Run any Skill in Manus with one click

dev-task-memory

Stars2

Forks0

UpdatedJune 19, 2026 at 03:29

5-layer persistent memory for in-progress dev tasks — survives context compression, /new, gateway restart. State file + git checkpoints + external memory + cross-session search.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

freedomw1987

freedomw1987/tree_monstor

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

File Explorer

15 files

SKILL.md

readonly

Dev Task Memory

解決 David 2026-06-06 hang fix 嘅延伸問題：context 處理完之後, dev task 點樣唔好被遺忘。 5-layer architecture 確保 task 嘅 decisions、current state、next steps 全部 persist 喺 file system, 唔靠 LLM memory (會被 compression 清), 都唔靠 user 記住 (不可靠)。

5 個 Layer

┌─────────────────────────────────────────────────────────────┐
│  Layer 1: TRIGGER (決定幾時 save)                            │
├─────────────────────────────────────────────────────────────┤
│  • 紅線 21-22 (long task, > 50 turns)                       │
│  • auto-compression event (gateway/run.py hook)             │
│  • gateway restart                                          │
│  • /new slash command                                       │
│  • User explicit "save state" / "checkpoint"               │
│  • **Subagent 做到一半 stop 嘅 WIP** (新 trigger — 2026-06-06)│
│    → working tree 有 untracked / modified files, 跟住用    │
│      `references/subagent-wip-pickup-recipe.md` 嘅 6-step   │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 2: STATE FILE (Single Source of Truth)               │
├─────────────────────────────────────────────────────────────┤
│  ~/www/<project>/docs/_meta/dev-task-state.md:             │
│  - Goal, Decisions (with WHY), Current state, Next steps,   │
│    Risks, Insights, Session lineage, References             │
│  - Gitignored via .gitignore snippet                        │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 3: GIT-TRACKED (可審計 + 可 rollback)                  │
├─────────────────────────────────────────────────────────────┤
│  • enable checkpoints: true (Hermes built-in)               │
│  • /rollback N for file rollback                            │
│  • git commit footer 包含 state summary                     │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 4: EXTERNAL MEMORY (cross-session fact recall)        │
├─────────────────────────────────────────────────────────────┤
│  • holographic (local, no API key) or mem0/honcho (API)     │
│  • Sync facts (decisions, insights, risks) 落 external store │
│  • Cross-session: 「之前你話用 X 點解?」即時 recall         │
│  • 當冇 external provider, fallback 落                      │
│    ~/.hermes/memories/dev-task-facts.jsonl                  │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 5: CROSS-SESSION SEARCH (session_search FTS5)        │
├─────────────────────────────────────────────────────────────┤
│  • 開新 session 自動 search "what was I working on"        │
│  • 注入 top-3 relevant past sessions                        │
│  • 用 hermes sessions list 找 candidate sessions           │
│  • 唔需要 manual session ID 記憶                            │
└─────────────────────────────────────────────────────────────┘

🔧 工具腳本

Script	用途	觸發時機
`scripts/save_state.py`	Save 當前 dev task state 落 `docs/_meta/dev-task-state.md`	Layer 1 trigger fires 佢
`scripts/load_state.py`	Read state file + 注入 context (for resume)	New session / /new / after-compression / after-restart
`scripts/sync_external.py`	Extract facts → push to external memory (or local fallback)	每次 save_state 完之後 fire

📂 支援 References

File	觸發時機
`references/subagent-wip-pickup-recipe.md`	Resume 一個 subagent 做到一半停咗嘅 task — working tree 有 uncommitted WIP, 跟住 6-step recipe: snapshot → commit history → read untracked → audit side effects → finish → verify-before-declaring-done。涵蓋 duplicate gitignore entries、batch-fix 嘅 anti-pattern、revert detection 等。
`references/recovery-template-limitation.md`	Day 11 lesson (2026-06-09): `recovery.sh` + `save_state.py` 嘅 generic template 永遠唔 fill 真正 state(Decisions / Files / Next Steps / Risks section 全部係 placeholder)。Resume 之前一定要手動 overwrite `dev-task-state.md`,fill 真正 state。Workaround recipe + root cause analysis + future fix blueprint 全部喺入面。
`references/working-tree-wip-detection.md`	Day 11 lesson (2026-06-09): Agent 同 David 共用 worktree 嘅時候,David 嘅 WIP(modified + untracked) 唔可以入 agent 嘅 commit。Selective `git add <files>` recipe + 3 條新紅線(34-36)+ Day 11 Phase 1 嘅 case study(crm-system 5M + 3?? WIP detection)。
`references/plan-execution-deviation-protocol.md`	Day 14.7 lesson (2026-06-07): Plan JSON 內部矛盾嘅處理 — 揀 primary intent 唔揀 literal text、3 處冗餘紀錄(code comment + commit + report)、anti-pattern(用戶不問照 ship、靜靜 rewrite plan)。Step 6 `/settings` 案例做 worked example。
`templates/handoff-doc-scaffold.md`	Day 14 lesson (2026-06-07): Human-readable handoff doc 嘅 scaffold,用於 `/new` / 收工 / 「之後呢」trigger 嗰陣,Agent 寫一份 David + fresh session agent 兩邊都讀得順嘅 commit-able doc(同 state file 互補)。Section structure 見 §5 嘅 recipe。

📖 標準使用流程

1) 開新 dev task (Plan → Build 開始)

# 進入 project directory
cd ~/www/crm-system

# 立即 save 一個 fresh state
python3 ~/.hermes/profiles/developer/skills/dev-task-memory/scripts/save_state.py \
    --project crm-system \
    --goal "實作 Companies 編輯頁加聯繫人 sub-row" \
    --trigger "task-start"

Agent 收到 trigger 應該:

立即 call save_state.py
Fill in docs/_meta/dev-task-state.md 嘅 Decisions、Next Steps、Insights section
Call sync_external.py push facts 落 external memory

2) Long task 中段 (每 30 min 或每 10 個 tool calls)

Agent 自動 trigger:

python3 scripts/save_state.py --project crm-system --trigger "auto-mid-task"
python3 scripts/sync_external.py --project crm-system

呢個確保就算 hang / compression / restart, state 都唔會 lost。

3) Context 壓縮 / Hang fix / /new / Restart 後恢復

User send "繼續" / "resume" / "where were we":

python3 scripts/load_state.py --project crm-system --search-sessions

Agent 收到 output 後:

Read state file content
Inject 入 conversation context
Resume 跟住 "Next 3-5 Steps" 繼續做

3b) Subagent WIP Pickup (新 — 2026-06-06)

Working tree 有 untracked + modified files 由前一個 subagent 留低, 唔係 from compression。唔好 call load_state.py (冇 state file, 從未 start), 唔好 re-delegate (浪費 70-90% 嘅 WIP)。跟 references/subagent-wip-pickup-recipe.md 嘅 6-step recipe:

git status + git diff --stat snapshot working tree
git log --oneline -10 check partial progress
Read untracked / modified files in full
Audit side effects (.gitignore duplicates, package.json, etc.)
Finish from actual pickup point
git log --oneline origin/main..HEAD empty + clean working tree 為止

4) 任務完成

刪除 state file (task done, no need to resume):

rm ~/www/<project>/docs/_meta/dev-task-state.md

或者 archive 落 docs/retros/YYYY-MM-DD-<task-name>.md。

5) Human-Readable Handoff Doc (Day 14 lesson — 2026-06-07)

When: User 話「之後呢」/「我休息下」/「你 stop 啦」/「/new 啦」,Agent 知道下一個 session 會係 David 自己(或者 fresh agent)讀,需要一份 人類可讀 + 可 commit + 可 grep 嘅文件。

唔可以只靠 state file:

dev-task-state.md 係 gitignored (紅線 28) — 下一個 session 嘅 fresh agent 唔知有呢個 file
內容係 machine-optimized,David 讀起上嚟唔順
冇 commit 嘅話,different branch / different machine 嘅 agent 攞唔到

Handoff doc 嘅特性(同 state file 對比):

特性	`docs/_meta/dev-task-state.md` (state)	`docs/retros/YYYY-MM-DD-<task>-handoff.md` (handoff)
Gitignored	✅ 紅線 28	❌ 正常 commit(可以 PR 一齊)
Reader	Machine + Agent	David + 新 session agent + reviewer
Format	Structured (Goal/Decisions/Next Steps)	Narrative + TL;DR + resume block
Content	Decisions with WHY, current step	Decisions, uncommitted WIP (App.tsx diff!), key env vars, first-3-commands
Lifetime	任務期間	任務結束後 audit trail(永遠保留)
Use case	`load_state.py` resume	`/new` 後手動 read

Recipe(2026-06-07 crm-system handoff 試過 work):

# 1. 收集 state(15 lines)
cd ~/www/<project>
git log --oneline -8 <branch>
git status --short
git diff --stat <base>..<branch> | tail -5

# 2. 寫 handoff doc (見 templates/handoff-doc-scaffold.md)
# 結構: TL;DR + Branch State + What's Done + What's NOT Done + Key Decisions
#       + Uncommitted Work + Critical Context + First 3 Commands

# 3. Save state(獨立, 唔係 handoff 嘅 substitute)
python3 scripts/save_state.py --project <name> --goal "..." --trigger task-start
python3 scripts/sync_external.py --project <name>

# 4. Verify 兩個 output 存在
ls -la docs/_meta/dev-task-state.md docs/retros/<date>-<task>-handoff.md

# 5. 印 resume command 畀 David(≤ 5 行)

Critical sections that MUST be in handoff doc(David 試過 miss 嘅):

Uncommitted WIP — git diff 結果(紅線 30 同 recovery 同理)
David 嘅 untracked files — 永遠唔可以「清理」,必須 list 出嚟
Key env vars / 32-char gate — 例如 JWT_SECRET 32-char hard-fail,新 session 要 re-export
Latent bugs fix 嘅 caveat — 例如 seed Role/RolePermission 喺 prod 唔存在,新 endpoint 會 403
First 3 commands — cd + load_state.py + start dev
Routing / runtime invariants (新 — 2026-06-07 lesson) — 當 step commit 包含 order-dependent routing(sibling route conflict 嗰類)但 smoke test 留到 ship step,必須喺 handoff doc 開獨立 sub-section,寫低「聲稱嘅行為 + 點 verify」。TS pass ≠ runtime pass。詳見 react-router-sibling-route-conflict skill。

Pitfall: 唔好將 handoff doc 寫到 docs/_meta/,會被 gitignore + 失去 audit trail 嘅 purpose。 Pitfall: 唔好 overwrite 已經有嘅 docs/retros/YYYY-MM-DD-<task>.md retro file — handoff 係中斷嘅 trail,retro 係完成嘅 trail,兩者 timing + content 唔同。

🔄 同 Hermes 內建功能嘅整合

Hermes feature	dev-task-memory 用法
`hermes --resume <id>` / `-c "<name>"`	state file 嘅 `Session Lineage` section 寫低 resume command
`hermes checkpoints` (filesystem)	Layer 3 自動 enable, 唔需要額外 config
`hermes memory status`	Layer 4 查 active provider
`session_search` (FTS5)	Layer 5 `load_state.py --search-sessions` 用佢
`/rollback N`	Layer 3 file rollback 配 state file restore

⚠️ 紅線 (跟 SOUL.md 紅線 24-28 對齊)

紅線 24:每次 long task 開始必須 save_state.py, 唔可以 rely on LLM memory
紅線 25:每 30 分鐘或 10 個 tool calls 必須 re-save (Layer 1 自動 trigger)
紅線 26:Decision 必須有 WHY — 唔可以淨寫 "Use Hono", 要寫 "Use Hono 4x 細, edge 啱用"
紅線 27:Resume 時必須先 load_state.py, 唔可以假設自己記得
紅線 28:State file 唔可以 commit 落 git — docs/_meta/* 必須喺 .gitignore
紅線 33:Resume 之前必 check dev-task-state.md 唔係 generic template(Day 11 lesson)。 Generic template 嘅 signals:有 <Decision 1> / <Risk 1> / **待填寫** / 全部 file 都係 M (now) 冇真實 file path。Hit 任何一個 → 手動 overwrite 真正 state 先 resume。詳見 references/recovery-template-limitation.md 嘅 workaround recipe。
紅線 34:Sibling subagent 並行 edit 同一 file 時必 check 雙方最後 diff 唔 overlap(Day 14.7 lesson)。 Patch tool / write_file 撞到 warning "modified by sibling subagent" / "last read with offset/limit pagination" → 唔好盲改。Recipe: (1) git diff <file> 睇 sibling 改咗邊段; (2) 自己嘅 patch 區段同 sibling 嘅 patch 區段唔可以重疊; (3) 如重疊(regex region 撞共同 functions / interfaces),手動 merge 兩邊嘅改動,唔好 git add + commit 直接撞; (4) Commit message 寫低「conflict-free with sibling X 嘅 Y 改動」。詳見 references/sibling-subagent-concurrent-edit-recipe.md(待建)。
紅線 35:Tying two patches to the same commit when they're in different sub-trees is safe; the dangerous case is two patches both touching the same function body or interface declaration. Quick test before committing sibling-collision files: git diff <file> | grep -E "^[+-]" | sort | uniq -c | sort -rn | head -20 — any line that appears as both + and - is a true merge conflict, not just proximity.

🛠️ 設定 (.gitignore snippet)

每個 project 嘅 .gitignore 加:

# dev-task-memory — runtime state, not source
docs/_meta/
dev-task-state.md

🧪 驗證指令

# 1. Save state
python3 scripts/save_state.py --project crm-system --goal "Test" --trigger "manual"

# 2. Sync to external
python3 scripts/sync_external.py --project crm-system

# 3. Read back
python3 scripts/sync_external.py --read --project crm-system

# 4. Search
python3 scripts/sync_external.py --search "Prisma"

# 5. Load (resume)
python3 scripts/load_state.py --project crm-system --search-sessions

📚 配合其他 skills

context-summarizer (existing): 自動每 30 分鐘壓 context。dev-task-memory 補佢嘅缺點 — 壓咗之後 decisions 唔會 lost
regression-guard (existing): 防舊 bug 翻發。state file 嘅 Risks section 配 RG-XXX
auto-doc-gen (existing): 自動生成 API doc。state file 嘅 Decisions 配 doc rationale

🐛 已知限制

save_state.py 嘅 "extract decisions from session" 而家係 stub — TODO: integrate with hermes_state.py 讀真正 session DB
External memory provider (mem0/honcho) API call 仲未 implement — 而家 fallback 落 local jsonl
Holographic provider 雖然係 local-only 但需要 config, 默認未啟用 — 可以之後 setup
State file 唔會 auto-update — 必須人手 call save_state.py (將來可以做 background curator 自動 trigger)
memory tool round-trip drift guard (2026-06-19 lesson) — Hermes 嘅 memory(action='add') 會 refuse write 當 ~/.hermes/profiles/developer/memories/MEMORY.md 有 manual edit / shell append / patch tool 嘅 content 唔可以 round-trip through memory() call。Detection: error message 提 "MEMORY.md.bak." + "Resolve the drift first — either rewrite the file as a clean §-delimited list of entries, or move the extra content out"。Resolution recipe:
1. cat ~/.hermes/profiles/developer/memories/MEMORY.md.bak.<timestamp> 拎 backup
2. Manual integrate missing entries into clean §-delimited list (one at a time via memory(action='add', content='...'))
3. Remove or rewrite original MEMORY.md to clean state
4. Retry memory(action='add')
唔可以 force overwrite — guard 嘅存在係防止 silent data loss (Hermes issue #26045)。唔可以 retry 5 次(浪費 token + 撞同一個 error)。Lesson codification: 如果個 lesson 必須 persist 但 memory tool 拒絕,fallback 落 skill(SKILL.md body)或者講畀 David 知要人手 sync。

🆕 Plan Z — Mid-Handoff Single-Step Commit Cycle (2026-06-07)

Trigger: User has multi-step work (≥5 steps) remaining, session has hit 紅線 22 (>50 turns) or 紅線 19 (tool-call noise from earlier hang). 3 paths to offer, ordered by safety:

Path	Description	Hang risk	/new cycles	Total wall time
X (Inline all)	Do every step in this session, accept hang	High (≥1 hang mid-run)	0 (until forced)	4-5 hr + 1 unplanned handoff
Y (Inline half)	Do first N/2 steps, then /new handoff	Medium	1 planned	4-5 hr, 1 clean checkpoint
Z (1-step cycles)	Do exactly 1 step + commit + handoff + /new, repeat	~Zero per cycle	1 per step	4-5 hr, multiple clean checkpoints

User pick pattern observed: David, when offered Z, replied "Z" within 1 turn. The deciding factor was proven hang history that session (今朝 6 點 3 小時 hang, 47-78 API calls stuck). When user has been burned by hang already, they prefer many small clean cycles over inline sprint.

Plan Z execution recipe (tested on crm-system 2026-06-07 Step 5):

Scope the next single step only — read handoff doc, pick the smallest step with concrete file paths (e.g. "Step 5: routes + API client", not "Step 5-8: all frontend").
Execute in ≤20 tool calls — hard budget. If you cross it, stop and handoff what you have.
Verify locally — tsc --noEmit, git status --short, type check. Defer browser/E2E smoke to the final ship step (Step 12 of the original plan).
Commit + push — single commit, descriptive footer pointing to next step.
Update handoff doc + state file — re-stamp the "What's Done" section with the new commit, mark the step as done in the "What's NOT Done" table.
Emit "📍 progress" + explicit /new recommendation — let user run /new themselves. Don't auto-/new (user has the shortcut key).
Tool-call budget enforcement: if a step needs >20 calls, split it across 2 cycles BEFORE starting, not mid-cycle.

Pitfall — routing conflict undetected (saw in 2026-06-07 crm-system Step 5): When adding sub-routes that share a prefix with a legacy top-level route (e.g. /settings/* + /settings), react-router v6 picks most-specific match but the build-time types don't catch this — you need browser smoke. Plan Z defers that to Step 12, so document the routing invariant in the handoff doc explicitly so the smoke step knows what to verify.

Pitfall — TS-only verification is not enough: tsc --noEmit passes for valid route trees that silently shadow each other at runtime. Treat the routing invariant claim in the handoff doc as a prediction to be tested, not a guarantee.

Why Z beats the 紅線 21 fallback (delegate_task to subagent): Z keeps the work in the main session lineage (so cross-session search, state file, retro doc all stay coherent). Subagent isolation (紅線 21) is better when you need to escape the main session context budget; Z is better when the work is progressing fine and you just need a clean checkpoint.

When NOT to use Z: trivial 1-shot tasks (just do it inline), tasks with hard external dependencies that won't survive a /new (e.g. holding an SSH tunnel), or tasks where the user explicitly says "do it all in one go".

🔴 反例 — 2026-06-07 crm-system code review + P0 fix sprint 漏 call save_state.py

情境:David 嗌我做「CRM code review (A+B+C) → 開 P0 security fix sprint」,我整咗 12-todo list + 7 commits + evidence report,全程冇 call save_state.py。最終冇 data loss(全部 commit 落 git),但中間有 2 個高風險 moment:

改 auth.ts register endpoint 嘅時候第一次寫錯咗 _actorId 嗰個 hack(routes 唔應該 pass 內部 data via body)。如果有 state file 紀錄「auth.ts P0-1: requirePermission('user:create') 嗰度 getUserIdFromRequest 嘅 helper 應該點用」,第二次 patch 就唔會走歪路。
Bun --env-file 喺 smoke test 撞 shadowing — 2-3 次 attempts 浪費。State file 嘅 Insights section 寫低「Bun 嘅 env-file 做 floor, 唔可以 shell override」就唔需要再 re-discover。

Lesson:任何 > 30 min 嘅 multi-phase dev task 必須喺 Layer 1 trigger 嗰度 fire。即係:

用戶話「做 A + B + C」→ 即刻 save_state.py --goal "A + B + C" --trigger task-start
中段每 30 min / 每 10 tool calls → save_state.py --trigger auto-mid-task + sync_external.py
Sprint end → 寫 docs/retros/YYYY-MM-DD-<task>.md 後,rm docs/_meta/dev-task-state.md(task 完了)

Skip 嘅後果:compression / /new 之後要 re-discover 同一個 Bun env shadowing pitfall、或者重寫同一個 _actorId hack,等於 burn 5-10 min 嘅 tool calls。

新增紅線 37:multi-phase dev task (>30 min, 3+ todos) 必須 task-start 即 call save_state.py。同紅線 24 對齊(任何 long task 開始必須 save),但 spec 出嚟:唔淨止「long」,要「multi-phase」 — 即 todo list 有明確 phase boundary(Think/Plan/Build/Review/Test/Ship)嘅 task。

🔴 反例 2 — 2026-06-07 crm-system Day 14.7 Plan → Build 過渡 wire-shape drift (Step 5 → Step 7)

情境:Day 14.7 System Settings refactor + Tax Rate feature 嘅 multi-phase plan(Steps 5-12,有 Plan doc + 3 個 options 嘅 ADR)。Step 5(commit eb1581f)寫 frontend settingsApi.getTax/putTax wrapper 嘅時候,直接跟 Plan JSON doc 嘅 wording 用 defaultTaxRate 做 wire field name。但 plan doc 同 backend actual route handler(apps/api/src/routes/settings.ts)唔同步 — backend 用 rate 做 field name(GET response + PUT body 兩邊都係)。PUT runtime 一定 400 Zod validation error,用戶零實機 catch 唔到。

根因:

Plan doc 嘅 description field 寫「Default tax rate (%)」,field name 我哋 plan 階段咁就命名 defaultTaxRate(亦合理 readable)
Backend dev 直接 implement prisma.systemConfig.upsert({ data: { value: rate } }),response 用 rate 短名(亦合理)
Plan 階段冇人 grep backend source code verify actual wire shape
Frontend wrapper commit 嗰陣冇 e2e 過(我喺 commit message 寫「smoke test deferred to Step 12」,紅線 — 紅線 12/16 講 P0 US 必須 Unit+Integration+E2E 三層,但呢個算「wire format bug」唔係 US)

第 7 步先 catch 因為 Step 7 寫實際 settings-tax.tsx 嘅 wire call 嗰陣,grep backend route handler 確認 actual wire shape,發現 defaultTaxRate 根本唔存在於 backend response。Commit bd1d107 fix 咗 TaxConfig interface + putTax payload。

Lesson + 新增紅線 38:Plan → Build 過渡期 wire-shape verification(適用於所有 plan-then-build multi-step task,尤其涉及 backend ↔ frontend contract):

Step N wrapper commit 之前,grep backend actual handler 確認 wire field name + shape(record 喺 Insights section 嘅 state file)
Plan doc 嘅 description / 變量名係人話命名,backend 嘅 field name 係 code 命名,唔可以假設 1:1
Plan doc 有「Options A/B/C 決定」但冇記 wire shape 嘅 section → 自己加,喺 Plan stage commit 嗰陣 grep 一次 backend
Build stage 寫 wrapper 嗰陣,跟 backend source 唔跟 plan doc(紅線:「按後端 source code 為最終依歸」)

Prevention recipe(references/wire-shape-verification-recipe.md 有詳細 case study):

Plan stage 寫完 doc → git grep backend route / zod schema → 對齊 wire field name
Plan doc 開一個 "Wire shape (verified against backend)" section 記低 GET /x → { fieldA, fieldB } / PUT /x body → { fieldA }
Build stage wrapper commit 之前再 grep 一次(backend 可能 plan stage 之後改咗)
第一次實際 wire call 嘅 commit(典型係「實作第一個 page 嗰個 step」),commit 之前 5-min smoke(就算冇 full E2E 都要 curl 一次 GET 200 確認 field name)

反例損耗:Step 5 wrapper 寫錯 → Step 7 catch → commit bd1d107 fix 浪費 1 commit boundary 但冇 runtime disaster(因為 spec 寫「Step 12 smoke test 先驗」,未到 user)。如果 David Day 14 開個新 tab 直接 trigger putTax 就會 400 — 但 Plan 範圍冇 expose settings UI until Step 7,所以零 user-facing impact。Prevention 嘅 ROI 唔係「災難避免」,係「淨 commit boundary」+「plan doc 嘅 wire-shape section 對日後 audit 有用」。

📐 Reusable sub-route + tab navigation recipe(Day 14.7 Step 6-8 印證)

當 plan 揀咗「option A — 一個 URL 切換」嘅 Tabs pattern(對應常見嘅 admin → settings → sub-tab 重構),templates/shadcn-tabs-wrapper.tsx 係 known-good starter。常見陷阱:

Radix <Tabs> controlled mode 配 NavLink → warning。每個 <TabsTrigger> 必須有對應 value 同 active trigger 配對;直接 NavLink 入 <TabsList> 會 trigger Radix warning "missing trigger"。Pick B:TabsTrigger + onClick={() => navigate('/settings/' + value)} 失去 middle-click open-in-new-tab,但避 warning。
Layout 包 page 之後 double-header。Page 本身 render 自己嘅 h1 + 自製 button-tab strip → Layout 又 render → 視覺垃圾。Step 6 整 layout 嗰陣順手清 page 內部 chrome(寫埋 Step 8 extract 嘅 plan)。
Legacy route URL 保留。<Route path="/old"><Navigate to="/new" replace /></Route> 做 backward-compat 5-10 個 redirect 一次過加,bookmark/email link 唔會死。
Query-string 預填 filter 嗰啲 page。例如 audit page 接受 ?action=SYSTEM_CONFIG_UPDATED 從 settings 跳過來即 filtered → 入 page 嗰陣 useSearchParams() 攞 initial state,1-2 行 code。

完整 7-step walkthrough + 對應 crm-system commit hash 對照 喺 references/wire-shape-verification-recipe.md 嘅 "Sub-route tab navigation" section。

📝 E2E test result (2026-06-06)

Save: 2715 bytes 寫入 ~/www/crm-system/docs/_meta/dev-task-state.md
Edit: 2 decisions + 1 insight injected
Sync: 4 facts → ~/.hermes/memories/dev-task-facts.jsonl
Read: 4 facts loaded back ✓
Search: "Prisma" → 1 hit ✓
Load: state file full render ✓

Subagent WIP Pickup test (2026-06-06)

入口: User send "現在繼續做" + git status show 1 untracked + 2 modified
Recipe: 6 steps, ~15 min total
結果: 1 commit (7dc56d0), push 成功, git log --oneline origin/main..HEAD empty
節省 vs re-delegate: ~15 min (WIP 95% correct, 淨係 consumer page wiring 缺)
新增 lessons: audit .gitignore for subagent duplicates, batch-fix noise 喺獨立 commit

name	dev-task-memory
description	5-layer persistent memory for in-progress dev tasks — survives context compression, /new, gateway restart. State file + git checkpoints + external memory + cross-session search.
trigger	context compression / /new / gateway restart / '繼續之前的任務' / 'where were we' / 'save state' / 'checkpoint task' / subagent 做到一半 stop 嘅 WIP / multi-phase dev task (>30 min, e.g. security review → fix-batch sprint)
version	2
category	devops

Dev Task Memory

5 個 Layer

┌─────────────────────────────────────────────────────────────┐
│  Layer 1: TRIGGER (決定幾時 save)                            │
├─────────────────────────────────────────────────────────────┤
│  • 紅線 21-22 (long task, > 50 turns)                       │
│  • auto-compression event (gateway/run.py hook)             │
│  • gateway restart                                          │
│  • /new slash command                                       │
│  • User explicit "save state" / "checkpoint"               │
│  • **Subagent 做到一半 stop 嘅 WIP** (新 trigger — 2026-06-06)│
│    → working tree 有 untracked / modified files, 跟住用    │
│      `references/subagent-wip-pickup-recipe.md` 嘅 6-step   │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 2: STATE FILE (Single Source of Truth)               │
├─────────────────────────────────────────────────────────────┤
│  ~/www/<project>/docs/_meta/dev-task-state.md:             │
│  - Goal, Decisions (with WHY), Current state, Next steps,   │
│    Risks, Insights, Session lineage, References             │
│  - Gitignored via .gitignore snippet                        │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 3: GIT-TRACKED (可審計 + 可 rollback)                  │
├─────────────────────────────────────────────────────────────┤
│  • enable checkpoints: true (Hermes built-in)               │
│  • /rollback N for file rollback                            │
│  • git commit footer 包含 state summary                     │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 4: EXTERNAL MEMORY (cross-session fact recall)        │
├─────────────────────────────────────────────────────────────┤
│  • holographic (local, no API key) or mem0/honcho (API)     │
│  • Sync facts (decisions, insights, risks) 落 external store │
│  • Cross-session: 「之前你話用 X 點解?」即時 recall         │
│  • 當冇 external provider, fallback 落                      │
│    ~/.hermes/memories/dev-task-facts.jsonl                  │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Layer 5: CROSS-SESSION SEARCH (session_search FTS5)        │
├─────────────────────────────────────────────────────────────┤
│  • 開新 session 自動 search "what was I working on"        │
│  • 注入 top-3 relevant past sessions                        │
│  • 用 hermes sessions list 找 candidate sessions           │
│  • 唔需要 manual session ID 記憶                            │
└─────────────────────────────────────────────────────────────┘

🔧 工具腳本

Script	用途	觸發時機
`scripts/save_state.py`	Save 當前 dev task state 落 `docs/_meta/dev-task-state.md`	Layer 1 trigger fires 佢
`scripts/load_state.py`	Read state file + 注入 context (for resume)	New session / /new / after-compression / after-restart
`scripts/sync_external.py`	Extract facts → push to external memory (or local fallback)	每次 save_state 完之後 fire

📂 支援 References

File	觸發時機
`references/subagent-wip-pickup-recipe.md`	Resume 一個 subagent 做到一半停咗嘅 task — working tree 有 uncommitted WIP, 跟住 6-step recipe: snapshot → commit history → read untracked → audit side effects → finish → verify-before-declaring-done。涵蓋 duplicate gitignore entries、batch-fix 嘅 anti-pattern、revert detection 等。
`references/recovery-template-limitation.md`	Day 11 lesson (2026-06-09): `recovery.sh` + `save_state.py` 嘅 generic template 永遠唔 fill 真正 state(Decisions / Files / Next Steps / Risks section 全部係 placeholder)。Resume 之前一定要手動 overwrite `dev-task-state.md`,fill 真正 state。Workaround recipe + root cause analysis + future fix blueprint 全部喺入面。
`references/working-tree-wip-detection.md`	Day 11 lesson (2026-06-09): Agent 同 David 共用 worktree 嘅時候,David 嘅 WIP(modified + untracked) 唔可以入 agent 嘅 commit。Selective `git add <files>` recipe + 3 條新紅線(34-36)+ Day 11 Phase 1 嘅 case study(crm-system 5M + 3?? WIP detection)。
`references/plan-execution-deviation-protocol.md`	Day 14.7 lesson (2026-06-07): Plan JSON 內部矛盾嘅處理 — 揀 primary intent 唔揀 literal text、3 處冗餘紀錄(code comment + commit + report)、anti-pattern(用戶不問照 ship、靜靜 rewrite plan)。Step 6 `/settings` 案例做 worked example。
`templates/handoff-doc-scaffold.md`	Day 14 lesson (2026-06-07): Human-readable handoff doc 嘅 scaffold,用於 `/new` / 收工 / 「之後呢」trigger 嗰陣,Agent 寫一份 David + fresh session agent 兩邊都讀得順嘅 commit-able doc(同 state file 互補)。Section structure 見 §5 嘅 recipe。

📖 標準使用流程

1) 開新 dev task (Plan → Build 開始)

# 進入 project directory
cd ~/www/crm-system

# 立即 save 一個 fresh state
python3 ~/.hermes/profiles/developer/skills/dev-task-memory/scripts/save_state.py \
    --project crm-system \
    --goal "實作 Companies 編輯頁加聯繫人 sub-row" \
    --trigger "task-start"

Agent 收到 trigger 應該:

立即 call save_state.py
Fill in docs/_meta/dev-task-state.md 嘅 Decisions、Next Steps、Insights section
Call sync_external.py push facts 落 external memory

2) Long task 中段 (每 30 min 或每 10 個 tool calls)

Agent 自動 trigger:

python3 scripts/save_state.py --project crm-system --trigger "auto-mid-task"
python3 scripts/sync_external.py --project crm-system

呢個確保就算 hang / compression / restart, state 都唔會 lost。

3) Context 壓縮 / Hang fix / /new / Restart 後恢復

User send "繼續" / "resume" / "where were we":

python3 scripts/load_state.py --project crm-system --search-sessions

Agent 收到 output 後:

Read state file content
Inject 入 conversation context
Resume 跟住 "Next 3-5 Steps" 繼續做

3b) Subagent WIP Pickup (新 — 2026-06-06)

git status + git diff --stat snapshot working tree
git log --oneline -10 check partial progress
Read untracked / modified files in full
Audit side effects (.gitignore duplicates, package.json, etc.)
Finish from actual pickup point
git log --oneline origin/main..HEAD empty + clean working tree 為止

4) 任務完成

刪除 state file (task done, no need to resume):

rm ~/www/<project>/docs/_meta/dev-task-state.md

或者 archive 落 docs/retros/YYYY-MM-DD-<task-name>.md。

5) Human-Readable Handoff Doc (Day 14 lesson — 2026-06-07)

唔可以只靠 state file:

dev-task-state.md 係 gitignored (紅線 28) — 下一個 session 嘅 fresh agent 唔知有呢個 file
內容係 machine-optimized,David 讀起上嚟唔順
冇 commit 嘅話,different branch / different machine 嘅 agent 攞唔到

Handoff doc 嘅特性(同 state file 對比):

特性	`docs/_meta/dev-task-state.md` (state)	`docs/retros/YYYY-MM-DD-<task>-handoff.md` (handoff)
Gitignored	✅ 紅線 28	❌ 正常 commit(可以 PR 一齊)
Reader	Machine + Agent	David + 新 session agent + reviewer
Format	Structured (Goal/Decisions/Next Steps)	Narrative + TL;DR + resume block
Content	Decisions with WHY, current step	Decisions, uncommitted WIP (App.tsx diff!), key env vars, first-3-commands
Lifetime	任務期間	任務結束後 audit trail(永遠保留)
Use case	`load_state.py` resume	`/new` 後手動 read

Recipe(2026-06-07 crm-system handoff 試過 work):

# 1. 收集 state(15 lines)
cd ~/www/<project>
git log --oneline -8 <branch>
git status --short
git diff --stat <base>..<branch> | tail -5

# 2. 寫 handoff doc (見 templates/handoff-doc-scaffold.md)
# 結構: TL;DR + Branch State + What's Done + What's NOT Done + Key Decisions
#       + Uncommitted Work + Critical Context + First 3 Commands

# 3. Save state(獨立, 唔係 handoff 嘅 substitute)
python3 scripts/save_state.py --project <name> --goal "..." --trigger task-start
python3 scripts/sync_external.py --project <name>

# 4. Verify 兩個 output 存在
ls -la docs/_meta/dev-task-state.md docs/retros/<date>-<task>-handoff.md

# 5. 印 resume command 畀 David(≤ 5 行)

Critical sections that MUST be in handoff doc(David 試過 miss 嘅):

Uncommitted WIP — git diff 結果(紅線 30 同 recovery 同理)
David 嘅 untracked files — 永遠唔可以「清理」,必須 list 出嚟
Key env vars / 32-char gate — 例如 JWT_SECRET 32-char hard-fail,新 session 要 re-export
Latent bugs fix 嘅 caveat — 例如 seed Role/RolePermission 喺 prod 唔存在,新 endpoint 會 403
First 3 commands — cd + load_state.py + start dev
Routing / runtime invariants (新 — 2026-06-07 lesson) — 當 step commit 包含 order-dependent routing(sibling route conflict 嗰類)但 smoke test 留到 ship step,必須喺 handoff doc 開獨立 sub-section,寫低「聲稱嘅行為 + 點 verify」。TS pass ≠ runtime pass。詳見 react-router-sibling-route-conflict skill。

🔄 同 Hermes 內建功能嘅整合

Hermes feature	dev-task-memory 用法
`hermes --resume <id>` / `-c "<name>"`	state file 嘅 `Session Lineage` section 寫低 resume command
`hermes checkpoints` (filesystem)	Layer 3 自動 enable, 唔需要額外 config
`hermes memory status`	Layer 4 查 active provider
`session_search` (FTS5)	Layer 5 `load_state.py --search-sessions` 用佢
`/rollback N`	Layer 3 file rollback 配 state file restore

⚠️ 紅線 (跟 SOUL.md 紅線 24-28 對齊)

紅線 24:每次 long task 開始必須 save_state.py, 唔可以 rely on LLM memory
紅線 25:每 30 分鐘或 10 個 tool calls 必須 re-save (Layer 1 自動 trigger)
紅線 26:Decision 必須有 WHY — 唔可以淨寫 "Use Hono", 要寫 "Use Hono 4x 細, edge 啱用"
紅線 27:Resume 時必須先 load_state.py, 唔可以假設自己記得
紅線 28:State file 唔可以 commit 落 git — docs/_meta/* 必須喺 .gitignore
紅線 33:Resume 之前必 check dev-task-state.md 唔係 generic template(Day 11 lesson)。 Generic template 嘅 signals:有 <Decision 1> / <Risk 1> / **待填寫** / 全部 file 都係 M (now) 冇真實 file path。Hit 任何一個 → 手動 overwrite 真正 state 先 resume。詳見 references/recovery-template-limitation.md 嘅 workaround recipe。
紅線 34:Sibling subagent 並行 edit 同一 file 時必 check 雙方最後 diff 唔 overlap(Day 14.7 lesson)。 Patch tool / write_file 撞到 warning "modified by sibling subagent" / "last read with offset/limit pagination" → 唔好盲改。Recipe: (1) git diff <file> 睇 sibling 改咗邊段; (2) 自己嘅 patch 區段同 sibling 嘅 patch 區段唔可以重疊; (3) 如重疊(regex region 撞共同 functions / interfaces),手動 merge 兩邊嘅改動,唔好 git add + commit 直接撞; (4) Commit message 寫低「conflict-free with sibling X 嘅 Y 改動」。詳見 references/sibling-subagent-concurrent-edit-recipe.md(待建)。
紅線 35:Tying two patches to the same commit when they're in different sub-trees is safe; the dangerous case is two patches both touching the same function body or interface declaration. Quick test before committing sibling-collision files: git diff <file> | grep -E "^[+-]" | sort | uniq -c | sort -rn | head -20 — any line that appears as both + and - is a true merge conflict, not just proximity.

🛠️ 設定 (.gitignore snippet)

每個 project 嘅 .gitignore 加:

# dev-task-memory — runtime state, not source
docs/_meta/
dev-task-state.md

🧪 驗證指令

# 1. Save state
python3 scripts/save_state.py --project crm-system --goal "Test" --trigger "manual"

# 2. Sync to external
python3 scripts/sync_external.py --project crm-system

# 3. Read back
python3 scripts/sync_external.py --read --project crm-system

# 4. Search
python3 scripts/sync_external.py --search "Prisma"

# 5. Load (resume)
python3 scripts/load_state.py --project crm-system --search-sessions

📚 配合其他 skills

context-summarizer (existing): 自動每 30 分鐘壓 context。dev-task-memory 補佢嘅缺點 — 壓咗之後 decisions 唔會 lost
regression-guard (existing): 防舊 bug 翻發。state file 嘅 Risks section 配 RG-XXX
auto-doc-gen (existing): 自動生成 API doc。state file 嘅 Decisions 配 doc rationale

🐛 已知限制

save_state.py 嘅 "extract decisions from session" 而家係 stub — TODO: integrate with hermes_state.py 讀真正 session DB
External memory provider (mem0/honcho) API call 仲未 implement — 而家 fallback 落 local jsonl
Holographic provider 雖然係 local-only 但需要 config, 默認未啟用 — 可以之後 setup
State file 唔會 auto-update — 必須人手 call save_state.py (將來可以做 background curator 自動 trigger)
memory tool round-trip drift guard (2026-06-19 lesson) — Hermes 嘅 memory(action='add') 會 refuse write 當 ~/.hermes/profiles/developer/memories/MEMORY.md 有 manual edit / shell append / patch tool 嘅 content 唔可以 round-trip through memory() call。Detection: error message 提 "MEMORY.md.bak." + "Resolve the drift first — either rewrite the file as a clean §-delimited list of entries, or move the extra content out"。Resolution recipe:
1. cat ~/.hermes/profiles/developer/memories/MEMORY.md.bak.<timestamp> 拎 backup
2. Manual integrate missing entries into clean §-delimited list (one at a time via memory(action='add', content='...'))
3. Remove or rewrite original MEMORY.md to clean state
4. Retry memory(action='add')
唔可以 force overwrite — guard 嘅存在係防止 silent data loss (Hermes issue #26045)。唔可以 retry 5 次(浪費 token + 撞同一個 error)。Lesson codification: 如果個 lesson 必須 persist 但 memory tool 拒絕,fallback 落 skill(SKILL.md body)或者講畀 David 知要人手 sync。

🆕 Plan Z — Mid-Handoff Single-Step Commit Cycle (2026-06-07)

Trigger: User has multi-step work (≥5 steps) remaining, session has hit 紅線 22 (>50 turns) or 紅線 19 (tool-call noise from earlier hang). 3 paths to offer, ordered by safety:

Path	Description	Hang risk	/new cycles	Total wall time
X (Inline all)	Do every step in this session, accept hang	High (≥1 hang mid-run)	0 (until forced)	4-5 hr + 1 unplanned handoff
Y (Inline half)	Do first N/2 steps, then /new handoff	Medium	1 planned	4-5 hr, 1 clean checkpoint
Z (1-step cycles)	Do exactly 1 step + commit + handoff + /new, repeat	~Zero per cycle	1 per step	4-5 hr, multiple clean checkpoints

Plan Z execution recipe (tested on crm-system 2026-06-07 Step 5):

Scope the next single step only — read handoff doc, pick the smallest step with concrete file paths (e.g. "Step 5: routes + API client", not "Step 5-8: all frontend").
Execute in ≤20 tool calls — hard budget. If you cross it, stop and handoff what you have.
Verify locally — tsc --noEmit, git status --short, type check. Defer browser/E2E smoke to the final ship step (Step 12 of the original plan).
Commit + push — single commit, descriptive footer pointing to next step.
Update handoff doc + state file — re-stamp the "What's Done" section with the new commit, mark the step as done in the "What's NOT Done" table.
Emit "📍 progress" + explicit /new recommendation — let user run /new themselves. Don't auto-/new (user has the shortcut key).
Tool-call budget enforcement: if a step needs >20 calls, split it across 2 cycles BEFORE starting, not mid-cycle.

🔴 反例 — 2026-06-07 crm-system code review + P0 fix sprint 漏 call save_state.py

改 auth.ts register endpoint 嘅時候第一次寫錯咗 _actorId 嗰個 hack(routes 唔應該 pass 內部 data via body)。如果有 state file 紀錄「auth.ts P0-1: requirePermission('user:create') 嗰度 getUserIdFromRequest 嘅 helper 應該點用」,第二次 patch 就唔會走歪路。
Bun --env-file 喺 smoke test 撞 shadowing — 2-3 次 attempts 浪費。State file 嘅 Insights section 寫低「Bun 嘅 env-file 做 floor, 唔可以 shell override」就唔需要再 re-discover。

Lesson:任何 > 30 min 嘅 multi-phase dev task 必須喺 Layer 1 trigger 嗰度 fire。即係:

用戶話「做 A + B + C」→ 即刻 save_state.py --goal "A + B + C" --trigger task-start
中段每 30 min / 每 10 tool calls → save_state.py --trigger auto-mid-task + sync_external.py
Sprint end → 寫 docs/retros/YYYY-MM-DD-<task>.md 後,rm docs/_meta/dev-task-state.md(task 完了)

Skip 嘅後果:compression / /new 之後要 re-discover 同一個 Bun env shadowing pitfall、或者重寫同一個 _actorId hack,等於 burn 5-10 min 嘅 tool calls。

🔴 反例 2 — 2026-06-07 crm-system Day 14.7 Plan → Build 過渡 wire-shape drift (Step 5 → Step 7)

根因:

Plan doc 嘅 description field 寫「Default tax rate (%)」,field name 我哋 plan 階段咁就命名 defaultTaxRate(亦合理 readable)
Backend dev 直接 implement prisma.systemConfig.upsert({ data: { value: rate } }),response 用 rate 短名(亦合理)
Plan 階段冇人 grep backend source code verify actual wire shape
Frontend wrapper commit 嗰陣冇 e2e 過(我喺 commit message 寫「smoke test deferred to Step 12」,紅線 — 紅線 12/16 講 P0 US 必須 Unit+Integration+E2E 三層,但呢個算「wire format bug」唔係 US)

Lesson + 新增紅線 38:Plan → Build 過渡期 wire-shape verification(適用於所有 plan-then-build multi-step task,尤其涉及 backend ↔ frontend contract):

Step N wrapper commit 之前,grep backend actual handler 確認 wire field name + shape(record 喺 Insights section 嘅 state file)
Plan doc 嘅 description / 變量名係人話命名,backend 嘅 field name 係 code 命名,唔可以假設 1:1
Plan doc 有「Options A/B/C 決定」但冇記 wire shape 嘅 section → 自己加,喺 Plan stage commit 嗰陣 grep 一次 backend
Build stage 寫 wrapper 嗰陣,跟 backend source 唔跟 plan doc(紅線:「按後端 source code 為最終依歸」)

Prevention recipe(references/wire-shape-verification-recipe.md 有詳細 case study):

Plan stage 寫完 doc → git grep backend route / zod schema → 對齊 wire field name
Plan doc 開一個 "Wire shape (verified against backend)" section 記低 GET /x → { fieldA, fieldB } / PUT /x body → { fieldA }
Build stage wrapper commit 之前再 grep 一次(backend 可能 plan stage 之後改咗)
第一次實際 wire call 嘅 commit(典型係「實作第一個 page 嗰個 step」),commit 之前 5-min smoke(就算冇 full E2E 都要 curl 一次 GET 200 確認 field name)

📐 Reusable sub-route + tab navigation recipe(Day 14.7 Step 6-8 印證)

當 plan 揀咗「option A — 一個 URL 切換」嘅 Tabs pattern(對應常見嘅 admin → settings → sub-tab 重構),templates/shadcn-tabs-wrapper.tsx 係 known-good starter。常見陷阱:

Radix <Tabs> controlled mode 配 NavLink → warning。每個 <TabsTrigger> 必須有對應 value 同 active trigger 配對;直接 NavLink 入 <TabsList> 會 trigger Radix warning "missing trigger"。Pick B:TabsTrigger + onClick={() => navigate('/settings/' + value)} 失去 middle-click open-in-new-tab,但避 warning。
Layout 包 page 之後 double-header。Page 本身 render 自己嘅 h1 + 自製 button-tab strip → Layout 又 render → 視覺垃圾。Step 6 整 layout 嗰陣順手清 page 內部 chrome(寫埋 Step 8 extract 嘅 plan)。
Legacy route URL 保留。<Route path="/old"><Navigate to="/new" replace /></Route> 做 backward-compat 5-10 個 redirect 一次過加,bookmark/email link 唔會死。
Query-string 預填 filter 嗰啲 page。例如 audit page 接受 ?action=SYSTEM_CONFIG_UPDATED 從 settings 跳過來即 filtered → 入 page 嗰陣 useSearchParams() 攞 initial state,1-2 行 code。

完整 7-step walkthrough + 對應 crm-system commit hash 對照 喺 references/wire-shape-verification-recipe.md 嘅 "Sub-route tab navigation" section。

📝 E2E test result (2026-06-06)

Save: 2715 bytes 寫入 ~/www/crm-system/docs/_meta/dev-task-state.md
Edit: 2 decisions + 1 insight injected
Sync: 4 facts → ~/.hermes/memories/dev-task-facts.jsonl
Read: 4 facts loaded back ✓
Search: "Prisma" → 1 hit ✓
Load: state file full render ✓

Subagent WIP Pickup test (2026-06-06)

入口: User send "現在繼續做" + git status show 1 untracked + 2 modified
Recipe: 6 steps, ~15 min total
結果: 1 commit (7dc56d0), push 成功, git log --oneline origin/main..HEAD empty
節省 vs re-delegate: ~15 min (WIP 95% correct, 淨係 consumer page wiring 缺)
新增 lessons: audit .gitignore for subagent duplicates, batch-fix noise 喺獨立 commit

dev-task-memory

More from this repository

More from this repository

Dev Task Memory

5 個 Layer

🔧 工具腳本

📂 支援 References

📖 標準使用流程

1) 開新 dev task (Plan → Build 開始)

2) Long task 中段 (每 30 min 或每 10 個 tool calls)

3) Context 壓縮 / Hang fix / /new / Restart 後恢復

3b) Subagent WIP Pickup (新 — 2026-06-06)

4) 任務完成

5) Human-Readable Handoff Doc (Day 14 lesson — 2026-06-07)

🔄 同 Hermes 內建功能嘅整合

⚠️ 紅線 (跟 SOUL.md 紅線 24-28 對齊)

🛠️ 設定 (.gitignore snippet)

🧪 驗證指令

📚 配合其他 skills

🐛 已知限制

🆕 Plan Z — Mid-Handoff Single-Step Commit Cycle (2026-06-07)

🔴 反例 — 2026-06-07 crm-system code review + P0 fix sprint 漏 call save_state.py

🔴 反例 2 — 2026-06-07 crm-system Day 14.7 Plan → Build 過渡 wire-shape drift (Step 5 → Step 7)

📐 Reusable sub-route + tab navigation recipe(Day 14.7 Step 6-8 印證)

📝 E2E test result (2026-06-06)

Subagent WIP Pickup test (2026-06-06)

Dev Task Memory

5 個 Layer

🔧 工具腳本

📂 支援 References

📖 標準使用流程

1) 開新 dev task (Plan → Build 開始)

2) Long task 中段 (每 30 min 或每 10 個 tool calls)

3) Context 壓縮 / Hang fix / /new / Restart 後恢復

3b) Subagent WIP Pickup (新 — 2026-06-06)

4) 任務完成

5) Human-Readable Handoff Doc (Day 14 lesson — 2026-06-07)

🔄 同 Hermes 內建功能嘅整合

⚠️ 紅線 (跟 SOUL.md 紅線 24-28 對齊)

🛠️ 設定 (.gitignore snippet)

🧪 驗證指令

📚 配合其他 skills

🐛 已知限制

🆕 Plan Z — Mid-Handoff Single-Step Commit Cycle (2026-06-07)

🔴 反例 — 2026-06-07 crm-system code review + P0 fix sprint 漏 call save_state.py

🔴 反例 2 — 2026-06-07 crm-system Day 14.7 Plan → Build 過渡 wire-shape drift (Step 5 → Step 7)

📐 Reusable sub-route + tab navigation recipe(Day 14.7 Step 6-8 印證)

📝 E2E test result (2026-06-06)

Subagent WIP Pickup test (2026-06-06)