feat(ms-ai-architect): C3.5 — courses-kolleksjon i decisions-io.mjs (UID-nøklet, Spor C) + kb-update §3c kurs-gate (TDD) [skip-docs]

Tredje ledger-kolleksjon (courses), additivt ved siden av decisions (URL,
Spor A) og actions (skill, Spor B). De tre rene helperne speiler
recordAction-trioen 1:1:
- recordCourseLead(led, uid, lead)            (ren, UID-nøklet)
- setCourseLeadStatus(led, uid, status, at)   (ren transisjon, kaster på ukjent UID)
- isCourseLeadDecided(led, uid)               (dedup policy A: any status)
createLedger()->courses:{}; loadDecisions backfiller courses ??= {} (bakoverkompat).

Strukturell ingen-ingest-invariant: apply-pathene leser ALDRI courses
(apply-skill-op.mjs->kun actions, discover-new-urls.mjs->kun decisions),
så et godkjent kurs-lead trigger aldri fetch/transform/KB-skriving.

commands/kb-update.md §3c: operatør-gate som leser course-detection-report.json,
dedup'er via isCourseLeadDecided, skriver godkjente leads via recordCourseLead
(--dry-run viser leads uten å skrive).

Tester (+16): test-decisions-courses.test.mjs. Eksisterende test-decisions-io
+ test-decisions-actions URØRT grønne (ikke-regresjons-bevis, samme mønster
som Spor B). kb-update 291->307 · validate 239/0 · kb-eval 100/0.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Kjell Tore Guttormsen 2026-06-23 14:38:56 +02:00
commit e0d2d05dbb
3 changed files with 352 additions and 8 deletions

View file

@ -135,6 +135,43 @@ d. **For `approved`:** registrér URLen i `url-registry` (gated) så den fanges
e. **Invariant:** `discover-new-urls.mjs` (deteksjon) skriver **aldri** ledgeren — den kun leser. Bare denne gaten skriver. Verifisert av `tests/kb-update/test-discover-invariant.test.mjs`.
### 3c. Kurs-deteksjon-gate — nye/endrede kurs via decision-ledger (courses-kolleksjon, Spor C / C3)
Operatør-gaten for kurs-sporet. **Et kurs-lead er et *signal* om at et tema finnes — aldri en doc-side som ingestes.** Å godkjenne et lead **henter/transformerer/skriver ingen KB-fil** (det ville vært auto-ingest, et eksplisitt ikke-mål) — det registrerer kun operatørens beslutning i den UID-nøklede `courses`-kolleksjonen. Dette er ENESTE skrivevei for kurs-beslutninger.
a. **Sjekk rapporten:** `data/course-detection-report.json`. Produsert av den Claude-frie detektoren (`detect-courses.mjs`) — enten schedulert (Tier 1/2) eller manuelt: `node scripts/kb-update/detect-courses.mjs`. Mangler fila, eller `status:"skipped"` (Keychain-creds mangler — C3 er opt-in inni opt-in): **hopp over denne gaten stille**. `status:"error"`: rapportér kort og hopp over.
b. **Les leads:** `new[]` + `updated[]` (hver `{uid, title, url, products, suggested_skill, suggested_category, updated_at}`). `removed[]` er **kun et informasjonssignal** — vis det som en notis, aldri som en beslutning (et retirert kurs er ikke et tema å dekke; spec §4.2).
c. **Dedup mot ledgeren (policy A):** utelat leads operatøren allerede har tatt stilling til. Bruk `isCourseLeadDecided` (UID er den stabile nøkkelen — URL kan endres):
```
import { loadDecisions, isCourseLeadDecided } from './scripts/kb-update/lib/decisions-io.mjs';
const led = loadDecisions();
const fresh = [...report.new, ...report.updated].filter((l) => !isCourseLeadDecided(led, l.uid));
```
d. **Presenter `fresh`** for operatør, gruppert per `suggested_skill` (vis `title`, `url`, `products`, `kind` new/updated). For hvert lead (eller batch): **approve** (relevant — verdt å dekke i KB-en), **reject** (ikke relevant — ikke vis igjen), eller **utsett** (`pending` — ikke avgjort).
Hvis `--dry-run`: stopp etter presentasjonen — **ikke skriv ledgeren**.
e. **Skriv beslutningene til ledgeren — ENESTE skrivevei.** Bruk `recordCourseLead` (ikke håndskriv JSON):
```
import { loadDecisions, recordCourseLead, saveDecisions } from './scripts/kb-update/lib/decisions-io.mjs';
let led = loadDecisions();
led = recordCourseLead(led, uid, {
kind, status: 'approved', title, url, products, suggested_skill, suggested_category,
updated_at, detected_at, decided_at: '<i dag>', note,
});
// ...én recordCourseLead per beslutning...
saveDecisions(led);
```
`decided_at` settes til dagens dato (caller-injisert — `recordCourseLead` er ren). `decisions.json` er tracket i git.
f. **Ingen ingest — strukturell invariant.** Et godkjent kurs-lead blir værende i `courses`-kolleksjonen og leses av **ingen** av apply-pathene: `apply-skill-op.mjs` leser kun `actions`, `discover-new-urls.mjs` kun `decisions`. En kurs-beslutning trigger derfor **aldri** fetch/transform/KB-skriving. Skal et kurs' tema faktisk dekkes med en KB-side, er det en **separat, bevisst** doc-discovery-handling gjennom §3b (finn doc-URLen) — ikke en automatisk konsekvens av å godkjenne kurset. Et godkjent kurs surfaces ved sesjonsstart (C3.6: «N nye / M endrede kurs i dekkede produkter») som en påminnelse til operatøren.
### 4. Per-fil oppdatering (etter brukerens `y`)
For hver fil i valgte prioriteter:

View file

@ -18,22 +18,29 @@ const DEFAULT_DATA_DIR = join(__dirname, '..', 'data');
/**
* Empty ledger scaffold.
*
* Two parallel collections, both written ONLY through the operator gate:
* Three parallel collections, all written ONLY through the operator gate:
* - decisions URL-keyed (Spor A): which Microsoft Learn page belongs where.
* - actions skill-keyed (Spor B / B3): skill-lifecycle ops (merge_skills,
* sanitize_skill, retire_skill, create_skill). Additive: version
* stays 1, and a pre-action ledger on disk normalizes cleanly
* (loadDecisions backfills the missing actions key).
* @returns {{version: number, updated_at: string|null, decisions: object, actions: object}}
* sanitize_skill, retire_skill, create_skill).
* - courses UID-keyed (Spor C / C3): training-course leads (new | updated)
* the operator may want covered in the KB. A SEPARATE collection
* on purpose an approved course lead must never reach the
* doc-transform/ingest pipeline (auto-ingest is a non-goal), so
* the apply-path (which reads only decisions/actions) never sees
* it. Dedup is per stable UID, not URL (spec §4.5).
* Additive: version stays 1, and a ledger written before any of the actions or
* courses layers existed normalizes cleanly (loadDecisions backfills the missing
* keys).
* @returns {{version: number, updated_at: string|null, decisions: object, actions: object, courses: object}}
*/
export function createLedger() {
return { version: 1, updated_at: null, decisions: {}, actions: {} };
return { version: 1, updated_at: null, decisions: {}, actions: {}, courses: {} };
}
/**
* Load the decision ledger from disk. Backward-compatible: a ledger written
* before the action layer existed (no `actions` key) is normalized to `{}` so
* callers never have to guard against undefined.
* before the action or course layers existed (no `actions` / `courses` key) is
* normalized to `{}` so callers never have to guard against undefined.
* @param {string} [dataDir] defaults to ../data/ relative to lib/
* @returns {object} parsed ledger or empty scaffold
*/
@ -42,6 +49,7 @@ export function loadDecisions(dataDir = DEFAULT_DATA_DIR) {
if (!existsSync(path)) return createLedger();
const led = JSON.parse(readFileSync(path, 'utf8'));
if (!led.actions) led.actions = {};
if (!led.courses) led.courses = {};
return led;
}
@ -184,3 +192,71 @@ export function setActionStatus(ledger, key, status, decided_at = null) {
},
};
}
// ===========================================================================
// Course-lead layer (Spor C / C3) — additive, UID-keyed, mirrors the action
// helpers. detect-courses.mjs (Claude-free detection) produces a report; the
// operator gate records APPROVED/REJECTED leads here. A course lead is a SIGNAL
// that a topic exists, never a doc to ingest — so it lives in its own `courses`
// collection that the apply-path (apply-skill-op.mjs / discover-new-urls.mjs)
// never reads. Dedup is per stable UID (spec §4.5: the course `id`, not its URL,
// is the durable key).
// ===========================================================================
/**
* Has the operator already decided on this course lead? Policy A: any entry
* (pending | approved | rejected) counts mirrors isDecided / isActionDecided.
* The gate filters detection-report candidates through this so a course the
* operator has touched never resurfaces.
* @param {object} ledger
* @param {string} uid the course's stable UID (module / learning-path id)
* @returns {boolean}
*/
export function isCourseLeadDecided(ledger, uid) {
return Boolean(ledger.courses && ledger.courses[uid]);
}
/**
* Record a course lead. Pure returns a new ledger, does not mutate. Keyed by
* the course's stable UID. Leaves the URL-keyed decisions and skill-keyed
* actions maps untouched.
* @param {object} ledger
* @param {string} uid the course's stable UID
* @param {{kind: string, status: string, decided_at?: string|null, title?: string,
* url?: string, products?: string[], suggested_skill?: string|null,
* suggested_category?: string, updated_at?: string, detected_at?: string,
* note?: string}} lead
* @returns {object} new ledger
*/
export function recordCourseLead(ledger, uid, lead) {
return {
...ledger,
updated_at: lead.decided_at ?? ledger.updated_at,
courses: { ...(ledger.courses ?? {}), [uid]: { ...lead } },
};
}
/**
* Transition a recorded course lead to a new status. Pure returns a new
* ledger, does not mutate. Every other field (title, url, products,
* suggested_skill, ) is preserved. Mirrors setActionStatus; lets the gate flip
* a pending lead approved/rejected without rebuilding the entry.
* @param {object} ledger
* @param {string} uid the course's stable UID
* @param {string} status new status (pending | approved | rejected)
* @param {string|null} [decided_at] when given, advances the entry date + updated_at
* @returns {object} new ledger
* @throws if no course lead exists under `uid`
*/
export function setCourseLeadStatus(ledger, uid, status, decided_at = null) {
const existing = ledger.courses?.[uid];
if (!existing) throw new Error(`setCourseLeadStatus: no course recorded under uid "${uid}"`);
return {
...ledger,
updated_at: decided_at ?? ledger.updated_at,
courses: {
...ledger.courses,
[uid]: { ...existing, status, decided_at: decided_at ?? existing.decided_at },
},
};
}

View file

@ -0,0 +1,231 @@
// tests/kb-update/test-decisions-courses.test.mjs
// Spor C / C3.5 (Sesjon 36): the ADDITIVE course-lead layer on the lag-2
// decision ledger. Courses are a THIRD entity type (UID-keyed) — the URL-keyed
// `decisions` map (Spor A) and the skill-keyed `actions` map (Spor B) are both
// untouched. Their contracts stay covered by test-decisions-io.test.mjs and
// test-decisions-actions.test.mjs, which remain green UNMODIFIED, proving this
// extension is non-regressive. Here we cover only the new `courses` collection:
// recordCourseLead (pure), isCourseLeadDecided (dedup policy A per UID),
// setCourseLeadStatus (pure transition), and backward-compatible normalization
// of pre-course ledgers.
//
// WHY a separate collection (spec §4.5): an approved course lead must NEVER
// reach the doc-transform/ingest pipeline (auto-ingest is an explicit non-goal).
// Keeping leads UID-keyed in `courses` holds them structurally out of the
// `decisions` stream the apply-path reads — the apply-path (apply-skill-op.mjs,
// discover-new-urls.mjs) reads only `actions`/`decisions`, never `courses`.
import { test } from 'node:test';
import assert from 'node:assert/strict';
import { mkdtempSync, rmSync, writeFileSync } from 'node:fs';
import { tmpdir } from 'node:os';
import { join } from 'node:path';
import {
createLedger,
loadDecisions,
saveDecisions,
recordDecision,
isDecided,
recordAction,
isActionDecided,
actionKey,
recordCourseLead,
isCourseLeadDecided,
setCourseLeadStatus,
} from '../../scripts/kb-update/lib/decisions-io.mjs';
function withTmp(fn) {
const dir = mkdtempSync(join(tmpdir(), 'dec-crs-test-'));
try {
return fn(dir);
} finally {
rmSync(dir, { recursive: true, force: true });
}
}
const UID = 'learn.wwl.introduction-to-azure-openai';
const LEAD = {
kind: 'new', // "new" | "updated"
status: 'pending', // pending | approved | rejected (dedup policy A)
title: 'Introduction to Azure OpenAI Service',
url: 'https://learn.microsoft.com/training/modules/intro-azure-openai/',
products: ['azure-openai'],
suggested_skill: 'ms-ai-engineering',
suggested_category: 'azure-ai-services',
updated_at: '2026-06-20T08:00:00Z',
detected_at: '2026-06-23T12:00:00Z',
decided_at: null,
note: '',
};
// --- scaffold now carries an empty courses collection (additive) ---
test('createLedger — additive courses:{} alongside untouched decisions:{} + actions:{}', () => {
const led = createLedger();
assert.equal(led.version, 1, 'version stays 1 — no migration needed');
assert.deepEqual(led.decisions, {}, 'URL-keyed map unchanged');
assert.deepEqual(led.actions, {}, 'skill-op map unchanged');
assert.deepEqual(led.courses, {}, 'new UID-keyed course-lead collection present and empty');
});
// --- isCourseLeadDecided — dedup gate (policy A: any status counts) ---
test('isCourseLeadDecided — false when absent, true for ANY recorded status (policy A)', () => {
let led = createLedger();
assert.equal(isCourseLeadDecided(led, UID), false, 'absent UID is undecided');
led = recordCourseLead(led, UID, { ...LEAD, status: 'rejected', decided_at: '2026-06-23' });
assert.equal(isCourseLeadDecided(led, UID), true, 'rejected lead still dedups (policy A)');
});
test('isCourseLeadDecided — true for pending and approved alike (policy A)', () => {
let led = createLedger();
led = recordCourseLead(led, 'uid-pending', { ...LEAD, status: 'pending', decided_at: '2026-06-23' });
led = recordCourseLead(led, 'uid-approved', { ...LEAD, status: 'approved', decided_at: '2026-06-23' });
assert.equal(isCourseLeadDecided(led, 'uid-pending'), true, 'pending dedups (policy A)');
assert.equal(isCourseLeadDecided(led, 'uid-approved'), true, 'approved dedups (policy A)');
});
// --- recordCourseLead — pure, mirrors recordAction ---
test('recordCourseLead — records under UID, returns new ledger, does not mutate input', () => {
const led = createLedger();
const next = recordCourseLead(led, UID, LEAD);
assert.deepEqual(led.courses, {}, 'input untouched');
assert.notEqual(next, led);
assert.equal(next.courses[UID].kind, 'new');
assert.equal(next.courses[UID].status, 'pending');
assert.equal(next.courses[UID].suggested_skill, 'ms-ai-engineering');
assert.deepEqual(next.courses[UID].products, ['azure-openai']);
});
test('recordCourseLead — advances updated_at to the decision date', () => {
const next = recordCourseLead(createLedger(), UID, {
...LEAD,
status: 'approved',
decided_at: '2026-06-23',
});
assert.equal(next.updated_at, '2026-06-23');
});
test('recordCourseLead — a pending lead (decided_at null) leaves updated_at unchanged', () => {
const next = recordCourseLead(createLedger(), UID, LEAD); // decided_at: null
assert.equal(next.updated_at, null, 'no decision date -> ledger updated_at untouched');
});
test('recordCourseLead — keyed by stable UID, not URL (dedup survives a URL change)', () => {
let led = recordCourseLead(createLedger(), UID, LEAD);
// Same course re-detected later with a changed URL but the SAME UID.
led = recordCourseLead(led, UID, { ...LEAD, url: 'https://learn.microsoft.com/training/modules/renamed/' });
assert.equal(Object.keys(led.courses).length, 1, 'same UID dedups to one entry despite URL change');
assert.equal(isCourseLeadDecided(led, UID), true);
});
test('recordCourseLead — does NOT touch the URL-keyed decisions or skill-keyed actions maps', () => {
let led = recordDecision(createLedger(), 'https://learn.microsoft.com/x', {
status: 'approved',
decided_at: '2026-06-19',
});
led = recordAction(led, {
operation_type: 'sanitize_skill',
status: 'approved',
decided_at: '2026-06-20',
targets: { skill: 'ms-ai-governance' },
});
const next = recordCourseLead(led, UID, LEAD);
assert.equal(isDecided(next, 'https://learn.microsoft.com/x'), true, 'URL decision survives');
assert.equal(isActionDecided(next, actionKey({ operation_type: 'sanitize_skill', targets: { skill: 'ms-ai-governance' } })), true, 'action survives');
assert.equal(isCourseLeadDecided(next, UID), true, 'course lead recorded alongside');
});
// --- ACCEPTANCE CRITERION (gate dedup): a decided course is NOT re-proposed ---
// Mirrors the discovery dedup-diff acceptance test (test-decisions-io.test.mjs):
// the gate filters report candidates through isCourseLeadDecided, so a course
// the operator has already touched (any status) never resurfaces.
test('dedup-diff — a decided course UID is surfaced in round 1 but filtered in round 2', () => {
const filterUndecidedCourses = (ledger, candidates) =>
candidates.filter((c) => !isCourseLeadDecided(ledger, c.uid));
// Round 1: empty ledger — the course is a fresh candidate.
let led = createLedger();
const round1 = filterUndecidedCourses(led, [{ uid: UID }]);
assert.deepEqual(round1.map((c) => c.uid), [UID], 'round 1 surfaces the new course');
// Operator rejects it through the gate — the ONLY write path.
led = recordCourseLead(led, UID, { ...LEAD, status: 'rejected', decided_at: '2026-06-23' });
// Round 2: same detection output, the decided course must be filtered out.
const round2 = filterUndecidedCourses(led, [{ uid: UID }]);
assert.deepEqual(round2, [], 'round 2 must NOT re-propose the decided course');
});
// --- setCourseLeadStatus — pure status transition (mirrors setActionStatus) ---
test('setCourseLeadStatus — transitions status, preserves every other field, is pure', () => {
const led = recordCourseLead(createLedger(), UID, { ...LEAD, status: 'pending' });
const next = setCourseLeadStatus(led, UID, 'approved', '2026-06-24');
assert.equal(led.courses[UID].status, 'pending', 'input ledger untouched (pure)');
assert.notEqual(next, led);
assert.equal(next.courses[UID].status, 'approved');
assert.equal(next.courses[UID].title, LEAD.title, 'title preserved');
assert.deepEqual(next.courses[UID].products, LEAD.products, 'products preserved');
assert.equal(next.courses[UID].suggested_skill, LEAD.suggested_skill, 'suggested_skill preserved');
});
test('setCourseLeadStatus — advances entry decided_at and ledger updated_at', () => {
const led = recordCourseLead(createLedger(), UID, { ...LEAD, status: 'pending' });
const next = setCourseLeadStatus(led, UID, 'approved', '2026-06-24');
assert.equal(next.courses[UID].decided_at, '2026-06-24', 'entry date advances');
assert.equal(next.updated_at, '2026-06-24');
});
test('setCourseLeadStatus — without a date keeps the entry decided_at and ledger updated_at', () => {
let led = recordCourseLead(createLedger(), UID, { ...LEAD, status: 'pending', decided_at: '2026-06-23' });
const next = setCourseLeadStatus(led, UID, 'approved');
assert.equal(next.courses[UID].decided_at, '2026-06-23', 'no date -> keep original');
});
test('setCourseLeadStatus — throws on an unknown UID (cannot transition a non-entry)', () => {
const led = recordCourseLead(createLedger(), UID, LEAD);
assert.throws(() => setCourseLeadStatus(led, 'uid-ghost', 'approved'), /no course|ghost/i);
});
test('setCourseLeadStatus — leaves decisions, actions, and sibling courses untouched', () => {
let led = recordDecision(createLedger(), 'https://learn.microsoft.com/x', {
status: 'approved',
decided_at: '2026-06-19',
});
led = recordCourseLead(led, UID, { ...LEAD, status: 'pending' });
led = recordCourseLead(led, 'uid-sibling', { ...LEAD, status: 'pending' });
const next = setCourseLeadStatus(led, UID, 'approved', '2026-06-24');
assert.equal(isDecided(next, 'https://learn.microsoft.com/x'), true, 'URL decision survives');
assert.equal(next.courses['uid-sibling'].status, 'pending', 'sibling course untouched');
});
// --- backward compatibility: a pre-course ledger on disk loads cleanly ---
test('loadDecisions — normalizes a pre-course (no courses key) ledger to courses:{}', () => {
withTmp((dir) => {
// Simulate an on-disk ledger written before the course layer existed.
writeFileSync(
join(dir, 'decisions.json'),
JSON.stringify({ version: 1, updated_at: null, decisions: {}, actions: {} }, null, 2),
);
const led = loadDecisions(dir);
assert.deepEqual(led.courses, {}, 'missing courses key is backfilled, not undefined');
// and recording a course lead on the loaded ledger works
const next = recordCourseLead(led, UID, LEAD);
assert.equal(isCourseLeadDecided(next, UID), true);
});
});
test('saveDecisions + loadDecisions — courses round-trip through disk', () => {
withTmp((dir) => {
const led = recordCourseLead(createLedger(), UID, LEAD);
saveDecisions(led, dir);
const back = loadDecisions(dir);
assert.deepEqual(back, led);
});
});