[META] codify 'AI design 階段過抽象,需要 third-party surfacing' as plugin-level principle (residue from #135)

[kiki830621]

Source

Residue from /idd-close #135 (Step 3.6,IC_R011 #105 checkpoint)。

#135 diagnosis comment ### Residue line 1:

Meta-rule about meta-design:本 session 「AI design 階段過抽象,需要 third-party surfacing」現象本身是否要 codify 成 plugin-level principle(類似 IDD MANIFESTO 加一條「skill 設計輪廓 audit」)。本 issue 不 operationalize 這個 — 它屬於 future iteration 議題。

Background

本 session 提供了具體 evidence:AI 在 #135 design 階段連 4 輪 over-abstract(獨立 skill → diagnose sub-step → idd-issue inline → 才到 composable primitive),user 連 4 輪 push back 才落地。Recursive case — AI 設計 /idd-clarify(catch terminology under-verification)時自己也 under-verified。

Open questions

• 是否在 MANIFESTO 加新 section「Skill design self-audit」/「AI over-abstraction discipline」?
• 用什麼 mechanism enforce?(spectra-discuss 階段強制 surface assumption + push-back round counter?)
• 跟既有 verify ensemble 怎麼互補?(verify catch implementation regression;這個 catch design over-abstract)

Type

enhancement / meta-principle design

Priority

P3 — wait for ≥3 instances of「AI design over-abstract」 caught in real spectra-discuss / spectra-propose sessions before codifying(rule of three)。Current session #135 is instance #1。

Refs #135

Current Status

Phase: diagnosed
Last updated: 2026-05-25 by idd-diagnose (batch)

Complexity

Spectra when triggered (parking lot)

Routing

See diagnosis comment Strategy section for explicit WAIT trigger condition + next-step path.

[kiki830621]

Diagnosis

Type

enhancement / meta-principle design — codify「AI design 階段過抽象,需要 third-party surfacing」現象成 plugin-level principle

Analysis

#135 residue 第 1 條 — recursive insight:/idd-clarify 設計目的是 catch issue body terminology gap,而 AI 設計 /idd-clarify 自己時卻過抽象(連 4 輪 design over-abstract,user push back 才落地)。

Instance count: 1(本 session #135 是第 1 個有 evidence 的 case)。Rule of three 還沒達標 → 不 codify。

但 issue 自己已列 3 個 open questions 推測未來實作方向:

1. MANIFESTO 新 section「Skill design self-audit」/「AI over-abstraction discipline」
2. Mechanism 候選:spectra-discuss 階段強制 surface assumption + push-back round counter
3. 互補 verify ensemble — verify catch implementation regression;這個 catch design over-abstract

Strategy

DEFER 至 ≥3 instances 出現:

• 紀錄 [META][idd-issue] terminology trust gap — source 用錯 domain term 時 AI 繼承錯誤 #135 session 為 instance idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1(diagnosis 已 audit)
• 待 instance idd-all: direct-commit mode 加 passive warning (dirty tree / on default branch) #2、idd-all SKILL.md: pseudocode quality (goto bash / branch-exists no-op / stale Failure Modes table) #3 出現(any spectra-discuss / spectra-propose session 出現「AI 過抽象 → user push back ≥3 round」pattern)→ 自動 surface 候選 case 到本 issue
• 達 3 instances → 重 diagnose + 進 Spectra discuss(MANIFESTO change 是 plugin-level principle = published contract)

現在能做的小事(non-blocking parking lot 維護):

• 在 [META] codify 'AI design 階段過抽象,需要 third-party surfacing' as plugin-level principle (residue from #135) #144 comment 紀錄 instance idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1 details([META][idd-issue] terminology trust gap — source 用錯 domain term 時 AI 繼承錯誤 #135 session 4-round over-abstract sequence)— 已隱含於本 diagnose,explicit comment 可加
• 觀察 future spectra session,instance idd-all: direct-commit mode 加 passive warning (dirty tree / on default branch) #2/idd-all SKILL.md: pseudocode quality (goto bash / branch-exists no-op / stale Failure Modes table) #3 出現時手動 comment 紀錄 evidence

Complexity

Spectra when triggered(MANIFESTO 是 plugin-level published contract,影響所有 user;架構性質決定 design 必走 Spectra discuss)。但目前 = parking lot,wait for instance #2 + #3。

Risks

• R1 Recursive over-codification:codify「不要 over-abstract」這個 principle 本身可能 over-abstract(meta-irony)。Mitigation:rule-of-three 強制 ≥3 evidence 才碰
• R2 Verify ensemble 已有 redundancy:verify catch implementation regression + DA reviewer challenge 群體盲點 — design over-abstract 可能由 DA reviewer 在 design.md 階段抓到(已部分覆蓋)。可能不需獨立 mechanism。Mitigation:wait + see
• R3 「Design over-abstract」definition fuzzy:何為「過抽象」是判斷不是 binary,可能難 codify 成 testable rule。Mitigation:rule-of-three 收集 evidence 後可分析共同 pattern

Residue

(NSQL §4.6):「self-aware AI design」這個 horizon — AI 設計 skill 時對自己 design tendency 的 reflexive awareness。本 issue 試圖 codify mechanism 來 catch 這個 tendency,但「self-awareness as discipline」本身需要 cultural enforcement(user/maintainer push back),不是 spec 能 force。

---

Vagueness Pre-check — Final

• **V1=3\n- **V4=3
• Triggered: no — both axes ≤ 3 (issue self-framed clear with explicit open questions + trigger condition)

Sister Concerns Filed (mid-diagnose, v2.47.0+ #528;v2.72.0+ default-flip #148)

(none surfaced — this issue IS itself a sister of #135 cluster (alongside #136 #137 #143 #144 #145 #146);cross-sister relationships already documented in each diagnose Strategy / Related section. Per IC_R011 §3 exemption "existing issue covers" — no net-new orphan to file.)