feature-workflow

name	feature-workflow
description	Per-issue SDD cycle: resolve a PM issue to its TASK spec, plan a BUILD/REVIEW pair, build per FRONTEND_RULE_COMMON, review via the review skill.

Feature Workflow

Activation Contract

Use this skill when

• You are given a GitHub issue number or link from asgard-ai-platform/asgard-sdk-pm
• You need the full per-issue SDD cycle: resolve → plan → build → review → done

Do NOT use for

• Small fixes with clear scope that do not trace to a PM issue (edit the file directly)
• Review-only runs (invoke /review BUILD-NNN directly instead)
• Issues already planned — check requirements/tasks/_index.md first

---

How it works

This skill is the per-issue orchestrator for the asgard-js-sdk SDD process. It operates in spec-only mode always (no running design-reference server, no screenshot comparison). The prototype reference under references/asgard-chat-kit-prototype is visual / behavioral background only — do not treat it as the implementation spec.

Lifecycle: Resolve → Plan → [PAUSE] → Build → Review → Done

No auto mode. The skill always pauses after Plan to present the task pair(s) and wait for explicit user confirmation before any implementation begins.

---

Stage 1: Resolve

Goal: Locate and read the PM TASK spec for the given issue.

Step 1.0 — Refresh design references（每次開工先執行）

在讀取任何 spec 之前，先把 PM spec 與 prototype/design submodule 更新到遠端最新，確保對到的是最新設計：

# 逐個把 submodule 更新到遠端最新（一個失敗不連坐其他）
git config -f .gitmodules --get-regexp '^submodule\..*\.path$' | awk '{print $2}' | while read -r m; do
  git submodule update --init --remote "$m" || echo "FAILED: $m"
  git -C "$m" submodule update --init --recursive || echo "FAILED (nested): $m"
done

# 驗收（必跑）：確認上面沒有 FAILED，且每個 submodule 都有 SHA
git submodule status

不要圖快改用 git submodule update --init --recursive --remote references。 --remote 需要知道要追哪個分支，.gitmodules 沒設 branch 時會 fallback 到該 clone 的 origin/HEAD；巢狀 submodule 常常沒有 origin/HEAD，會噴 fatal: Unable to find refs/remotes/origin/HEAD revision in submodule path ...，而 git 遇到這個是直接中止整趟迭代——排在後面的 submodule 靜默沒被更新，你會拿舊 spec 開工。

跑完不要再補一條不帶 --remote 的 git submodule update。 那條的語意是「把每個 submodule reset 回父 repo 記錄的 pin」，會把剛才 --remote 的成果整個倒退回去。巢狀 submodule 已經由上面 git -C "$m" ... 那行補齊了。

驗收怎麼看：git submodule status 的 + 前綴代表 checkout 已超前父 repo 的 pin（--remote 生效）；沒有 + 代表 pin 本來就是最新——兩者都正常。- 前綴才是沒初始化。

註：--remote 會把 submodule checkout 到遠端最新，父 repo 的 pin 會顯示為已變更。這個 pin bump commit 進 feature 分支完全無妨、也不影響開發——references/ 只是背景參考（不被 app 編譯，實作以 distill 進 requirements/ 的內容為準），且 CI 未開 submodules: true。因此就讓 pin 始終浮到最新、bump 自然留在分支即可，不必刻意避免 commit。

Step 1.1 — Fetch the issue

gh issue view <n> -R asgard-ai-platform/asgard-sdk-pm --json title,body,labels

Inspect the body field for a tracking reference — typically under "相關票" or similar heading. The reference is a path like tracking/asgard-js-sdk/tasks/TASK-NNN-<slug>.md (or a features/F-NNN-*.md / use-cases/UC-NNN-*.md path).

Step 1.2 — Read the TASK spec

Read references/asgard-sdk-pm/<that-path>.

If the file is still not present after Step 1.0's refresh, search by TASK number or title:

find references/asgard-sdk-pm/tracking/asgard-js-sdk -name "*NNN*.md" 2>/dev/null
# or search by keyword:
grep -rl "<keyword-from-title>" references/asgard-sdk-pm/tracking/asgard-js-sdk/ 2>/dev/null

Step 1.3 — Resolve Gate

• Issue title, labels, and body retrieved
• TASK / feature / use-case spec file path confirmed and file read
• Acceptance criteria or behavioral requirements identified from the spec

If the spec references other upstream specs under references/asgard-sdk-pm (features ↔ use-cases), read those too as background. Do not implement from references/ directly — the spec must be distilled into requirements/tasks/ files first.

---

Stage 2: Plan

Goal: Distill the spec into one BUILD + REVIEW pair (or several pairs for large features), create the task files, and PAUSE for user confirmation.

Step 2.1 — Determine task numbering

Find the next BUILD number from existing files:

ls requirements/tasks/BUILD-*.md 2>/dev/null | sort | tail -1

If no BUILD tasks exist yet, start at BUILD-001. Otherwise increment the highest number found by 1.

Step 2.2 — Decide how many BUILD tasks

Default: one BUILD + one REVIEW pair per issue.

Split into multiple pairs only when the issue covers two or more clearly separable capabilities (e.g. a core-only stream change AND an independent react template) that cannot reasonably be reviewed together. When in doubt, keep it one pair.

Step 2.3 — Create BUILD-NNN.md

Copy requirements/tasks/_build_template.md to requirements/tasks/BUILD-NNN-<slug>.md.

Fill in these fields (remove template HTML comments after filling):

• Task ID: BUILD-NNN
• Status: draft
• Issue: the GitHub issue URL (https://github.com/asgard-ai-platform/asgard-sdk-pm/issues/<n>)
• Source spec: the references/asgard-sdk-pm/tracking/asgard-js-sdk/<path> spec path
• Complexity: S (single function / small change) | M (standard client method / template) | L (cross-package flow or streaming/protocol change)
• Brief: 2–5 sentences — the developer-visible outcome and the changed packages / modules / components; note anything that already exists to be reused or extended.
• Relevant Rules: the pre-filled table from the template (sourced from FRONTEND_RULE_COMMON.md) — do not delete rows; add extra rows if this task involves patterns not in the default table.
• Acceptance Criteria: translate the spec's acceptance conditions into EARS R# identifiers. Every R# must be independently verifiable. The last R# is always a build + demo/test smoke check.
• Implementation Tasks: ordered T# steps mapped to R#; include npm run lint:packages, npm run format:check, and npm run build:core && npm run build:react in the pre-done step.
• Coverage — both Use Cases and Files: leave as [filled during build].
• Execution Log: add YYYY-MM-DD: BUILD task created from <issue URL> (Status: draft).

Step 2.4 — Create REVIEW-NNN.md

Copy requirements/tasks/_review_template.md to requirements/tasks/REVIEW-NNN-<slug>.md.

Fill in:

• Task ID: REVIEW-NNN
• Status: draft
• BUILD Task: BUILD-NNN
• Leave Reviewed commit and Reviewed branch blank — filled at review time.
• Execution Log: add YYYY-MM-DD: REVIEW task created, paired with BUILD-NNN (Status: draft).

Step 2.5 — Register in _index.md

Add both tasks to the Task Queue in requirements/tasks/_index.md with Status draft.

Step 2.6 — PAUSE and present plan for user confirmation

Mandatory stop. Do not begin Stage 3 until the user explicitly confirms.

Present:

計畫完成 — 請確認後再開始實作。

Issue: <title> (#<n>)
Source spec: references/asgard-sdk-pm/<path>

Task Queue (新增):
| Task ID    | Title          | Complexity | Status |
|------------|----------------|------------|--------|
| BUILD-NNN  | <title>        | S/M/L      | draft  |
| REVIEW-NNN | Review: <...>  | —          | draft  |

Acceptance Criteria (BUILD-NNN):
- R1: When ..., the system shall ...
- R2: ...
- RN: (Smoke check) When the developer runs `npm run build:core && npm run build:react`
      and exercises <feature> in the react-demo (http://localhost:4200) / Vitest,
      the system shall <expected observable behavior> with no build errors.

有問題或需要調整（例如拆成多個 task、修改 R#）請告知。確認後我開始實作。

After user confirms, update BUILD-NNN.md Status: draft → ready. Update _index.md to ready. If no other task is currently in-progress, advance ## ▶ Next Task to BUILD-NNN.

---

Stage 3: Build

Goal: Implement per FRONTEND_RULE_COMMON.md; fill ## Coverage; keep _index.md status current.

On user go

1. Update BUILD-NNN.md Status: ready → in-progress. Update _index.md.
2. Read .claude/skills/feature-workflow/FRONTEND_RULE_COMMON.md in full before writing any code. The ## Relevant Rules table in BUILD-NNN.md is a distilled reference, but the full corpus catches edge cases.
3. Work through ## Implementation Tasks in order, checking off each T# as completed.
4. Implement per FRONTEND_RULE_COMMON.md §1–§7:
   • Types / enums in packages/core/src/types + constants before first use (§2.3)
   • Keep the package boundary: @asgard-js/core has no React/DOM deps; @asgard-js/react imports core via its public entry only (§1.6 / §2.1)
   • No any, no @ts-ignore / eslint-disable to bypass type errors (§1.1–§1.2)
   • Clean teardown for every RxJS subscription / EventSource / timer (§1.5)
   • Public API changes go through @deprecated, not removal (§1.7); export new API from the package entry (§2.2)
   • React components: typed props, theming via CSS variables (§4.1–§4.2)
5. Fill ## Coverage in BUILD-NNN.md:
   • Use Cases: list R# identifiers satisfied (or — if purely structural with no behavioral validation)
   • Files: list every .ts / .tsx file created or modified (note the package)
6. Run static checks:

npm run lint:packages
npm run format:check
npm run build:core && npm run build:react

7. Smoke check: build core+react, then exercise the feature — run relevant Vitest, and/or npm run serve:react-demo (http://localhost:4200) and walk through all R# (e.g. /templates). Take a screenshot to .github/screenshots/ if the change is visually significant.
8. Update BUILD-NNN.md Status: in-progress → done. Fill Execution Log. Update _index.md.
9. Update REVIEW-NNN.md Status: draft → ready. Update _index.md.

Blocking rule

若 build 過程有需要人為判斷、且無法從 spec 或 FRONTEND_RULE_COMMON.md 決定的事項，立即停下來問，不要自行假設。

---

Stage 4: Review

Goal: Run the review skill on BUILD-NNN; route findings back to build; loop until all findings pass.

Step 4.1 — Invoke the review skill

/review BUILD-NNN

The review skill runs §1 Static Code Review and §3 Functional Validation (per REVIEW_RULE.md) and writes findings to REVIEW-NNN.md. It reads BUILD-NNN.md ## Coverage to determine scope.

註：若該 BUILD task 的 Coverage.Use Cases 為 —，review skill 會跳過 §3 功能驗收，REVIEW-NNN 仍可達 done。

Step 4.2 — Route findings

On review skill completion:

• Zero BLOCKERs → proceed to Stage 5.
• §1 violations (static rule breach): fix each violation in source, re-run npm run build:core && npm run build:react + npm run lint:packages, then invoke the review skill again on BUILD-NNN.
• §3 failures (functional): describe [actual behavior] vs [expected behavior].
  • If fixable by implementation change → fix in source, re-run test / demo, invoke the review skill again.
  • If the gap requires a spec change or new R# → stop and ask the user before touching any file.

Repeat the loop until the review skill reports no BLOCKERs and marks REVIEW-NNN done.

---

Stage 5: Done

Goal: Close the cycle only after explicit user authorization.

After the review skill marks REVIEW-NNN done:

1. Present summary:

Review 通過 — 等待授權完成此 cycle。

BUILD-NNN: <title> — done (§1 ✅ / §3 ✅)
REVIEW-NNN: done

確認後我更新 _index.md 並標記本 cycle 完成。

2. Wait for explicit user authorization.
3. On authorization: confirm both BUILD-NNN and REVIEW-NNN are done in _index.md. Advance ## ▶ Next Task to the next ready task, or set to None — awaiting task selection if none remain.

---

State and Status Rules

• Mode: always spec-only. No running design-reference server, no screenshot comparison against a reference build. The prototype at references/asgard-chat-kit-prototype is visual / behavioral background only.
• Status flow: draft → ready → in-progress → done
• Forbidden status form: never write in_progress (underscore). Always use in-progress (hyphen).
• The Status field in a task file and the matching row in _index.md Task Queue must always be in sync.

---

When to Stop and Ask

Situation	Stage
The tracking path is absent from the issue body and find returns nothing	1
The spec has no identifiable acceptance criteria or behavioral requirements	1
User requests task split / merge / R# change after seeing the plan	2
A build decision requires human judgment not covered by the spec or FRONTEND_RULE_COMMON.md	3
A §3 failure requires a spec change rather than an implementation fix	4
User has not yet explicitly authorized Done	5