-merge union with no loss.
|
|
const movedIdentities = diff.fileMoves.map((m) => m.to.replace('skills/skill-a/references/', ''));
|
|
const survivingIdentities = new Set([...CTX.skillRefs['skill-a'], ...movedIdentities]);
|
|
for (const r of [...CTX.skillRefs['skill-a'], ...CTX.skillRefs['skill-b']]) {
|
|
assert.ok(survivingIdentities.has(r), `${r} must survive the merge`);
|
|
}
|
|
assert.equal(survivingIdentities.size, guardrail.expectedPostCount);
|
|
});
|
|
|
|
test('planMergeSkills — guardrail FAILS (ok=false) when ref paths would collide/clobber', () => {
|
|
const colliding = {
|
|
...CTX,
|
|
skillRefs: { 'skill-a': ['cat1/shared.md', 'cat1/a1.md'], 'skill-b': ['cat1/shared.md'] },
|
|
};
|
|
const { guardrail } = planMergeSkills('skill-a', 'skill-b', colliding);
|
|
assert.equal(guardrail.ok, false, 'a clobber must not be silently approved');
|
|
assert.deepEqual(guardrail.collisions, ['cat1/shared.md']);
|
|
assert.equal(guardrail.countInvariant, false, 'postCount < expected -> loss detected');
|
|
});
|
|
|
|
test('planMergeSkills — entry is a PENDING merge_skills action with both targets', () => {
|
|
const { entry } = planMergeSkills('skill-a', 'skill-b', CTX);
|
|
assert.equal(entry.operation_type, MERGE_OPERATION);
|
|
assert.equal(entry.status, 'pending', 'planner only proposes — never approved');
|
|
assert.deepEqual(entry.targets, { absorber: 'skill-a', absorbed: 'skill-b' });
|
|
assert.equal(entry.decided_at, '2026-06-20');
|
|
assert.equal(entry.guardrail.ok, true);
|
|
});
|
|
|
|
test('planMergeSkills — diff carries moves, taxonomy reassignment, retire, and manual description reconciliation', () => {
|
|
const { diff } = planMergeSkills('skill-a', 'skill-b', CTX);
|
|
assert.deepEqual(diff.fileMoves, [
|
|
{ from: 'skills/skill-b/references/cat3/b1.md', to: 'skills/skill-a/references/cat3/b1.md' },
|
|
{ from: 'skills/skill-b/references/cat3/b2.md', to: 'skills/skill-a/references/cat3/b2.md' },
|
|
]);
|
|
assert.deepEqual(diff.taxonomyReassignments, [{ category: 'cat3', from: 'skill-b', to: 'skill-a' }]);
|
|
assert.equal(diff.retire, 'skill-b');
|
|
assert.equal(diff.descriptionReconciliation.absorber, 'desc A');
|
|
assert.match(diff.descriptionReconciliation.note, /reconcil/i, 'semantic merge flagged as manual, not auto-done');
|
|
});
|
|
|
|
test('planMergeSkills — throws on self-merge', () => {
|
|
assert.throws(() => planMergeSkills('skill-a', 'skill-a', CTX), /self|same/i);
|
|
});
|
|
|
|
test('planMergeSkills — missing ref entry defaults to empty (no throw)', () => {
|
|
const { guardrail } = planMergeSkills('skill-a', 'ghost', { skillRefs: { 'skill-a': ['c/x.md'] } });
|
|
assert.equal(guardrail.preCountAbsorbed, 0);
|
|
assert.equal(guardrail.expectedPostCount, 1);
|
|
assert.equal(guardrail.ok, true);
|
|
});
|
|
|
|
// ===========================================================================
|
|
// (B) Impure shell — loadSkillRefs + runMergePlan (read-only + gated ledger write)
|
|
// ===========================================================================
|
|
|
|
// Build a throwaway skills/ fixture: two skills with a few non-colliding refs.
|
|
function withSkillsFixture(fn) {
|
|
const root = mkdtempSync(join(tmpdir(), 'skillops-'));
|
|
const mk = (skill, cat, file, body) => {
|
|
const dir = join(root, 'skills', skill, 'references', cat);
|
|
mkdirSync(dir, { recursive: true });
|
|
writeFileSync(join(dir, file), body);
|
|
};
|
|
mk('skill-a', 'cat1', 'a1.md', '# A1\n');
|
|
mk('skill-a', 'cat1', 'a2.md', '# A2\n');
|
|
mk('skill-b', 'cat3', 'b1.md', '# B1\n');
|
|
const dataDir = join(root, 'data');
|
|
mkdirSync(dataDir, { recursive: true });
|
|
try {
|
|
return fn({ skillsDir: join(root, 'skills'), dataDir });
|
|
} finally {
|
|
rmSync(root, { recursive: true, force: true });
|
|
}
|
|
}
|
|
|
|
// Snapshot every file under a dir as path -> mtimeMs+size, for change detection.
|
|
function snapshot(dir) {
|
|
const out = {};
|
|
const walk = (d) => {
|
|
for (const e of readdirSync(d, { withFileTypes: true })) {
|
|
const p = join(d, e.name);
|
|
if (e.isDirectory()) walk(p);
|
|
else { const s = statSync(p); out[p] = `${s.size}:${s.mtimeMs}`; }
|
|
}
|
|
};
|
|
walk(dir);
|
|
return out;
|
|
}
|
|
|
|
test('loadSkillRefs — read-only, returns references-relative paths per skill', () => {
|
|
withSkillsFixture(({ skillsDir }) => {
|
|
const refs = loadSkillRefs(skillsDir);
|
|
assert.deepEqual(refs['skill-a'].sort(), ['cat1/a1.md', 'cat1/a2.md']);
|
|
assert.deepEqual(refs['skill-b'], ['cat3/b1.md']);
|
|
});
|
|
});
|
|
|
|
test('runMergePlan --write — records a PENDING entry in the ledger and touches NOTHING under skills/', () => {
|
|
withSkillsFixture(({ skillsDir, dataDir }) => {
|
|
const before = snapshot(skillsDir);
|
|
const { entry, guardrail } = runMergePlan('skill-a', 'skill-b', { skillsDir, dataDir, write: true, decided_at: '2026-06-20' });
|
|
|
|
// proposal is correct
|
|
assert.equal(entry.status, 'pending');
|
|
assert.equal(guardrail.ok, true);
|
|
assert.equal(guardrail.expectedPostCount, 3);
|
|
|
|
// ledger received the pending action (the gate's only write path)
|
|
const led = loadDecisions(dataDir);
|
|
assert.equal(isActionDecided(led, actionKey(entry)), true);
|
|
assert.equal(led.actions[actionKey(entry)].status, 'pending');
|
|
|
|
// THE DESTRUCTIVE INVARIANT: skills/ is byte-for-byte unchanged
|
|
assert.deepEqual(snapshot(skillsDir), before, 'dry-run must not write under skills/');
|
|
});
|
|
});
|
|
|
|
test('runMergePlan without --write — pure preview, no ledger file created', () => {
|
|
withSkillsFixture(({ skillsDir, dataDir }) => {
|
|
const { entry } = runMergePlan('skill-a', 'skill-b', { skillsDir, dataDir, write: false });
|
|
assert.equal(entry.status, 'pending');
|
|
assert.equal(existsSync(join(dataDir, 'decisions.json')), false, 'no write without --write');
|
|
});
|
|
});
|