dgk-dev/dgk-gpt/test

Test

Use this as a verification-and-fix mode after implementation or during bug hunts. This is not just "run one test command". It is a persistent loop that keeps narrowing failures and rerunning checks until the requested behavior is verified.

Default Loop

1. Define the verification target.

• affected files, routes, surfaces, and user-visible claims
• exact stopping condition for signoff

1. Build the smallest credible test matrix from the repo and task.

• code-level checks first
• integration and e2e next
• live runtime QA last

1. Run the relevant existing automated checks before inventing custom ones.

• prefer repo-native scripts and targeted specs or files
• use the stack already present: Vitest, Jest, Playwright, Cypress, pytest, go test, etc.
• widen scope only after the narrow failing checks pass

1. If the surface is interactive, verify the real runtime too.

• web: run relevant e2e flows or drive the app directly
• before browser DevTools inspection, probe http://127.0.0.1:9333/json/version
• if it is already live, reuse it
• if it is not live and a chrome launcher exists, start it automatically, wait for the port, then use chrome-devtools
• browser diagnosis: use chrome-devtools for console, network, DOM, accessibility, or performance inspection
• before Pix desktop DevTools inspection, probe http://127.0.0.1:9334/json/version
• if it is already live, reuse it
• if it is not live and a tauri-pix launcher exists, start it automatically, wait for the port, then use tauri-devtools
• Tauri desktop: use tauri-devtools and the fixed Tauri debug path when diagnosis is needed
• direct browser control: use the repo's Playwright setup or a persistent js_repl Playwright session when targeted manual proof is still needed

1. On failure, fix and rerun.

• repeat inspect -> patch -> rerun narrow checks -> rerun broader confidence checks
• do not stop after the first or second failed attempt

1. Before signoff, rerun the checks that support the final claims.

Priority Rules

• Existing repo tests beat ad hoc manual clicks.
• Narrow targeted checks beat whole-suite runs unless the task truly warrants broader coverage.
• Browser or desktop verification complements automated tests; it does not replace them.
• When the repo has both unit-level tests and Playwright e2e, usually run the relevant unit or component checks first, then the relevant Playwright flow, then live diagnosis only if something is still unclear.
• When the required DevTools launcher exists, do not wait for the user to open it manually; check first, auto-launch if absent, and continue.
• If the user implies "I am stepping away" or "keep going until it is clean", keep iterating without asking after each failure.

Good Targets

• "/test"
• "나 점심 먹고 올 테니 테스트랑 검증 끝까지 돌려"
• "버그 없어질 때까지 계속 확인해"
• "회귀까지 보고 진짜 끝내줘"
• "웹, 모바일, Tauri까지 확인하고 막아줘"

Avoid

• Do not stop at "probably fixed".
• Do not run the whole monorepo matrix by reflex.
• Do not skip existing automated tests just because manual QA looked okay.
• Do not claim success if any required surface was not actually checked.
• Do not keep brute-forcing when a real external blocker is the limit; name it clearly.

Finish

Return:

• verification target
• automated checks run
• runtime surfaces checked
• pass or fail
• remaining blockers or residual risk