Sprint 159 — AI Analysis Data Parsing Hotfix + Base Image CVE Deploy Unblock

Goal

• Hotfix a bug where the AlgoSu study user dh4m's recent submission's AI analysis result page exposed raw JSON/markdown text as-is
• Immediately after merging the main hotfix PR, post-merge Trivy Scan — ai-analysis failed, causing Update GitOps manifests + Deploy Notification to skip → production deploy blocked → unblock via forward-fix
• Formal retrospective record for an urgent hotfix sprint that proceeded without /start invocation

Decisions

• Triple-defense pattern adopted: Instead of a single-point fix, applied (A) backend envelope + (B) frontend string-aware brace counter + (C) frontend friendly fallback as three layers. If one layer is bypassed, the next layer defends. Direct continuation of Sprint 155's three-layer safety net pattern
• Auto-Critic P2 option-branch judgment: Critic R1 P2 finding (1 case: raw_excerpt logger exposure) → Oracle chose between option A (allow) / option B (reject) → Option B adopted. The sprint goal = block raw exposure, so logger exposure persistence is self-contradictory. Decision driven by sprint goal (Sprint 155 pattern carry-over)
• "Main merge failed" forward-fix adopted: PR #272 merge itself succeeded, but post-merge Trivy fail blocked the deploy pipeline. Instead of revert, added a base image patch PR for unblock. Forward-fix is correct since the code itself is normal
• Base image patch approach: Applied apt-get update && upgrade -y to both stages instead of switching to alpine. Alpine carries build-time dependency change risk and belongs to a separate sprint scope. .trivyignore rejected as self-contradictory (bypassing security scan in a security exposure fix sprint)
• Self-contradiction detection possible after Critic 1st pass: P1 envelope introduction was OK, but raw_excerpt persistence in logger contradicted the sprint goal. Codex cross-validation's hotfix sprint consistency guard effectiveness reaffirmed

Implementation (2 PRs squash merge, origin/main a73c596 → 2ec3747)

PR #272 Details — AI Analysis Parsing Hotfix (5 commits squash)

Symptom: dh4m's submission AI analysis result page exposed raw JSON/markdown text instead of structured categories/summary

Root cause chain:

1. Claude API returned malformed JSON response (e.g., pattern like 'broken{json "totalScore": 90 more broken')
2. Backend services/ai-analysis/src/claude_client.py:_parse_response (line 216~344) all 4 fallback stages failed (markdown strip / numeric quote sanitize / optimizedCode nullification / first JSON object extraction)
3. catch-all block (line 311~344) stored raw_text directly in feedback field + returned status=completed (when totalScore regex extraction succeeded)
4. Frontend frontend/src/lib/feedback.ts:parseFeedback (line 109~146) catch block fallback exposed raw text via summary: feedback
5. Additional defect: Frontend cleanAndExtractJson brace counter (line 9599) did not recognize string-internal braces — robustness mismatch vs backend _extract_first_json_object (line 497534, with in_string + escape tracking)

Triple-defense applied:

(A) Backend envelope (services/ai-analysis/src/claude_client.py:311~362):

• catch-all fallback always stores valid JSON envelope instead of raw text
• Envelope structure: {"totalScore": <regex>, "summary": <friendly message>, "categories": [], "optimizedCode": null}
• Summary varies by score: score>0 "AI analysis result parsing encountered a temporary error... only score available" / score=0 "Cannot display AI analysis result. Please try again later"
• Critic P2 follow-up: raw_excerpt removed from logger extra, only metadata error/score_extracted/raw_length exposed

(B) Frontend string-aware brace counter (frontend/src/lib/feedback.ts:78~118):

• New findJsonObjectEnd function — 1:1 mapping with backend _extract_first_json_object's in_string + escape tracking
• Correctly handles braces inside strings (e.g., "optimizedCode": "def f(): return {1,2}")

(C) Frontend friendly fallback (frontend/src/lib/feedback.ts:144~158):

• parseFeedback catch block no longer exposes raw feedback text via summary
• Replaced with user-friendly message "Cannot display AI analysis result. Please try again later."

PR #272 Details — Critic P2 Hotfix Cycle

Critic R1 (Codex gpt-5, --base main):

• 1 P2 finding: services/ai-analysis/src/claude_client.py:317~320's raw_excerpt = raw_text.strip()[:200] followed by logger.warning(..., extra={"raw_excerpt": raw_excerpt}) exposes raw Claude response (potentially containing user code/markdown/secrets) to logs/log aggregation
• Self-contradictory with this hotfix's goal (block raw text exposure)

Oracle judgment: Option B accepted (block maintained) — sprint goal is raw exposure blocking, so logger exposure must also be blocked

Architect re-delegation result (commit 33cad49):

• raw_excerpt field removed, logger extra retains only metadata (error[:100], score_extracted, raw_length)
• score extraction logic moved above logger to utilize score_extracted label
• New test test_fallback_does_not_leak_raw_text_to_logs (caplog-based)

Critic R2: clean ✅ "did not find a discrete functional, security, or maintainability regression"

PR #272 Tests (Backend 3 updated + 2 new / Frontend 1 updated + 2 new)

Backend (services/ai-analysis/tests/test_claude_client.py):

• Updated: test_parse_invalid_json_returns_envelope (renamed from _returns_raw), test_fallback_total_failure_regex_score, test_fallback_total_failure_no_score — verifies feedback is valid JSON envelope + raw text never exposed in envelope/feedback
• New: test_fallback_envelope_is_valid_json — feedback always parses as valid JSON dict even with mixed raw markdown response
• New: test_fallback_does_not_leak_raw_text_to_logs (caplog-based) — verifies raw_text/potential secret tokens never exposed in logger anywhere + only metadata fields exist

Frontend (frontend/src/lib/__tests__/feedback.test.ts):

• Updated: parseFeedback fallback friendly message verification
• New: cleanAndExtractJson handles braces inside string values — string-internal braces parsed correctly
• New: parseFeedback fallback shows friendly message not raw dump — verifies raw text not included in summary

Verification:

• pytest tests/test_claude_client.py: 78/78 PASS (claude_client.py coverage 99%)
• jest feedback.test.ts: 44/44 PASS + tsc clean
• ruff check src/: All checks passed
• ruff format src/: 9 files already formatted (after commit 44eee44)

PR #273 Details — Base Image CVE Deploy Unblock

Symptom: PR #272 merge itself succeeded (origin/main 505568a), but post-merge CI run 26071766405 Trivy Scan — ai-analysis failed → Update GitOps manifests + Deploy Notification skipped → production deploy blocked. The reality of "main merge failure".

Trivy detected 3 CVEs (all base image python:3.13-slim = Debian 13 Bookworm OS packages):

Security surface comparison:

• ai-analysis: python:3.13-slim (Debian 13) ← affected
• 5 NestJS services: node:22-alpine ← unaffected (alpine, debian-independent)
• Python packages (Pydantic, FastAPI, anthropic, etc.): all clean (0 vulnerabilities)

Changes (services/ai-analysis/Dockerfile, +10):

• Added OS security patch layer to both builder + runner stages:

RUN apt-get update \
    && apt-get upgrade -y --no-install-recommends \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*

• debian security repo's deb13u1 patches auto-applied
• --no-install-recommends + cache cleanup minimizes image size impact

Verification:

• PR #273 CI: 29 SUCCESS / 0 FAIL (Trivy SKIPPED on PR stage, runs only post-merge)
• Critic R1: clean ✅ "package index update/upgrade/cleanup steps to both Docker build stages... no discrete regression"
• Post-merge CI run 26072326494: ALL SUCCESS including:
  • Trivy Scan — ai-analysis: ✅ success
  • Update GitOps manifests: ✅ success (previously SKIPPED → unblock)
  • Deploy Notification: ✅ success (previously SKIPPED → unblock)

Verification

New Patterns

1. Triple-defense pattern — backend envelope + frontend string-aware parser + user-friendly fallback. If one layer is bypassed, the next defends. Direct continuation of Sprint 155's three-layer safety net (plan + pre-push + CI lint), but this sprint applies backend/frontend layer separation
2. Auto-Critic P2 hotfix sprint self-contradiction detection — P1 envelope introduction successfully blocks raw exposure, but raw_excerpt persistence in logger self-contradicts sprint goal. Codex cross-validation's sprint consistency guard effectiveness. Continuation of Sprint 117~ Auto-Critic establishment pattern
3. Oracle option-branch sprint-goal-driven decision — between Critic-proposed options A/B, sprint goal = raw exposure blocking → option B (maintain block) adopted. Direct continuation of Sprint 155 pattern
4. "Main merge failed" forward-fix pattern — when git merge succeeds + post-merge security gate fails blocking deploy pipeline, avoid revert and add base image patch PR for unblock. Maintains code normality + separates security patch as distinct PR for clear history
5. PR-stage Trivy SKIP + post-merge fail pattern exposed — matrix conditional causes PR-stage SKIP, post-merge actually runs. Base image regressions undetectable pre-merge → post-fact forward-fix required. Sprint 160 seed: PR-stage Trivy activation
6. Python (Debian) vs NestJS (alpine) security surface difference — Python services' base image exposes OS CVEs more frequently. Base image regular update automation needed. Sprint 160 seed: Dependabot Dockerfile updater or weekly cron
7. Formal retrospective for urgent hotfix sprint without /start — abnormal flow but with actual work merged to main, user /stop invocation uses Sprint 159 slot + normal retrospective written. Status guard bypass policy

Lessons

1. Backend regex score fallback robustness ≠ user visibility — totalScore regex extraction + status=completed makes analysis itself appear success, but raw exposure in feedback field renders it useless to users. Envelope consistency required all the way to frontend
2. Frontend brace counter must 1:1 map with backend string-aware logic — defense in depth. If only one side is string-aware, the other receives broken JSON and enters fallback. Sprint 159 hotfix's nature exposes this tech debt
3. Self-contradiction detection possible after Critic 1st pass — P1 envelope introduction succeeded, but raw_excerpt persistence in logger self-contradicts. Auto-Critic auto-queue (Sprint 117~) effectiveness reaffirmed this sprint. Sprint goal consistency guard at same level as message/symbol/exit code consistency
4. PR-stage Trivy SKIP cannot detect base image regressions pre-merge → post-fact forward-fix required. Sprint 160 seed activation to move detection earlier
5. Successful git merge ≠ successful deploy — when post-merge security gate fails, GitOps update + Deploy Notification skipped means no production deployment. "Main merge failed" reality = deploy pipeline blocked. User-facing message is ambiguous, needs clearer guard
6. python:3.13-slim Debian base may regularly expose OS CVEs — alpine switch carries build-time dependency change risk, but base image regular update is essential. Sprint 160 seed: Dependabot Dockerfile updater or weekly cron for automation
7. _parse_group_response shares same raw_text fallback pattern — identified in architect report. Outside this sprint's scope, but group analysis usage carries same exposure risk. Sprint 160 carry-over seed
8. apt-get upgrade on both stages is standard pattern — safer than base image tag change (regression risk). --no-install-recommends + cache cleanup minimizes image size impact (estimated +5~10MB)
9. PR merge timing is not deploy completion — needs explicit "merge ≠ deploy" distinction in user-facing alerts/Discord, etc. (Sprint 160+ UX seed)

Sprint 160 Carry-Over Seeds

New Automation Candidates (Derived from Sprint 159)

• Seed #new1: PR-stage Trivy scan activation — change matrix conditional for pre-merge base image regression detection (avoid this sprint's forward-fix cycle)
• Seed #new2: Base image regular update automation — Dependabot Dockerfile updater activation or weekly cron apt-get upgrade rebuild (continuation of Sprint 156 weekly cron pattern)
• Seed #new3: Apply same envelope to _parse_group_response raw_text fallback — extend this sprint's dh4m fix envelope protection to group analysis (architect-identified)
• Seed #new4: Strengthen deploy pipeline blocking alerts — clearly separate "git merge success + deploy block" cases in Discord/Slack notifications (avoid user's ambiguous "main merge failed" message)

Sprint 158 Carry-Over

• New automation candidates:
  • Seed #30: Build output Korean residue auto-validation CI step (allowlist-based)
  • Seed #31: i18n matching checklist 3-layer (meta/UI/body) plan template auto
• Sprint 157 carry-over:
  • Seed #24: plan template i18n dual-side mandatory checklist auto
  • Seed #26: docs/adr/README.md paths filter negation
  • Seed #27: CI build-blog ls out/ output existence verification step
  • Seed #28: check-adr-links.mjs ROOT auto-detection
  • Seed #29: plan template "probe step companion mandatory for new CI step"
• UAT user-direct (16 sprints cumulative):
  • Seed #5: Programmers re-submission scoring pass confirmation
  • Seed #9: English environment + production Grafana CB dashboard ai-analysis visual integrity
• Carry-over maintained:
  • Seed #18: Blog post pre-merge domain fact cross-check automation
  • Seed #23: plan template "post-rebase cumulative count fix" checklist
• Follow-up (optional):
  • create/edit page.tsx category UI
  • Programmers URL auto category inference
  • Existing SQL problem data backfill
  • coverage-gate skipped allowance removal (Sprint 156 Phase B option B)
  • post-merge pre-deploy gate (Sprint 156 Phase B option C)
  • prom-client Case B~D inspection automation
  • .claude-tools/ Phase 2 actual deletion (after trigger path verification)
  • (adr) layout split (KR + EN override) — alternative to Sprint 158 description unification

Branch Discipline

✅ 27 sprints consecutive compliance — both PRs new branches + Squash merge, 0 main direct commits, 0 --no-verify uses. This sprint proceeded without /start, but formal process maintained: work branch separation + PR merge + Critic cross-validation.

References

• Previous sprint: sprint-158.md
• PR #272: https://github.com/tpals0409/AlgoSu/pull/272
• PR #273: https://github.com/tpals0409/AlgoSu/pull/273
• main HEAD: a73c596 → 2ec3747