| name | qa-tracker-us-closure |
| description | Per-US test closure cadence for closing out NONE/PARTIAL P0 user stories in a project's QA-TRACKER.md. Class-level, applies to any project that uses the US ↔ Test 對照 tracker pattern (PM-System, crm-system, llm-acp, future projects). Trigger when a project's QA tracker shows rows with NONE / PARTIAL / DRAFT for P0 user stories, or when the user says "close out P0 US test gap", "Sprint N P0 test push", "做嗰啲 NONE 嘅 US", "ship 嗰啲 P0 test by test", or "audit NONE US". Per-US rhythm is audit source, write derive-or-source test, run + fix, commit + tracker sync + push. Distinct from a multi-axis code review (see `code-review-pipeline`) and from full Sprint retrospective writing. |
QA Tracker — Per-US Test Closure Cadence
Close out the P0 user stories that show as NONE / PARTIAL in docs/QA-TRACKER.md, one commit at a time, with a per-US test file and a tracker sync.
Why this skill exists
docs/QA-TRACKER.md (see docs/qa-tracker.md for the format) is the source of truth for which US has which test. When rows show NONE for P0 US, the project can't ship those features with full coverage. Closing them one by one is a discrete, repeatable workflow, NOT a single big bang (which 紅線 12 + 16 forbid for P0 because the per-US test attribution gets lost).
When to use vs. not
| Situation | Use this skill? | Why |
|---|
| Tracker has 5+ P0 US with NONE | ✅ Yes | Per-US closure is exactly the pattern |
| User says "ship all remaining P0 test" | ✅ Yes | Split into per-US commits, audit each |
| One P0 US needs a feature + test together | ✅ Yes | Same per-US rhythm applies |
| User asks "review the codebase" | ❌ No | That's code-review-pipeline (multi-axis) |
| User asks "write retro for Sprint N" | ❌ No | That's the existing retro workflow |
| P1/P2 US gap closure | ⚠️ Yes but lower priority — confirm with user first | |
| User says "補: [ ] ..." 格式 = retro follow-up 段 list 出嚟 | ✅ Yes (docs-only path) | See §Sprint follow-up registration below — pure docs patch, no code/test, register retro "What's NOT done" 段入 tracker |
The 6-step per-US rhythm
For each P0 US, do these in order. Hard rule: one US = one commit + one tracker sync + one push + one revert-detection check. Do NOT bundle multiple US into one commit — David needs the per-US commit attribution for ship-gate audit and the per-US RG entries.
1. Audit the US scope (1-2 tool calls)
Before writing a single test, read the actual source code for that US's route + helpers. Check:
- Endpoint (route file + method), e.g.
backend/src/routes/worklogs.ts:24-62 for US-6.4
- Helper exports — is the logic in an exported helper (
export function ...) or inline in the route?
- Exported helper → import directly in test (more stable, less duplication)
- Inline logic → derive a pure function in the test file that mirrors the source
- Existing tests — read
*.test.ts next to the route file, some US invariants are already partially covered
- Frontend code, e.g.
RichTextEditor.tsx for US-3.5 had a <p></p> → '' invariant worth testing
Audit output goes into the commit message body, NOT a separate file.
2. Decide: derive or source-import
| Helper situation | Use |
|---|
export function already exists in route | import { ... } from './routeFile' |
Helper is const inside route file (not exported) | Derive a mirror function in the test file, link to source line in comment |
Logic spread across route + multiple if branches | Derive a single function that takes the relevant query/user args |
PM-System examples:
tasks.ts exports buildTaskListWhere + resolveTaskProjectId → source-import (commit 3b93dba)worklogs.ts has all logic inline → derive buildWorkLogFilterWhere (commit df8c4dd)requirements.ts accepts description: string from frontend RichTextEditor → derive normalizeRichTextDescription based on the frontend's <p></p> → '' invariant (commit e45626b)
3. Write the tests (target 5-10 cases per US)
Test cases must cover the actual invariant, NOT trivial smoke. Good categories:
| Category | What to test |
|---|
| RBAC gate | "non-admin tries to set userId filter" → ignored. "admin sets departmentId" → applied. |
| Null safety | undefined / null / empty-string input → graceful return, no throw |
| Round-trip fidelity | US-3.5: any HTML in → same HTML out (Tiptap <p></p> → '' is the only exception) |
| Cross-entity guards | US-4.4: multiple linked requirements across different projects → error |
| Composition | "filter A + filter B can both be applied" |
| Edge cases | boundary values (date 5th vs 4th), empty input, single-element array |
Aim for 5-10 cases per US. Fewer = undertested, more = redundant.
4. Run + fix (always run the full backend suite, not just the new file)
bun test runs the whole backend suite in ~3s. ALWAYS run it after writing tests — even if your new test file passes, the suite-wide run catches unintended regressions in shared helpers (computePagination, serializeWorkLog, etc.).
cd backend && bun test 2>&1 | tail -5
If 0 fail + N increased by your test count, proceed to commit. If fail, fix before committing.
5. Commit + tracker sync (5 spots in QA-TRACKER.md, all required)
Commit message format:
test(<area>): US-X.Y <short description> (Sprint N)
N 個新 [derive|source] test 守住 [US-X.Y] 嘅 [<invariant categories>]:
- <bullet 1: invariant + behavior>
- <bullet 2: invariant + behavior>
- ...
US-X.Y 由 NONE → PASS-UNIT[+E2E] (Sprint N: 計數 +N)。
[Optional: full-suite regression result]
QA-TRACKER.md 同步更新 <5 個 spot 嘅具體改動>。
QA-TRACKER.md has 5 spots that must sync per US (紅線 11). Easy to miss one:
- The US row itself — change
NONE → **PASS-UNIT** 🟢 (Sprint N: <N> tests — <summary>) in the comparison table
- Metrics row "P0 US PASS-UNIT only" — increment count
- Metrics row "Unit tests 總數" — update count + add
+N delta
- Sprint history entry (bottom of file) — add a row in the date-sorted table summarizing the US push
- Status header
> **Status**: AND > **Update**: lines — append to the Update line on EVERY commit, not just the first of the sprint. The Update line is cumulative — by sprint end, it tells the full story in one read.
Example Update-line progression (累積 pattern, NOT rewrite):
# After commit 1 (US-6.4)
> **Status**: 🟢 2026-06-10 — Sprint 10 進行中,P0 remaining US test push(US-6.4 已 PASS-UNIT)
> **Update**: 2026-06-10 Sprint 10 — US-6.4 worklogs filter RBAC 由 NONE → PASS-UNIT(9 個 test,non-admin 強制 userId + admin departmentId gate),Unit 549→558(+9)
# After commit 2 (US-3.5)
> **Update**: 2026-06-10 Sprint 10 — US-6.4 ...;US-3.5(requirement rich-text)由 NONE → PASS-UNIT(11 個 test,...);Unit 558→569(+11)
Each commit appends its US closure to the existing Update string (use ;US-X.Y(...) separator). The Status line only changes wording on the first commit of a sprint (e.g. add "Sprint N 進行中") and on sprint closure (change to "Sprint N 收工").
6. Push + revert-detection
After commit:
git push origin master
git log --oneline origin/master..HEAD
git log --oneline -3 origin/master
Per memory entry "Revert detection 鐵律": David has done git revert mid-session in past projects and the assistant never noticed. Always do the 3-second revert check before claiming a US is done.
Sprint follow-up registration (docs-only, no-code) — TRIGGER: "補: [ ] ..." 格式
Trigger signal: User message 開頭係 補:(或 + / add: / 英文 add:)後面 list 住啲 task,通常係 retro doc "What's NOT done" 段嘅 copy-paste(eg PM-System retro 2026-06-09-sprint-11-...md line 75-79)。冇 coding request,只係要將 follow-up register 入 tracker 等下個 sprint 開工嗰陣 agent 開 file 就見到。
呢個係 pure docs-only update,唔入 6-step per-US rhythm — 唔寫 test, 唔 commit code,只 patch docs/QA-TRACKER.md。0 個 code commit,1 個 docs commit。
The 4-step docs-only rhythm
1. 讀 retro doc 嘅 "What's NOT done" 段(0-1 tool call)
如果 user message 入面有 retro 嘅 reference(file path 或 file 內文),讀 docs/retros/<date>-sprint-N-...md 嘅 bottom section(通常叫 "What's NOT done" / "Out of scope" / "Follow-up")。
2. 將每條 follow-up 歸類入 3 個 bucket(0 tool calls,純分類)
| Bucket | 信號 | Tracker 處理 |
|---|
| A: New test follow-up | 「E2E test 喺 X」「補 test」「X spec」 | 對應 US row 嘅 E2E cell 改 ❌(DRAFT Txx, Sprint N planned — <一句解釋>),+ Status 保持 PARTIAL 或 NONE |
| B: New US-level follow-up | 「US-X.Y full-text」「[新 area] search」 | 對應 US row Status 改 NONE-HOLD 🟠(加 hold 嘅理由 + 點解紅線 12 唔適用:P1/P2) |
| C: Refactor / infra follow-up | 「抽 component」「對齊 pattern」 | 唔入 US row, 加去新 section「🟠 Open follow-ups」最後一張 table |
3. Txx 命名(sequence 跟上一個 sprint, 唔跳號)
| Sprint 最後一個 T-ID | 下次 first T-ID | Example |
|---|
| T14h (Sprint 9) | T15a | PM-System Sprint 11 follow-up (2026-06-09) |
| T15x (Sprint 11) | T16a | next sprint |
Rule: 一個 US 嘅多條 follow-up → T15a + T15b + T15c(sub-letter), 唔開 T16 唔同 US,即使係隔離。Sub-task 一定 share 同一個 US row。Example:
- T15a: E2E
ProjectDetailPage bug tab create + rich text + image paste (US-5.6 sub-task 1)
- T15b: E2E
ProjectDetailPage bug tab search filter (US-5.6 sub-task 2)
兩個都入 US-5.6 row 嘅 E2E cell:❌(DRAFT T15a + T15b, Sprint 11 planned — ...) — 唔開 row。
4. Patch 落 tracker(3 spots,5-15 lines 改動)
| Spot | 點改 |
|---|
| 1. US row (Bucket A/B) | E2E cell 加 (DRAFT T15a + T15b, Sprint 11 planned — ...) 或 Status cell 加 NONE-HOLD 🟠 |
| 2. 新增 section「🟠 Open follow-ups」 | 喺 health metrics table 下面加一張 table(ID / US / Owner / Status),list 晒全部 Bucket A + B + C 嘅 follow-up, 等下次開 Sprint 12 個 worker 一打開 tracker 就見到 |
| 3. Status banner | 加一行 > **Update**: <date> <sprint> 收工 — Retro Sprint <N+1> follow-up registration: <一句 summary> |
5個 spot 嘅 full US closure rhythm 唔適用 — 因為冇 code, 冇 test count, 冇 metrics increment。只改呢 3 個 spot。
Example commit message (docs-only, no code)
docs(qa-tracker): register Sprint 11 follow-ups (T15a/T15b/US-10.3 hold)
Retro 2026-06-09 sprint 11 收工 "What's NOT done" 段 3 條 follow-up 全部
register 入 tracker, 等下個 sprint 開工對住做:
- T15a (DRAFT): E2E `ProjectDetailPage` bug tab create + rich text +
image paste — US-5.6 PARTIAL, 覆 ProjectDetailPage 入口(舊
CreateBugModal 5 tests 2026-06-09 skip, 唔可以重新 deprecate)
- T15b (DRAFT): E2E `ProjectDetailPage` bug tab search filter —
US-5.6 PARTIAL sub-task 2
- US-10.3 NONE-HOLD 🟠: Wiki full-text search hold, scope 較大
需要 Postgres tsvector 或 MeiliSearch, 留俾下個 epic 決定
(Sprint 11 client-side title search 已 done in WikiTab,
P1 非關鍵 紅線 12 唔適用)
Patch:
- 1 file / +13 / -2
- 0 code, 0 test
- 1 docs commit
Common pitfalls (docs-only)
- 唔好 default 開 US-5.6 row 2 條 — 用同一 row 嘅 E2E cell 寫
(DRAFT T15a + T15b) 即夠。Multi-row 表 = 噪音, David 一打開 tracker 見到 2 row US-5.6 會以為係唔同 US。
- NONE-HOLD 唔等於 DEPRECATED — DEPRECATED ⚫ = 拎走,紅線都放棄;NONE-HOLD 🟠 = 留低,只係 explicit 「呢個暫時唔做,等將來決定」,status 仍然 active, David 寫新 epic 嗰陣可以 query 「咩 US 係 HOLD 嘅」做 input。
- Status banner 嘅 Update line 用
;US-X.Y() separator 唔 work 喺 docs-only — 因為冇 US closure event, 改用一句 narrative Retro Sprint N follow-up registration: <summary>。6-step rhythm 嘅累積 pattern 唔適用, docs-only commit 開新 Update line。
- 冇 commit 都係 0 個 fix / 0 個 feature — David 嘅「睇唔到 = 冇做」鐵律伸延到「冇 git diff = 冇做」,docs-only commit 仍然要 push + git log 確認 HEAD 喺度。
- 唔好加多過 1 個 "Open follow-ups" section — 整個 tracker 應該得 1 個(喺 health metrics 下面)。如果再 patch 發現 row 重複, grep
Open follow-ups 確認。
Common pitfalls
-
Bundling multiple US in one commit — David loses the per-US attribution for ship-gate audit. Default: one US = one commit. Exception: 2 US sharing the SAME route file + SAME derive pattern (e.g. US-2.4 + US-2.3 both in projects.ts and both derive project helpers) can ship as one commit — the shared describe blocks preserve per-US attribution in the test file even when the commit message groups them. PM-System Sprint 10 commit bfb8618 did exactly this. If unsure, ask the user before bundling.
-
Deriving logic when source is exported — wastes effort, drifts over time. Always check export first.
-
Testing trivial invariants — expect(add(1, 2)).toBe(3) is a smoke test, not a coverage test. The point of these tests is to lock in invariants (RBAC gates, round-trip fidelity, cross-entity guards) so future refactors don't silently break them.
-
Forgetting one of the 5 tracker spots — the sprint history row at the bottom is the one most often missed. The Update-line append is the second most often missed.
-
Not running the full suite — bun test <file> is faster but misses regressions in shared helpers. Run rhythm per commit:
bun test <new_file> — confirm new test file passes (fast feedback)bun test — full backend suite, must end with N pass 0 fail before commit- (E2E US only)
cd e2e && ../frontend/node_modules/.bin/tsc --noEmit --target es2020 --module commonjs --moduleResolution bundler --esModuleInterop --skipLibCheck --strict tests/<new>.spec.ts — TypeScript pre-flight because e2e/ has no own tsc install and npx tsc collides with the wrong wrapper (see SOUL.md npm pitfall).
-
Re-running git pull mid-sprint — fine, but if it brings new changes that touch a route you just audited, re-audit before next US.
-
Derive function null-safety vs source's "caller guards" — when source uses user?.role !== 'admin' and a null user is theoretically reachable, your derive function will still output {members: {some: {userId: undefined}}} for null users (because user?.role is undefined, which !== 'admin' is true). Do NOT write a test that expects "null user → no OR scope" — that's a different function. Two options:
- Mirror source's output (accept the
userId: undefined row) and document "caller must guard null user" in the test comment. The route handler does this upstream (e.g. projects.ts:29 if (!user) return { projects: [] }).
- Add an explicit null guard at the top of the derive function (
if (!user) return {}) and document "derive adds a defensive guard the route already had" — only do this if you're also patching the source.
PM-System Sprint 10 commit bfb8618 US-2.4 hit this — the test had to be rewritten after a 1-fail run, expecting {members: {some: {userId: undefined}}} and noting "caller 必須 guard 返". This is correct behavior, not a bug.
-
E2E spec for frontend-heavy US without dev stack — some US have no cheap unit boundary (drag-drop Kanban, multi-modal editor flows). The user may opt to ship a .spec.ts file with TypeScript clean + 5/6 tests that run on real stack + 1 placeholder (test.skip) for the drag-drop interaction that needs keyboard-accessible rewrite. This is a valid PASS-E2E status — the spec exists, the file compiles, and the API invariant is locked down. The drag-drop rewrite is a follow-up sprint task. PM-System Sprint 10 commit eaae48a US-4.5 used this pattern; the file is e2e/tests/project-kanban.spec.ts (6 tests, 5 functional, 1 drag-drop skip).
Multiple US in one commit (rare but legitimate)
When 2 US share the same source file + same derive pattern and the user's preferred commit cadence allows it, bundle them. PM-System Sprint 10 commit bfb8618 did this for US-2.4 + US-2.3, both in projects.ts — the derive pattern is identical, the test categories are distinct, and shipping them together saved 1 round of full-suite + revert check. Always tell the user this is a bundled commit in the commit message body.
Session start checklist (when triggered)
cd <project>; git pullgrep -E '^\|.*NONE' docs/QA-TRACKER.md | head -20 — list of US to close- Ask user: "逐個 / 一次過開 5-6 個 sprint? (default: 1 sprint, 5-7 US, P0 first)"
- For each US: audit → test → commit → tracker sync → push → revert check
- Sprint closure: update Sprint history entry, mark Sprint N as "Sprint N closure", next sprint in "in progress"
Reference
references/pm-system-sprint-10-example.md — worked example of 3 US closures from PM-System Sprint 10 (2026-06-10), with audit decisions, test categories, tracker sync spots, and pattern observationsreferences/pm-system-sprint-10-commits-4-5.md — Sprint 10 commits #4 + #5 (US-2.4+US-2.3 bundled, US-4.5 E2E spec draft). Covers derive null-safety pitfall, multiple-US-per-commit rare exception, E2E spec draft pattern, TypeScript pre-flight for e2e/ specs.references/sprint-followup-registration-example.md — Sprint 11 (2026-06-09) docs-only follow-up registration: 3 條 retro "What's NOT done" 段 → tracker 嘅 2 row patch + 1 new section + T15a/T15b sub-letter naming + NONE-HOLD 🟠 marker。When to use: 用戶 message 開頭係 補: / +: 格式 list 住嘅 task,純 docs-only,冇 coding intent。
Related skills
code-review-pipeline — for multi-axis review before ship (different cadence)regression-guard — for any bug fix that ships (per-US test push rarely hits this; only if a test reveals a real bug)tech-debt-register — if the audit reveals debt that doesn't fit a US closure (e.g. missing migration)context-summarizer — Sprint 10+ long sessions, decisions get auto-summarized; per-US commit messages are the durable recordinterruption-recovery — pair with the docs-only rhythm if David 收工前要中途停(register 咗 1-2 條 Txx 然後 /new,下個 session 用 load_state 恢復)
