Gsx Vrf Review

gsx-vrf-review

Codex Adapter

In Codex, translate this Claude wrapper instead of rewriting it:

• $ARGUMENTS = text after $gsx-*.
• /gsd: ... or Skill("gsd:", args="...") = run .codex/skills/gsd-/SKILL.md with args as {{GSD_ARGS}}.
• AskUserQuestion: only map to request_user_input if your Codex build truly renders an interactive picker; otherwise (the usual Codex case) do not force it. Before any choice among options/plans, first print each option's core content in your reply, then list the choices as a numbered Markdown list and stop and wait for the user's text reply. Never self-answer a prompt the user never saw.
• Claude tool names are intent labels; use equivalent Codex tools.
• Preserve wrapper boundaries; if GSD owns edits, do not edit source here.

At the end of a GSD checkpoint:human-verify round, audit what changed across frontend / backend / database, and surface a hidden risk: problems that should have been fixed in backend/DB but were patched in the frontend. These look fine in-browser but fail when the API is called directly — validation absent, translations/defaults missing, sensitive fields still exposed.

Flow: summarize by layer → flag risks → interview user → record decisions. No code edits, no gsd-executor, no commits, no file writes — conversation only. Acting on flagged items is the user's call (follow up with /gsd:quick).

0 — Locate scope (read-only)

$ARGUMENTS given → use it. Else auto-locate:

1. Read .planning/STATE.md for the executing active phase.
2. In .planning/phases//, find *-PLAN.md with `; read its + ` — the "should-be" baseline.
3. Not found (no .planning/, or no checkpoint task) → "Could not locate a checkpoint:human-verify task — please specify the phase / checkpoint." and stop. Do not guess.

Acceptance period = everything done in this conversation since the user began running `` and reporting issues, plus all fixes you made. Conversation context is the primary source; git is a cross-check. Don't fabricate.

1 — Reconstruct and layer-classify changes

List every change this period; cross-check git:

git status --short      # uncommitted
git diff --stat         # uncommitted file list
git log --oneline -20   # commits this period

| Layer | Criteria | Typical paths | |-------|----------|---------------| | Frontend | React pages/components/API clients/utils/styles | dap-frontend/src/... | | Backend | Controller/Service/Mapper/DTO·Request·Response/enums/config | dap-server/dap-*/src/main/java/... | | Database | table create/alter/index/migration/seed SQL | *.sql, migration dirs, docker init |

Per layer, a small table: file:line · what changed · which feedback it addressed. No changes → "(no changes this round)".

2 — Risk audit: frontend-only patches that should go deeper

Core step. For every frontend change, ask: is this covering a backend/DB deficiency?

Before flagging, actually read the backend code (Grep/Read the Request DTO / Controller / Service / enum / entity) to confirm the backend lacks the guard. Backend already guarded = compliant, not a risk. Only flag when the backend genuinely lacks it.

| Frontend changed | Risk | Should land | |-----------------|------|-------------| | Field validation (required/format/length/range) | direct API call bypasses browser → validation absent | DTO @NotBlank/@Pattern/@Size/@Min/@Max; hard constraints + DB NOT NULL/UNIQUE/CHECK | | Enum/code → label (zh↔en) | hardcoded code→中文 map; every consumer rewrites; new values drift | backend enum (@Getter label) → xxxLabel at VO/Response assembly; or dict table | | Default values | direct API call → null | DTO default / Service fallback / DB column DEFAULT | | Uniqueness/dedup (frontend scans loaded list) | races; only current page; not a real constraint | DB UNIQUE uk_xxx + backend catches conflict → BizException | | List filter/sort/paginate (in-memory frontend) | only current page; breaks at scale; unfiltered data sent | backend query params + SQL WHERE/ORDER BY/LIMIT | | Masking/sensitive hide (frontend redacts) | full plaintext still in response, visible in DevTools | backend strips/masks before return, or Response DTO omits field | | Permission/button visibility (frontend hides/disables) | API still callable directly | Sa-Token @SaCheckPermission/@SaCheckRole | | Error messages (frontend fabricates copy for status code) | copy drifts; other clients see raw code | backend ErrorCode Chinese defaultMessage | | Computed/derived fields (frontend computes total/count/status) | each client computes independently, may disagree | backend computes at VO/Response assembly |

NOT a risk (don't false-flag): layout/spacing/colors/animations, pure display formatting (dates as 2026-04-23), interaction details (drawer/focus/toast timing), routing, frontend-owned i18n text — frontend's job by design.

Distinguish intentional design from oversight: if ` or threat_model` states a behavior is by design (e.g. SK plaintext returned once at creation), not a risk.

3 — Interview per flagged risk

No risks → skip to step 4, note "No risks found".

Per risk, AskUserQuestion one at a time (≤4/call; chain calls). Explain: what changed (frontend), why a risk, where it should land (endpoint / Controller.method / VO field / table column).

Options:

1. Agree — mark as backend/database TODO (first, "(Recommended)" when genuine)
2. Frontend is fine — backend already safe / acceptable here
3. Defer — pending discussion

User may use "Other". Record only, do not act.

4 — Output summary (conversation only, Chinese)

## Acceptance Summary — 

### I. Changes This Round
**Frontend** (N changes)
| file:line | what changed | feedback addressed |
**Backend** (N changes / no changes this round)
**Database** (N changes / no changes this round)

### II. Risk Audit (frontend-only patches that should go deeper)
> No risks found — (or list each:)
**Risk 1: **
- Current state: frontend  applied 
- Risk: 
- Should land at: 
- Your decision: 

### III. Follow-up Recommendations
> Items the user "agreed" to address, as a checklist (not dispatched — follow up via /gsd:quick).

Guardrails

• Read and interview only. No Edit/Write, no file writes, no gsd-executor, no commit, no branch switch.
• Read backend code before flagging. Under-report over over-report; uncertain → "Defer — pending discussion".
• One checkpoint per run.
• Reply in Chinese.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: JinsYin
• Source: JinsYin/skills
• License: MIT

Install and usage instructions live in the source repository linked above.