[che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名）

[kiki830621]

Problem

使用者原文（2026-04-14 對話中發現）：
「archive mail沒有順便把附件抓下來嗎？」

現有 commands/archive-mail.md 的 Step 5 生成 Markdown 時，只把附件檔名寫進 Markdown 正文（例如 Attachments: Taxometric_Analysis_Submission_v3_20260408.docx, Figures & Tables20260408.docx），沒有實際下載附件到本地磁碟。

2026-04-08 歸檔的 Tatsuma 信件即為例證：communications/2026-04-08_Re--Shall-we-resume-work-on-the-Taxometric-Analysis-paper-al-1.md 原本最後一行只有檔名字串，原始 .docx 仍停留在 Apple Mail.app 內部儲存區；使用者若日後要存取，必須手動從 Mail 取出。

MCP server 本身有附件能力但 command 完全沒用：

• mcp__plugin_che-apple-mail-mcp_mail__list_attachments
• mcp__plugin_che-apple-mail-mcp_mail__list_attachments_batch
• mcp__plugin_che-apple-mail-mcp_mail__save_attachment

Type

feature

Baseline from manual backfill (2026-04-14)

2026-04-14 session 已在 tatsuma 專案手動實作整套 pattern 一次，驗證 7 個附件、3 個 email、跨 thread 2 週時間範圍。commits: bcac7ee (init), 784eac6 (backfill)。以下是已驗證可行的約定，正式實作時應直接採用：

1. 目錄結構（確定）

{output_dir}/attachments/{主郵件 markdown stem}/{原檔名}

其中 markdown stem 是主信歸檔 md 的檔名去掉 .md。例如：

主信 md	附件目錄
2026-04-08_Re--Shall-we-resume-work-on-the-Taxometric-Analysis-paper-al-1.md	attachments/2026-04-08_Re--Shall-we-resume-work-on-the-Taxometric-Analysis-paper-al-1/
2026-03-17_Re--Shall-we-resume-work-on-the-Taxometric-Analysis-paper-Ab.md	attachments/2026-03-17_Re--Shall-we-resume-work-on-the-Taxometric-Analysis-paper-Ab/

好處：一對一對應、可 ls 查找、支援回填（md 已存在、附件後補）。

2. 檔名保留原始（確定）

附件檔名不套 #16 的 subject hyphenation rule，保留原始 bytes（含空白、&、中日文、emoji）。磁碟 I/O 用原字串寫入，在 Markdown link URL 才做 URL encoding。

3. URL encoding（確定）

Markdown 的 [label](url) 裡只需要處理：

• 空白 → %20
• & → %26
• 其他一律保留（中日文 Unicode 不動，GitHub/VS Code 都能 render）

顯示文字（[label]）保留原檔名完整可讀。範例：

- [Figures & Tables20260408.docx](attachments/2026-04-08_.../Figures%20%26%20Tables20260408.docx) (93 KB)

4. Attachments 區塊在 md 中的位置（確定）

原信簽名結束後（例如 Tatsuma 的 ∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞ 分隔線之後），引用 thread history（差出人: / 寄件者: 段落）之前。這樣邏輯上屬於「這封信」而非被混到 thread quote 裡。

... Tatsuma's body ...
Kind regards,
Tatsuma
∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞
... signature ...
∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞∞

Attachments:
- [file1](...) (146 KB)
- [file2](...) (93 KB)

差出人: Lay, Keng-Ling <kllay@ntu.edu.tw>
... thread history ...

5. 回信無附件時的處理（確定）

回信引用原信時，原信的附件標記（例如 <Taxometric_Analysis_Submission_v3.docx>）會被 Mail client 留在 text quote 裡，但 list_attachments(reply_email) 會回傳空陣列。兩者不一致時以 list_attachments 為準。回信 md 的 Attachments: 區塊建議寫 cross-reference：

(Attachments on the original email from Tatsuma — see `2026-04-08_..._al-1.md`)

6. Backfill（既有歸檔補附件）（確定）

若歷史 md 已歸檔但附件沒抓，可以後補：

1. 從 list_attachments 取得檔案清單
2. save_attachment 到 attachments/{既有 md stem}/
3. Edit md 在簽名後插入 Attachments: 區塊

本次 backfill 範例：2026-03-17 的 v2 原稿（4 檔 882 KB）+ 2026-03-23 的 Keng-Ling track-changes 版（1 檔 143 KB）。

7. Thread 覆蓋度檢查（新發現）

執行 /archive-mail 後，若要保證附件完整，應對已歸檔的同 thread email 全部跑 list_attachments_batch 檢查是否有漏抓的附件。今天發現 3/23 整封 email 漏歸檔（date_from=2026-03-24 太晚），需要擴大 thread scope。

Expected（spec）

在 Step 5 生成 Markdown 前新增 Step 4.5「下載附件」：

1. list_attachments_batch([所有待歸檔 email]) → 一次取得所有附件元資料
2. 對每個有附件的 email：
   • mkdir -p {output_dir}/attachments/{主郵件 markdown stem}/
   • 對每個附件 save_attachment(...) 到該目錄（保留原檔名）
3. Step 5 生成 Markdown 時在簽名後插入 Attachments 區塊（見 baseline §4）

Actual

Step 5 只描述 Markdown 內容格式，完全沒提附件下載。

Impact

• 歸檔的價值打折：幾年後想看當時討論 "draft v3" 具體內容，需要回 Apple Mail 找，Mail 若已清理或帳號異動就失去附件
• 對合作者來說歸檔不完整
• 與 /archive-mail 的「完整歸檔到本地」語意不符

Open Questions

Baseline 已回答 Q1 (目錄)、Q4 (檔名)、新增了 position / encoding / backfill 規則。剩下 3 個需在 diagnose 階段決定：

• Q1 目錄結構 → Decided: attachments/{main-email-markdown-stem}/
• Q2 去重：同一附件被多封信 forward 時如何處理（hash 檢查 / 忽略 / 每封信各存一份）？
• Q3 大小限制：超過 N MB 的附件是否跳過？是否接 git-lfs？（baseline 實作的最大檔案為 439 KB，未踩到限制）
• Q4 附件檔名 → Decided: 保留原檔名，僅 URL encoding
• Q5 跳過規則：簽名檔裡的 logo 圖片、email client 的 inline image 是否要歸檔？
• Q6 .claude/emails.md 新增欄位：archive_attachments: bool（預設 true）、attachment_size_limit_mb: int（預設 25）、skip_inline_images: bool（預設 true）？
• Q7 （新）Thread 覆蓋度：/archive-mail 是否該自動對同 thread 已歸檔 email 做 backfill 檢查？或另外提供 /archive-mail --backfill-attachments？

Dependencies

• 建議在 [che-apple-mail-mcp] archive-mail 不支援零參數模式（無法讀取 .claude/emails.md） #13 .claude/emails.md schema 定稿後再實作（需要新增 attachment 相關 config 欄位）
• 與 [che-apple-mail-mcp] archive-mail Markdown 內容格式過度複雜（元資料表+摘要+待辦，實務用簡單版） #17 Markdown 內容格式簡化一起考慮（Attachments: 欄位設計）

Reference implementation

Tatsuma 專案 2026-04-14 backfill：

• PsychQuant/collaborations_tatsuma commits bcac7ee (init) + 784eac6 (backfill)
• 3 個目錄、7 個附件、2 個 md 更新 + 1 個新 md
• 檔名從 54 KB docx 到 439 KB pdf 皆驗證 OK

[kiki830621]

Body updated with baseline from manual backfill

本次更新把 2026-04-14 session 中手動實作的完整 pattern 寫進 body：

• 確定 3 個規則：目錄結構 attachments/{md stem}/、原檔名保留、URL encoding 只處理空白與 &
• 新增 4 條約定：Attachments 區塊在 md 中的位置、回信無附件的 cross-reference、backfill 流程、thread 覆蓋度檢查
• 回答 2 個 open questions（Q1 + Q4），剩下 5 個待 diagnose

實作可直接從 baseline 開始，降低 diagnose 成本。

Reference commits：

• bcac7ee feat(communications): initialize email archive
• 784eac6 feat(communications): backfill v2 attachments + 2026-03-23 Keng-Ling email

[kiki830621]

🪢 Cross-repo tracking

This issue is part of the Manuscript review automation umbrella tracked at PsychQuant/macdoc#75.

See that issue for the full dependency tree and other related child issues across ooxml-swift, che-word-mcp, psychquant-claude-plugins, and che-apple-mail-mcp.

[kiki830621]

Diagnosis (consolidated — #23 #24 #25 #26)

Type

mixed: bug (#24) + 3 features (#23 #25 #26), all in the same skill file commands/archive-mail.md and interdependent. Analyzed together as one bundle.

Context

Scanned the current skill at /Users/che/.claude/plugins/marketplaces/psychquant-claude-plugins/plugins/che-apple-mail-mcp/commands/archive-mail.md (202 lines, 7 steps). The skill today:

1. Takes a single filter string (sender/to substring)
2. Searches via mail__search_emails with query: filter
3. Generates Markdown per email with attachment filenames as plain text (no download, no MCP save_attachment call)
4. No post-hoc coverage check

MCP capabilities exist but skill never uses them: list_attachments, list_attachments_batch, save_attachment.

Dependency analysis

All 4 issues touch the same skill file. They are interdependent:

#24 filter asymmetry (bug)      ──┐
   └─ without this, thread       │
      completeness is broken     │
                                 ├──►  need to design together
#23 attachment download ─┐       │
   └─ Downloading makes  │       │
      coverage auditable ├──►    │
                         │       │
#25 coverage audit ──────┘       │
   └─ validates #23 did its job  │
                                 │
#26 attachment routing ──────────┘
   └─ decides WHERE #23 writes to
      (data/ vs correspondence/attachments/)

#23 is the foundational feature. #26 extends #23's output logic. #25 validates #23's output. #24 expands the input set that #23 operates on.

Root cause (for #24 bug)

Skill uses mail__search_emails with query: filter — matches sender or to. But collaboration threads include internal multi-party discussions where the external collaborator is not in sender/to (e.g., Che ↔ Lay ↔ Pei-Chi without Tatsuma). Those threads are invisible to the current filter.

Requirements breakdown (for #23/#25/#26)

#23 — attachment download:

• Use save_attachment MCP for each email with attachments
• Directory: {output_dir}/attachments/{email_md_stem}/{original_filename} (per baseline established in 2026-04-14 tatsuma backfill)
• Filename preservation: keep original bytes on disk; URL-encode %20 / %26 only in Markdown link
• Attachment block placement: after signature, before thread quote
• Reply without attachments: cross-reference original email's md (no duplicate download)

#25 — coverage audit:

• After Step 6, add Step 7: list_attachments_batch on archived emails, compare with disk
• 3 phases: (Phase 1) undownloaded-attachment warn, (Phase 2) thread-aware orphan detection, (Phase 3) --backfill-attachments / --backfill-thread flags for auto-repair

#26 — attachment routing:

• Extension-based rules: .csv/.sav/.dta/.parquet/.xlsx → data/raw/; .pdf/.docx/.txt/.md → attachments/
• Keyword override: filenames containing data|raw|indicators|codebook → data/; Submission|Figures|Manuscript|draft → attachments/
• Config: .claude/emails.md YAML attachment_routing block for per-project overrides
• Report: "15 attachments archived: 4 to data/raw/, 11 to attachments/"

Impact

• Affected file: plugins/che-apple-mail-mcp/commands/archive-mail.md (single file, 202 lines → likely ~350-400 after all 4)
• Affected config schema: .claude/emails.md gains 3 new optional fields:
  • subject_keywords (for [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24)
  • participant_aliases (for [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24)
  • attachment_routing (for [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26)
• Affected users: any caller of /archive-mail. Current behavior preserved when new fields are absent — additive.

Strategy

• [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24: expand Step 3 search with subject_keywords + optional participant_aliases matching; reference-chain follow via In-Reply-To (if MCP supports)
• [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23: add Step 6b (between Markdown generation and index update) that calls save_attachment for each email with attachments, writes to {output_dir}/attachments/{email_stem}/{filename}
• [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23: Markdown generation updates — insert Attachments block between signature and thread quote with URL-encoded links
• [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26: attachment routing logic (extension + keyword + YAML config) applied inside the [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 download loop; output to data/raw/ or attachments/ based on classification
• [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25: add Step 8 coverage audit — run list_attachments_batch on archived emails, cross-check with disk, print report with warnings; optionally recursively expand thread siblings
• Add --backfill-attachments and --backfill-thread CLI flags for retroactive repair ([che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 Phase 3)
• Update skill version + CHANGELOG; deploy via /plugin-tools:plugin-deploy

Complexity

SDD-warranted, reasons:

1. 4 GitHub issues bundled — shared file + shared design decisions; proposing individually would re-litigate the same choices 4 times
2. New config schema — .claude/emails.md gains 3 new blocks with rules and precedence
3. Architecture decision: filter expansion strategy — subject matching vs thread-reference vs participant aliases. Multiple valid approaches with trade-offs
4. Architecture decision: routing rules precedence — extension vs keyword vs YAML config, which wins?
5. Strategy has 7+ steps with order dependencies: [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24 before [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 (expanded filter determines what gets downloaded); [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 before [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 (audit validates downloads); [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26 is orthogonal to [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 but shares code with [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23

Risks

• Skill length explosion: 202 lines → 400+. Consider factoring: extract routing table to a separate reference markdown if skill gets unwieldy
• MCP capability unknowns: mail__search_emails may not support subject-only or in-reply-to search. Need to verify before committing to [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24 strategy
• Filter expansion false positives: "taxometric" as subject keyword could match unrelated emails. YAML config must support both allowlist AND denylist
• Attachment download size: large attachments (video, big PDFs) could explode correspondence/ disk usage. Consider a size warn threshold in config
• /plugin-tools:plugin-deploy workflow: after skill update, deployment touches marketplace + cache. One-step plugin-deploy is documented; any manual deviations break sync

Recommended scope split

Two options:

Option	Shape	Trade-off
A. One bundled change archive-mail-attachment-coverage	All 4 in one Spectra change	Coherent story, one review pass; but ~40 tasks, harder to ship incrementally
B. Two-phase split	Phase 1: archive-mail-downloads-and-routing (#23 + #26, foundational). Phase 2: archive-mail-filter-and-audit (#24 + #25, validation layer).	Ships value incrementally — downloads work first, then filter expansion + audit catches gaps. Two smaller changes.

Recommendation: Option B (two-phase). Same rationale as ooxml-swift#1's A+D → B+C split worked well (validate fast, refine later). #23 + #26 are the "table stakes" (download + correct location); #24 + #25 are "robustness" (don't miss anything, verify what was got).

Per-issue status

• [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 (attachment download) — foundational, ships in Phase 1
• [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24 (filter asymmetry bug) — ships in Phase 2; requires thread/subject search strategy
• [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 (coverage audit) — ships in Phase 2; consumes [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23's output and [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24's expanded corpus
• [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26 (attachment routing) — ships in Phase 1 alongside [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 (same download code path)

[kiki830621]

Diagnosis (consolidated — #23 #24 #25 #26)

Type

mixed: bug (#24) + 3 features (#23 #25 #26), all in the same skill file commands/archive-mail.md and interdependent. Analyzed together as one bundle.

Context

Scanned the current skill at /Users/che/.claude/plugins/marketplaces/psychquant-claude-plugins/plugins/che-apple-mail-mcp/commands/archive-mail.md (202 lines, 7 steps). The skill today:

1. Takes a single filter string (sender/to substring)
2. Searches via mail__search_emails with query: filter
3. Generates Markdown per email with attachment filenames as plain text (no download, no MCP save_attachment call)
4. No post-hoc coverage check

MCP capabilities exist but skill never uses them: list_attachments, list_attachments_batch, save_attachment.

Dependency analysis

All 4 issues touch the same skill file. They are interdependent:

#24 filter asymmetry (bug)      ──┐
   └─ without this, thread       │
      completeness is broken     │
                                 ├──►  need to design together
#23 attachment download ─┐       │
   └─ Downloading makes  │       │
      coverage auditable ├──►    │
                         │       │
#25 coverage audit ──────┘       │
   └─ validates #23 did its job  │
                                 │
#26 attachment routing ──────────┘
   └─ decides WHERE #23 writes to
      (data/ vs correspondence/attachments/)

#23 is the foundational feature. #26 extends #23's output logic. #25 validates #23's output. #24 expands the input set that #23 operates on.

Root cause (for #24 bug)

Skill uses mail__search_emails with query: filter — matches sender or to. But collaboration threads include internal multi-party discussions where the external collaborator is not in sender/to (e.g., Che ↔ Lay ↔ Pei-Chi without Tatsuma). Those threads are invisible to the current filter.

Requirements breakdown (for #23/#25/#26)

#23 — attachment download:

• Use save_attachment MCP for each email with attachments
• Directory: {output_dir}/attachments/{email_md_stem}/{original_filename} (per baseline established in 2026-04-14 tatsuma backfill)
• Filename preservation: keep original bytes on disk; URL-encode %20 / %26 only in Markdown link
• Attachment block placement: after signature, before thread quote
• Reply without attachments: cross-reference original email's md (no duplicate download)

#25 — coverage audit:

• After Step 6, add Step 7: list_attachments_batch on archived emails, compare with disk
• 3 phases: (Phase 1) undownloaded-attachment warn, (Phase 2) thread-aware orphan detection, (Phase 3) --backfill-attachments / --backfill-thread flags for auto-repair

#26 — attachment routing:

• Extension-based rules: .csv/.sav/.dta/.parquet/.xlsx → data/raw/; .pdf/.docx/.txt/.md → attachments/
• Keyword override: filenames containing data|raw|indicators|codebook → data/; Submission|Figures|Manuscript|draft → attachments/
• Config: .claude/emails.md YAML attachment_routing block for per-project overrides
• Report: "15 attachments archived: 4 to data/raw/, 11 to attachments/"

Impact

• Affected file: plugins/che-apple-mail-mcp/commands/archive-mail.md (single file, 202 lines → likely ~350-400 after all 4)
• Affected config schema: .claude/emails.md gains 3 new optional fields:
  • subject_keywords (for [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24)
  • participant_aliases (for [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24)
  • attachment_routing (for [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26)
• Affected users: any caller of /archive-mail. Current behavior preserved when new fields are absent — additive.

Strategy

• [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24: expand Step 3 search with subject_keywords + optional participant_aliases matching; reference-chain follow via In-Reply-To (if MCP supports)
• [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23: add Step 6b (between Markdown generation and index update) that calls save_attachment for each email with attachments, writes to {output_dir}/attachments/{email_stem}/{filename}
• [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23: Markdown generation updates — insert Attachments block between signature and thread quote with URL-encoded links
• [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26: attachment routing logic (extension + keyword + YAML config) applied inside the [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 download loop; output to data/raw/ or attachments/ based on classification
• [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25: add Step 8 coverage audit — run list_attachments_batch on archived emails, cross-check with disk, print report with warnings; optionally recursively expand thread siblings
• Add --backfill-attachments and --backfill-thread CLI flags for retroactive repair ([che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 Phase 3)
• Update skill version + CHANGELOG; deploy via /plugin-tools:plugin-deploy

Complexity

SDD-warranted, reasons:

1. 4 GitHub issues bundled — shared file + shared design decisions; proposing individually would re-litigate the same choices 4 times
2. New config schema — .claude/emails.md gains 3 new blocks with rules and precedence
3. Architecture decision: filter expansion strategy — subject matching vs thread-reference vs participant aliases. Multiple valid approaches with trade-offs
4. Architecture decision: routing rules precedence — extension vs keyword vs YAML config, which wins?
5. Strategy has 7+ steps with order dependencies: [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24 before [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 (expanded filter determines what gets downloaded); [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 before [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 (audit validates downloads); [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26 is orthogonal to [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 but shares code with [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23

Risks

• Skill length explosion: 202 lines → 400+. Consider factoring: extract routing table to a separate reference markdown if skill gets unwieldy
• MCP capability unknowns: mail__search_emails may not support subject-only or in-reply-to search. Need to verify before committing to [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24 strategy
• Filter expansion false positives: "taxometric" as subject keyword could match unrelated emails. YAML config must support both allowlist AND denylist
• Attachment download size: large attachments (video, big PDFs) could explode correspondence/ disk usage. Consider a size warn threshold in config
• /plugin-tools:plugin-deploy workflow: after skill update, deployment touches marketplace + cache. One-step plugin-deploy is documented; any manual deviations break sync

Recommended scope split

Two options:

Option	Shape	Trade-off
A. One bundled change archive-mail-attachment-coverage	All 4 in one Spectra change	Coherent story, one review pass; but ~40 tasks, harder to ship incrementally
B. Two-phase split	Phase 1: archive-mail-downloads-and-routing (#23 + #26, foundational). Phase 2: archive-mail-filter-and-audit (#24 + #25, validation layer).	Ships value incrementally — downloads work first, then filter expansion + audit catches gaps. Two smaller changes.

Recommendation: Option B (two-phase). Same rationale as ooxml-swift#1's A+D → B+C split worked well (validate fast, refine later). #23 + #26 are the "table stakes" (download + correct location); #24 + #25 are "robustness" (don't miss anything, verify what was got).

Per-issue status

• [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 (attachment download) — foundational, ships in Phase 1
• [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24 (filter asymmetry bug) — ships in Phase 2; requires thread/subject search strategy
• [che-apple-mail-mcp] archive-mail 缺 thread 完整性 verify 步驟（漏 14 個附件未察覺） #25 (coverage audit) — ships in Phase 2; consumes [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23's output and [che-apple-mail-mcp] archive-mail filter 不對稱：抓不到三方內部討論（漏 21 封 + 14 附件） #24's expanded corpus
• [che-apple-mail-mcp] archive-mail 應分流附件：研究資料（xlsx/csv）vs 文件附件（PDF/docx） #26 (attachment routing) — ships in Phase 1 alongside [che-apple-mail-mcp] archive-mail 未下載附件（僅記錄檔名） #23 (same download code path)