POSA_Copyrighter/docs/plans/2026-06-03-operator-console-quality-improvement-plan.md
유창욱 3f7b3a9cf2 chore: initial commit of copyrighter (rights_filter)
Image rights / copyright detection system: SQLite store, HTTP app,
search integrations (Naver, Google Custom Search, Google Cloud Vision
web detection), image analysis (fingerprints, face/person detection,
evidence enrichment, risk scoring), an admin/review layer, governance
and retention policies, batch jobs, and a browser-based operator GUI.

This baseline incorporates a full code-review remediation pass
(46 fixes; 358 tests passing). Highlights:

CRITICAL
- Prevent evidence cascade-delete during the schema-constraint
  migration by disabling FK enforcement around the table rebuild.

Security
- Sandbox served media (neutralize stored XSS from uploaded/collected
  SVGs) via CSP + nosniff on the untrusted media routes.
- Strip embedded EXIF/GPS from external image derivatives before they
  are sent to third-party APIs.
- Return a clean 404 (not an uncaught StopIteration) for PATCH on an
  unknown provider.

Correctness
- LLM-summary failures no longer add +30 to the risk score.
- Decode only explicit JS escapes so Korean image URLs are not mangled.
- Consume search quota only after a successful request.
- Naver/Google adapters map responses inside the failure boundary, so a
  malformed response degrades to evidence instead of crashing enrichment.
- Domain-aware provider attribution; face-box IoU de-duplication; count
  searches (not result items); per-box crop isolation; clamp evidence
  confidence and Google CSE num; real submittedEpoch; and more.

Robustness
- Offline LLM connect fast-fails (short connect timeout) so seed/reload
  requests are not stalled; full read timeout preserved for generation.
- Malformed numeric env vars fall back to defaults instead of crashing
  startup.

Performance
- Per-submission evidence reads (no full-table scan per rescore),
  audit-log LIMIT, lazy active-store lookup, hoisted timestamps.

Tests
- ~24 regression tests added pinning the above fixes.

Runtime data (data/, outputs/, *.sqlite3, *.log), secrets (.env), and
node_modules are gitignored.
2026-06-09 09:50:31 +09:00

7.7 KiB

status created type title
active 2026-06-03 quality Operator console quality improvement plan

Operator Console Quality Improvement Plan

Scope

This plan responds to the concrete quality issues raised during review of the current operator console:

  • static tests rely too heavily on string presence;
  • visible Korean copy is partially corrupted;
  • web/operator-gui/app.js carries too many responsibilities;
  • the case workbench still needs stronger next-action guidance;
  • submission upload and folder reload UX needs clearer operator feedback.
  • the product purpose is not explicit enough in the UI.

1. Strengthen Tests Beyond String Presence

Problem: The operator GUI tests catch missing selectors and accidental deletion, but they do not prove the main flows work in a browser.

Plan:

  • Keep static tests as fast contract checks.
  • Add server-backed behavior tests for upload, reload, and manual search.
  • Add future browser smoke tests for queue selection, workbench tab switching, and suggested-query fill behavior.

Immediate action: Completed:

  • tests/operator_gui/test_browser_smoke.py now verifies the suggested-query browser flow: selecting a recommendation fills the manual query input without running a search.
  • tests/operator_gui/test_browser_smoke.py now verifies the real browser upload flow through the local HTTP server: file selection, /api/submissions/upload-image, saved image file, new queue item, workbench navigation, and selected uploaded submission.
  • tests/rights_filter/server/test_http_app.py::test_http_server_uploads_submission_image_into_active_queue verifies the upload HTTP contract saves the image into the active queue and returns the uploaded submission in the refreshed bootstrap payload.
  • tests/operator_gui/test_static_workbench.py retains fast structural checks and now also protects module load order, mojibake prevention, product purpose copy, workflow guidance, and upload UX contracts.

2. Repair Corrupted Korean Copy

Problem: Several operator-facing status strings in app.js are mojibake. This weakens trust in a review tool.

Plan:

  • Prioritize visible workflow copy over legacy compatibility strings.
  • Replace high-use status text first: reload, folder import, image upload, bulk rerun, manual search, and decision memo errors.
  • Keep compatibility-only tokens isolated until tests can be rewritten around clean labels.

Immediate action: Completed:

  • Reload, folder import, upload, bulk rerun, manual search, evidence, provider, knowledge DB, candidate, correction, and empty-state copy were normalized to readable Korean.
  • A mojibake regression check now scans the operator GUI files for common corrupted UTF-8 fragments.
  • The operator GUI JavaScript files pass node --check.

3. Split app.js Responsibilities

Problem: app.js owns queue rendering, workbench rendering, search, knowledge DB, provider controls, upload, and audit events.

Plan:

  • First extract pure helpers: file payload reading, evidence sufficiency, query suggestions, and import status formatting.
  • Then split rendering domains into small static modules if the deployment can safely move to module scripts.
  • Avoid a broad framework migration.

Immediate action: Completed:

  • web/operator-gui/operator-labels.js owns visible label dictionaries.
  • web/operator-gui/submission-import.js owns submission import/upload file payload helpers and import status message helpers.
  • web/operator-gui/evidence-guidance.js owns evidence sufficiency checks, follow-up reasons, query seed normalization, and suggested evidence queries.
  • web/operator-gui/operator-search.js owns query status labels, query strategy labels, and manual search provider normalization.
  • web/operator-gui/app.js now delegates those concerns to the extracted helpers while retaining rendering and event orchestration.

Remaining:

  • Rendering domains are still mostly in app.js. Further extraction should focus on evidence rendering and knowledge/candidate rendering only if needed, because the current no-framework static runtime is now split enough to reduce the most immediate risk.

4. Improve Next-Action Guidance

Problem: The workbench shows evidence, but does not always tell the operator what to do when evidence is thin.

Plan:

  • Keep the existing "근거 보강 추천" panel.
  • Expand the sufficiency heuristic with clearer reasons: no direct match, too few searchable results, provider empty state, or duplicate queries.
  • Add a browser smoke test to ensure clicking a recommendation fills the manual query form without running a search.

Immediate action: Completed:

  • The "근거 보강 추천" panel now explains concrete insufficiency reasons such as missing direct/page evidence, fewer than two searchable results, empty provider results, and failed provider history.
  • Suggested queries fill the manual search input and switch to the query tab without automatically running a search.
  • Browser smoke coverage verifies this non-automatic behavior.

5. Clarify Upload And Folder Reload UX

Problem: Operators should not need to understand active queue internals. After adding a photo or changing folders, the UI should say what happened.

Plan:

  • Provide a file picker for adding a photo to the active queue.
  • Save uploaded files into the active queue folder and rescan immediately.
  • Select the uploaded submission when available.
  • Show clear status text for imported count, selected ID, upload progress, and active folder label.

Immediate action: Completed:

  • The upload endpoint and UI are implemented.
  • Upload success selects the uploaded submission, resets queue filters, moves to the workbench evidence tab, and tells the operator the new case was selected.
  • The queue header now explains that adding a photo creates a new review case and opens it for case review.
  • Browser smoke coverage verifies the real upload flow end to end.

6. Clarify Product Purpose And Workflow

Problem: The console exposed many implementation concepts and actions, but the product purpose and operator flow were not explicit enough.

Plan:

  • State the product purpose in the top bar.
  • Show the operator's expected flow on the queue screen.
  • Keep this copy stable with static tests.

Completed:

  • The top bar now states: "이미지 저작권 위험 심사" and explains that submitted images, external search evidence, and the internal criteria DB are reviewed together.
  • The queue screen now shows a three-step operating flow: "심사 건 추가", "근거 보강", and "운영 결정".
  • Static tests protect the purpose copy and workflow copy.

Verification

Run:

python -m pytest tests\operator_gui\test_static_workbench.py
python -m pytest tests\operator_gui\test_browser_smoke.py
python -m pytest tests\rights_filter\server\test_http_app.py::test_http_server_uploads_submission_image_into_active_queue
node --check web\operator-gui\app.js
node --check web\operator-gui\operator-labels.js
node --check web\operator-gui\submission-import.js
node --check web\operator-gui\evidence-guidance.js
node --check web\operator-gui\operator-search.js

Latest local verification:

  • python -m pytest tests\operator_gui\test_static_workbench.py tests\operator_gui\test_browser_smoke.py -> 39 passed.
  • python -m pytest tests\rights_filter\server\test_http_app.py::test_http_server_uploads_submission_image_into_active_queue -> 1 passed.
  • node --check passed for app.js, operator-labels.js, submission-import.js, evidence-guidance.js, and operator-search.js.

Future verification:

  • If more rendering code is extracted from app.js, add a browser smoke check for that affected flow before relying on static checks.
  • If the visual layout changes again, refresh the ui-overhaul-final-results.json overflow audit and screenshots.