Recommendation Loading Recovery + Platform Toggle Scoping — Deployment Gap Diagnosis and platform Filter

Date: 2026-07-22

Goal

Recover the recommendation feature added in Sprint 254 after a report that recommendations could not be loaded in production. Then improve the previously platform-mixed recommendation exposure by scoping it to the problem-add modal's platform toggle (BOJ/Programmers) so only the selected platform's problems are suggested.

Background / Constraints: Reproduced against the actual deployment URL (https://algo-su.com). The source-code path (FE hook → gateway → controller route priority → service Tier3 seed fallback) had no defect under static analysis → hypothesis set the cause as a runtime/deployment gap rather than code.

Decisions

D1. Root cause = deployment gap (not a code bug)

Confirmed with physical evidence from CI logs. After PR #479 (4b60583) merged, only the Trivy Scan — problem job failed on main CI, and the deploy job's service-scoped fail-closed gate skipped the GitOps image-tag update for the Trivy-failed service (log: ⚠ algosu-problem SKIPPED (Trivy status: fail — service-scoped security gate)). As a result, the deployed problem-service stayed pinned to the old ee4d4fe (Sprint 251) → the /recommendations route was absent → the new FE called a nonexistent route and failed. The frontend deployed normally, producing a new/old mismatch.

• The main-4b60583 problem image built and pushed to GHCR successfully (only the tag update was skipped) — the image itself exists.

D2. Recovery = unblock the gate by resolving transitive CVEs (Option A)

The Trivy failure cause was two newly-published transitive-dependency HIGH CVEs in problem-service (feature-unrelated drift):

• brace-expansion → CVE-2026-13149 (DoS)
• js-yaml → CVE-2026-59869 (DoS)

Instead of a temporary .trivyignore entry (Option B, fast but defers the DoS), we adopted pinning patched versions via services/problem/package.json overrides (Option A, real security fix + unblock at once) — service-stability-first. Same pattern as the existing multer·tmp·ajv overrides:

• brace-expansion@>=3.0.0 <=5.0.6 → 5.0.7, brace-expansion@<=1.1.15 → 1.1.16
• js-yaml@>=4.0.0 <=4.2.0 → 4.3.0, js-yaml@<=3.14.2 → 3.15.0

PR #481 c889092 merged → CI green → Trivy problem passed → GitOps tag updated → problem-service redeployed → recommendation route restored.

D3. Scope recommendations to the platform toggle (feature improvement)

The problem-add modal has a platform toggle, but recommendations ignored it and mixed all platforms. Scope recommendations to the toggle so only the selected platform is suggested:

• BE: Add a platform parameter to RecommendQueryDto → controller → delegate to recommendForStudy. Apply a conditional sourcePlatform filter in the DB where of the cross-study candidate query (findRecommendationCandidates), and filter the Tier3 seed fallback by platform too. When unspecified, expose all (backward compatible).
• Cold start: Add a BOJ representative-problem seed set (recommendation-seeds.ts). Do not assert precise solved.ac tiers — use level=null, difficulty macro-category only.
• FE: The use-problem-recommendation hook takes a platform option and refetches on tab switch (resetting exposure history). SearchStep injects the current tab's platform.

PR #482 e14e8c6 merged.

D4. Critic gate = foreground blocking execution (avoiding ACP reap)

The background PTY-wrapper approach was reaped twice in a row in this ACP environment (SIGKILL without a verdict, no .done marker created). Rather than a third identical retry, switched to foreground blocking execution (macOS script PTY emulation, base c889092, -c model="gpt-5.5" pin) → completed without a reap. Verdict: 1 P2 (a race where, if a tab switch occurs while a previous platform query is in-flight, the loadingRef guard drops the refetch so the stale-platform response overwrites the screen). Resolved with an epoch-ref-based stale-discard pattern + a regression test (86b041d).

Completed Items

• PR #481 c889092 — problem-service Trivy HIGH CVE override resolution (completed first by a parallel Oracle session). Unblocked the deployment gap.
• PR #482 e14e8c6 — recommendation platform-toggle scoping
  • BE(services/problem): RecommendQueryDto.platform + controller delegation + recommendForStudy platform filter + findRecommendationCandidates conditional sourcePlatform where + new BOJ seeds in recommendation-seeds.ts
  • FE: use-problem-recommendation.ts (platform option + refetch on tab switch + epoch stale discard), SearchStep.tsx current-tab platform injection

Verification (Oracle direct re-check — distrust self-reports): problem BE tsc (feature clean)·ESLint 0·jest 252 pass. FE tsc 0·ESLint 0·jest 98 pass, recommendation hook 14/14 (including the new race regression test).

Critic (Codex gpt-5.5, base c889092, foreground): 1 P2 → resolved with the epoch pattern (86b041d). Re-check closed on green local gates.

Incidents

1. Deployment-gap misdiagnosis risk: "the code is fine but recommendations don't show" was easy to misread as a code defect from static analysis alone. Reproducing against the live deployment URL + tracing CI-log physical evidence (per-job results, GitOps tags) pinned down the problem-service old-version freeze.
2. Critic ACP reap twice in a row: the background PTY wrapper lost the verdict to SIGKILL at session boundaries repeatedly (isomorphic to Sprint 254). Solved with foreground blocking execution.
3. Parallel Oracle session first (recovery part): just before starting Option A, a parallel session had already completed and merged the same overrides / PR #481 → switched from rebuild to CI/redeploy verification.