Problem Add [Recommendation] Feature — Internal-Pool P2 Hybrid Recommendation + Refresh Rotation

SummaryAdded a [Recommendation] problem suggestion to the search step of the add-problem modal. To handle cold start (few real users), adopted a P2 hybrid: internal problem-pool 3-tier matching (difficulty+tags → difficulty-only → seed fallback) with server bundle prefetch (8) + client-side [Refresh] rotation. Shows 1 by default. PR #479 `4b60583` (squash), 23 files +1729/−16. All 3 Critic rounds of findings (P2·P2·P3) fixed; the final round lost its verdict to an ACP turn-reap → local verdict adopted (all findings closed). A parallel Oracle session had already completed the full pipeline, so Oracle switched to independent verification→Critic→merge without rebuilding.

Date
ImpactCritical

Date: 2026-07-21

Goal

Add a recommended-problem suggestion to the search step of the add-problem modal (a search → confirm 2-step flow). Suggest difficulty/tag-based candidates so users quickly find problems suited to their study.

Background / constraint: AlgoSu has few real users, so the internal problem pool is thin (cold start). There is also no official recommendation API from external platforms (Programmers/BOJ) → empty/sparse recommendations are a UX risk.

Decisions

D1. Recommendation source = internal problem pool (P2), no external API dependency

Suggest, among already-registered problems from other studies/weeks, those not present in this study and matching its difficulty/tags. The Problem entity's difficulty·level·tags·category provide the criteria data. Stable — no external crawling / solved.ac workaround.

D2. Cold-start fallback = 3-tier graceful degradation

recommendForStudy (services/problem/src/problem/problem.service.ts):

  1. Tier 1 — internal problems matching the difficulty band + tags
  2. If insufficient — drop the tag condition, match difficulty band only
  3. If still insufficient — fall back to a curated seed list (recommendation-seeds.ts, representative problems per difficulty, static constants)

The seed fallback means zero external-API dependency and no "empty screen" even with few users. As the pool grows, tiers 1·2 naturally dominate → the structure isn't wasted by building it now.

D3. Show 1 + [Refresh] rotation, hybrid prefetch

  • Show only 1 recommendation by default (strongest for cold start — only one candidate needed).
  • [Refresh] rotates to the next candidate. Hybrid: on modal open, fetch a bundle (8) of candidates once from the server (use-problem-recommendation.ts) → refresh is client-side rotation (instant, 0 queries); when the bundle is exhausted, re-fetch the next bundle. Removes the downsides of pure per-refresh queries (latency) and pure single-bundle (only repeats after exhaustion).

D4. Security — cross-study candidate scoping

Internal candidate query (findRecommendationCandidates): a select whitelist excludes description (prevents problem-body leakage, aligning with the Sprint 252 policy), studyId: Not(excludeStudyId) excludes the own study, and In([]) full-scan is guarded. The exclude (already-seen) list is capped by @ArrayMaxSize(100) (RecommendQueryDto).

Completed Items

  • PR #479 4b60583 (squash), 23 files +1729/−16
    • BE (services/problem): recommend endpoint (GET /recommendations, RecommendQueryDto) + recommendForStudy 3-tier + recommendation-seeds.ts + dual-write update
    • FE: use-problem-recommendation.ts (prefetch-8 rotation + shownUrls capping), SearchStep.tsx recommendation section, ConfirmStep.tsx platform alignment, problem-search.utils.ts, i18n ko/en

Verification (Oracle direct re-check — distrust self-reports): BE tsc (feature clean)·ESLint 0·jest 19 suites/245 pass (problem.service.ts 99.05%/97.97%, seeds 100%, controller 100%). FE tsc 0·lint Errors 0·jest 6 suites/95 pass.

Critic (Codex gpt-5.5, base 9c5d188): all 3 rounds of findings fixed — re-delegation to Curator per the rule.

  • R1 [P2] SearchStep.tsx — on recommendation select, the source platform followed the current tab → fixed to follow the recommendation data (d2345c6)
  • R2 [P2] use-problem-recommendation.ts — shownUrls grew unbounded on repeated refresh → capped below the backend limit (100) (e25ef1b)
  • R3 [P3] ConfirmStep.tsx — confirm display followed the tab (ignoring the recommendation platform) → render by effectivePlatform, the actual platform (ecf19b1)
  • R4 final — verdict not emitted due to ACP turn-reap → local verdict adopted (all prior findings closed, diff identical)

Incidents

  1. Parallel Oracle session finished first: while Oracle was assembling the delegation context, a parallel session had already completed the full pipeline (2 commits → push → PR #479), exactly matching the confirmed spec (P2 3-tier·seed·prefetch 8·rotation) → Oracle switched to independent verification→Critic→merge without a blind rebuild.
  2. Repeated Critic ACP turn-reap: both run_in_background and fully-detached nohup were SIGKILLed at session boundaries → unstable verdict success (3 of 4 rounds emitted a verdict, 1 lost). The watchdog's "abnormal termination detected" alert must be distinguished from a completion alert.
  3. Trivy Scan — problem FAILURE: brace-expansion·js-yaml HIGH DoS transitive-dependency CVEs — unrelated to this feature (not in the diff), drift also present on main. Non-required gate, so mergeable; bump planned as a separate PR.

Backlog

  • Bump Trivy transitive-dependency CVEs (brace-expansion·js-yaml) — feature-unrelated, separate PR (not mixed into the feature)
  • GA4 admin Enhanced Measurement OFF / production UAT / data stream URL alignment (user-direct)
  • Server redeploy + live SEO verification (ops, Sprint 212/213 deliverables)
  • Harness checkup --full CI scheduled automation (monthly cron) review (Sprint 209 follow-up)