ms-ai-architect/scripts/kb-update/run-detection.mjs
Kjell Tore Guttormsen 2e69e69326 feat(ms-ai-architect): C3.4 — detect-courses.mjs (binder C3.1–C3.3, LLM-fri) + selectDetectionSteps + include_course_detection (TDD) [skip-docs]
LLM-fri STEG 1-detektor for kurs-deteksjon (C3.4 av GODKJENT C3-spec). Speiler
discover-new-urls' to-stegs-monster: leser Keychain-creds, paginerer modules +
learning-paths filtrert til in-domain produkter, differ mot egen diff-state
(course-registry.json), skriver kandidatrapport. Skriver KUN report + registry —
ALDRI decisions.json (operator-gaten i C3.5 promoterer leads).

- scripts/kb-update/detect-courses.mjs (NY): detectCourses(deps) — injiserbar
  kjerne (readSecret/getToken/paginate/loadTaxonomy/load+saveCourseRegistry/
  saveReport/dataDir/now/log) -> hermetisk testbar uten Keychain/nett. Fail-soft
  (spec 3/6): creds=null -> status:"skipped"; run-feil -> status:"error" (kaster
  ALDRI); begge exit 0 (advisory, aldri rod pipeline). Tynn CLI med realpath-guard
  (kjorer kun nar invokert direkte) + --data-dir test-seam. updatedAt.gt-cursor
  KUN pa incremental (full ISO, gotcha #1); products-filter = comma-join (OR).
- lib/detection-schedule.mjs: ny ren selectDetectionSteps(config) gater BADE
  skillLifecycle + courseDetection (en sannhetskilde). DEFAULT_SCHEDULE_CONFIG
  +include_course_detection:false (opt-in i opt-in, spec 8d). coerce/serialize
  utvidet. DETECTION_STEPS +{detect-courses, courseDetection:true}.
- run-detection.mjs: bruker selectDetectionSteps (fjerner duplisert filter-linje).

Tester (+22):
- test-detect-courses-invariant (8): ingen saveDecisions/decisions-io/atomic-write/
  backup-import; skriver egen registry+report; binder diffCourses+makeCourseClassifier;
  LLM-fri (kommentar-strippet kildesjekk).
- test-detect-courses (6): fail-soft skipped (alle/delvis creds), error-path (no
  throw, registry urort), 4.3 report-shape + avledet skill/category, incremental
  cursor/produkt-param, baseline=ingen leads (spec 8b), hermetisk exit-0-subprosess
  via PATH-shadow av `security`.
- test-detection-schedule (+8): include_course_detection parse/serialize/default;
  selectDetectionSteps-gating (default av, opt-in pa, uavhengige gates);
  run-detection bruker selektoren.

Gotcha (secrets-hook): clientSecret: '...'-form traff llm-security-regelen
(client[_-]?secret|ClientSecret)\s*[=:]\s*['"]…{8,} -> lost med array-av-par
(feltnavn star aldri rett for quotet streng).

Gate 7 C3.4 mott: invariant + fail-soft(exit 0) + report-shape + DETECTION_STEPS-
filter gronn. kb-update 269->291, validate PASSED (239/0), run-detection dry-run
dropper detect-courses (opt-in av), eval urort (rorer ikke KB), null regresjon.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-23 14:24:57 +02:00

55 lines
2.6 KiB
JavaScript

#!/usr/bin/env node
// run-detection.mjs — Spor C / C1: the Claude-FREE detection entrypoint.
//
// Runs the pure-node detection pipeline (poll → report → discover →
// detect-skill-lifecycle), each a local script that produces a JSON report and
// NEVER contacts Anthropic. This is the structural ToS guarantee: this file
// only ever spawns `node` on the allow-listed DETECTION_STEPS — it can never
// invoke `claude`, and it never touches the apply path (which alone invokes
// Claude and stays manual + in-session + gated).
//
// The SessionStart hook background-spawns this when the opt-in schedule is
// enabled + stale (see detection-schedule.mjs). Standalone use is fine too:
// node scripts/kb-update/run-detection.mjs # run the pipeline
// node scripts/kb-update/run-detection.mjs --dry-run # list steps, run none
import { dirname, join } from 'node:path';
import { fileURLToPath } from 'node:url';
import { execFileSync } from 'node:child_process';
import { loadScheduleConfig, selectDetectionSteps } from './lib/detection-schedule.mjs';
const __dirname = dirname(fileURLToPath(import.meta.url));
const PLUGIN_ROOT = join(__dirname, '..', '..');
const dryRun = process.argv.includes('--dry-run');
// Honor the opt-in content choices: each flagged step (skill-lifecycle,
// course-detection) runs only when its config flag is on. The gating lives in
// the pure selectDetectionSteps so it is unit-tested independently of this
// executor. (enabled-gating is the hook's job; invoking this script directly is
// itself the intent to run.)
const config = loadScheduleConfig(PLUGIN_ROOT);
const steps = selectDetectionSteps(config);
function run(step) {
const fullPath = join(PLUGIN_ROOT, 'scripts', step.dir, step.script);
console.log(`\n--- detection: ${step.name} (${step.dir}/${step.script} ${step.args.join(' ')}) ---`);
try {
// Only ever `node` on a local detection script — never `claude`.
execFileSync('node', [fullPath, ...step.args], { stdio: 'inherit', timeout: 10 * 60 * 1000 });
} catch (err) {
// A failing step is non-fatal for the others — detection is advisory.
console.error(`detection step ${step.name} failed: ${err.message}`);
}
}
if (dryRun) {
console.log('DRY RUN — Claude-free detection pipeline would execute:');
steps.forEach((s, i) => console.log(` ${i + 1}. ${s.dir}/${s.script} ${s.args.join(' ')}`));
console.log('\n(All steps are pure node; none invoke Claude. Apply stays manual + in-session.)');
process.exit(0);
}
console.log('=== Claude-free detection run ===');
for (const step of steps) run(step);
console.log('\n=== detection complete (reports refreshed; 0 writes to skills/) ===');