idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion

[kiki830621]

Problem

Original request (使用者於 thesis project session 中提出，2026-05-03):
「idd-all 我想要有一種是直接 commit 的方案，因為 PR 不一定很方便，我其實是在旁邊，這樣的 all 中間還是可以問我問題，總之 all 可以分成 human-in-the-loop 跟不是的情況」

當前 idd-all v2.40.0 強制走 PR path 且強制 unattended（所有 sub-skill 的 AskUserQuestion 都被 suppress）。這個設計對「fire-and-forget orchestration」很合適，但對「使用者就在旁邊、想監控進度也想被諮詢」的常見場景反而是 friction：

1. PR 對小型 / 個人 repo 是 overhead — 沒有 review 流程的 solo project，每個 issue 開 feature branch + PR 是純 ceremony
2. Sub-skill 的 attended 預設被一律 override — 像 spectra-discuss 的多輪對話、spectra-propose 的 Park/Apply 抉擇、idd-implement 的 plan-tier approval gate，全被靜默跳過。當使用者就在旁邊時，這些 checkpoint 反而是它們存在的意義

Type

feature / enhancement

Proposed design (待 diagnose 細化)

兩種模式

Mode	Path	Sub-skill questions	Stop point	Use case
Unattended (current v2.40.0)	PR (forced)	All suppressed	Verify → stop	Background automation, scheduled cron, /loop
Attended (HITL) (NEW)	Direct-commit (current branch)	Allowed at natural points	TBD: verify-stop or full-cycle?	使用者就在旁邊、想被諮詢

Open design questions (留給 diagnose)

1. Flag 設計 — --hitl / --attended / --no-pr？要不要乾脆把預設改成 attended（因為直接從 CLI 喊的用例多半是現場互動），unattended 變成 opt-in via --unattended？
2. Branch 策略 — Attended 模式還要不要強制 feature branch？還是允許在當前 branch 直接 commit？若使用者已經在 feature branch（例如自己手動切過去），維持當前 branch 是合理的；若在 main，仍然應該建 feature branch（保險）？
3. Verify 之後 — Attended 模式既然使用者就在旁邊，要不要直接連 idd-close 也跑？還是維持「verify 後停下、由人決定 close」？
4. Spectra path interaction — Attended 模式的 sub-skill suppress hint 整個拿掉？還是仍然保留某些 checkpoint suppress（例如 spectra-propose Step 10 Park/Apply）？「都讓 sub-skill 自己決定」是最乾淨的，但要 audit 每個 sub-skill 的 attended 預設是否合理。
5. 混用情境 — Attended 模式中途使用者可能離開 keyboard，sub-skill 卡在 AskUserQuestion 等不到回應 — 是否需要 timeout fallback？或者明確記錄 attended 模式假設「使用者全程在線」？
6. Backward compat — 現有 idd-all #N --pr 的 caller（包含 /loop 自動化）期望 unattended 行為。若預設改成 attended，需要 deprecation 期 + warning。
7. 與 idd-implement --no-pr 的關係 — idd-implement 已經透過 pr_policy config 支援 PR / direct-commit / ask 三選一。Attended idd-all 是不是直接用 pr_policy 而不是新增 flag？這樣 config 介面比較一致。

Impact

• 新使用者第一次跑 idd-all 會比較直覺（不用先理解「為什麼一定要開 PR」）
• 解放 sub-skill 的 attended 設計（spectra-discuss 多輪對話、Plan tier EnterPlanMode 等）— 讓 unattended override 變成 opt-in 而不是 default
• 對 solo repo / personal project 的 friction 大幅降低

Acceptance criteria（draft，待 diagnose 確認）

• idd-all --no-pr #N 走 direct-commit + 允許 sub-skill AskUserQuestion — naming chosen as --no-pr (reuse idd-implement vocab); SKILL.md Step 0.2/Phase 0.5; smoke test Case B
• 既有 idd-all #N --pr 行為不變（unattended PR path）— smoke test Cases A + H + F (absent config /loop default); regression reviewer round-1 confirmed byte-equivalent v2.40.0 behavior
• Documentation 清楚說明兩種模式的契約差異 + 何時用哪個 — SKILL.md frontmatter + 核心原則 + 2 example traces; pr-flow.md "idd-all path resolution" section; usecase-routing.md row 11+11b; CLAUDE.md skill summary
• Sub-skill args 的 unattended hint 機制改成 opt-in（不是 always-on）— Phase 3a/3b/4 conditional if [ "$INTERACTION" = "unattended" ] guards; spectra-apply attended branch SHALL NOT inject any directive (round-2 fix)
• 新增 / 修改的測試或 dry-run trace 證明兩種模式都按設計走 — openspec/changes/idd-all-hitl-mode/smoke-test.sh (28/28 assertions PASS); smoke-test-trace.txt captured
• 至少一個 sample run 在每個模式各跑通一個 issue — smoke test Cases A/H (PR mode) + Case B (direct-commit mode) executes resolution + sub-skill arg construction end-to-end with deterministic assertions

Related

• 當前 idd-all skill: plugins/issue-driven-dev/skills/idd-all/SKILL.md (v2.40.0)
• Sub-skill 的 attended 預設文件: plugins/issue-driven-dev/references/pr-flow.md
• v2.40.0 release notes: cross-repo --cwd 支援（先前 milestone）

---

Current Status

Phase: closed
Last updated: 2026-05-04 by idd-close

Key Decisions

• Closing summary posted: idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1 (comment)
• Round-1 verify FAIL → fix → Round-2 verify PASS → smoke test 28/28 PASS — full audit chain
• 設計重構為兩軸正交化（path × interaction 從 pr_policy + --pr/--no-pr 同源推導），不引入新 --hitl flag
• pr_policy 缺省 → idd-all 默認 (PR, unattended) (v2.40.0 backward compat); 顯式 pr_policy: ask 才走 AskUserQuestion — idd-implement vs idd-all 在缺省時刻意 diverge (intentional, documented in pr-flow.md)
• 不自動 close — idd-all 兩 mode 都停在 verified;closing summary 由人寫(本 closing comment 即實踐)
• HITL mode 不加 silent timeout;明示「attended assumes user in session」
• Backward compat: /loop 等顯式 --pr caller 零受影響;新缺省規則保護無 config 的 caller 不 hang

Scope Changes

• direct-commit mode 容許 dirty working tree (verify 標 P2 scope creep) → 已開 idd-all: direct-commit mode 加 passive warning (dirty tree / on default branch) #2 follow-up
• spec.md 加入兩個 backward-compat scenarios (no-flag-no-config, explicit-ask) 對齊 round-2 implementation

Blocking

• (none) — 所有 27 spec tasks done;3 P0 + 5 P1 round-1 blockers closed via 5fc39fc;AC idd-all Phase 0.5: fork detection 改 fail-closed (gh repo view 失敗時預設 PR mode) #5/idd-all SLUG derivation: BSD/GNU sed alternation (\| not portable) #6 closed via smoke test
• 12 個 P2/P3 follow-up 整合為 9 個 focused issue (idd-all: direct-commit mode 加 passive warning (dirty tree / on default branch) #2-idd-all Examples: 跨 repo invocation 段重複描述,合併 #10),全部 ref Related: idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1

Commits

• f7230b8 test: idd-all Phase 0.5 smoke test — 28/28 assertions PASS, AC idd-all Phase 0.5: fork detection 改 fail-closed (gh repo view 失敗時預設 PR mode) #5/idd-all SLUG derivation: BSD/GNU sed alternation (\| not portable) #6 PASSED (Refs idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1)
• 5fc39fc fix: idd-all HITL mode — close P0/P1 verify FAIL blockers (Refs idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1)
• 98b7a71 feat: idd-all HITL mode — Phase 0.5 (path, interaction) tuple resolution (Refs idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1)

[kiki830621]

Diagnosis

Type

feature / enhancement (orchestrator behavior + config surface extension)

Analysis

Reframing: 這個需求其實混了 兩個正交軸 — 把它們拆開看，design space 比 issue body 描述的更清楚。

軸	兩端	影響
A. Persistence path	--pr (feature branch + PR) ↔ --no-pr (current branch direct commit)	影響 git 結構、CI gating、code review
B. Interaction style	--unattended (suppress all sub-skill questions) ↔ --attended (allow natural pacing)	影響 sub-skill AskUserQuestion / EnterPlanMode 是否啟用

現狀 (v2.40.0): idd-all 把這兩軸強制綁定 — 寫死 --pr + --unattended。所以 issue body 提的「HITL 模式」實際上是要解開這個綁定。

重要發現: references/pr-flow.md 已經為 idd-implement 定義了 pr_policy config (always|never|ask)。idd-all 目前完全不讀 pr_policy，硬覆蓋成 PR path。這就是 Q7 直指的不一致 — 而且不只命名不一致，是 idd-all 直接無視 user 對 repo 的 path 偏好設定。

Strategy

核心設計決定: 兩軸正交化 + 從 pr_policy 衍生預設。

設計 (proposed)

Config	Per-invocation flag	預設行為
pr_policy: always	--pr	unattended PR path（現狀，自動化用）
pr_policy: never	--no-pr	HITL direct-commit path（NEW）
pr_policy: ask (or absent)	--pr / --no-pr 二選一	AskUserQuestion at Phase 0

衍生規則: pr_policy 同時推導 path 和 interaction style:

• pr_policy: never → 暗示「solo / 個人 repo」→ user 通常在現場 → 預設 attended
• pr_policy: always → 暗示「自動化 / fire-and-forget」→ 預設 unattended

override: 罕見情境（attended PR / unattended direct）可用獨立 flag:

• --attended / --unattended 顯式覆蓋 interaction style
• --pr / --no-pr 顯式覆蓋 path

實作上，先做 derived defaults（覆蓋 80% 用例），interaction-style flag 留 v2.41+ 評估必要性。

對 7 個 open design questions 的回答

Q	答	理由
1 flag 命名	重用既有 --pr / --no-pr，不新增 --hitl	與 idd-implement 對齊；users 已經會
2 branch 策略	直接套用 pr-flow.md 既有規則：PR path 強制 feature branch from default; direct-commit path 維持 current branch	避免 idd-all 跟 idd-implement 兩套 branch 邏輯
3 verify 後 close	NO — 仍停在 verify。Closing summary 含 root cause / solution narrative，人寫才有審計價值	跟現狀一致；HITL 不等於 auto-close-everything
4 Spectra path interaction	direct-commit + attended 模式下，不傳 unattended hint 給 sub-skills，讓 spectra-discuss 自然多輪 / spectra-propose 問 Park/Apply / idd-implement plan-tier 走 EnterPlanMode	這就是 user 要的
5 中途離開 keyboard	不加 timeout。明確記錄「attended mode assumes user in session」	timeout 設多久都會踩雷；放棄遠端執行情境
6 backward compat	預設不變（即 pr_policy: ask 時，AskUserQuestion 第一次仍預設 PR path）。/loop 等自動化呼叫者顯式帶 --pr 不受影響	零 breaking change
7 與 idd-implement --no-pr 關係	重用 pr_policy — idd-all 不再硬覆蓋，按 pr_policy 解析	一致介面

Impact

修改檔案:

1. plugins/issue-driven-dev/skills/idd-all/SKILL.md — 主要改動：拆掉 hardcoded "Always PR path"，改讀 pr_policy；Phase 3a/3b 的 sub-skill args 移除 always-unattended hint，改 conditional
2. plugins/issue-driven-dev/references/pr-flow.md — 補充 idd-all 也消費 pr_policy 的契約（目前文件只說 idd-implement）
3. plugins/issue-driven-dev/references/usecase-routing.md (新) — 新增 attended vs unattended decision matrix
4. plugins/issue-driven-dev/skills/idd-all/SKILL.md Phase 3b（spectra chain）— 改成 conditional unattended hint

不需要改的:

• idd-implement / idd-verify / idd-close — 這些已經有自己的 path resolution 跟 attended 預設，沒被打破
• spectra-* skills — 它們的 attended 預設是對的；改的只是 idd-all 是否覆蓋
• references/pr-flow.md 的核心 algorithm — 已 generic，只需加註 idd-all 也適用

Implementation tasks

• 修改 idd-all/SKILL.md Phase 0.5 (新增 path resolution from pr_policy) — 98b7a71 + refined in 5fc39fc (split absent vs explicit ask)
• 修改 idd-all/SKILL.md Phase 3a (idd-implement 呼叫條件化 --pr/--no-pr + UNATTENDED MODE) — 98b7a71
• 修改 idd-all/SKILL.md Phase 3b (spectra chain conditional unattended hint) — 98b7a71 + spectra-apply attended branch fixed in 5fc39fc (SHALL NOT inject)
• 修改 idd-all/SKILL.md Phase 5 (PR creation conditional, direct-commit short-circuits to Phase 6) — 98b7a71
• 修改 idd-all/SKILL.md 核心原則段 — 98b7a71 (replaced "Always PR path" / "Unattended assumption" / "Stop before close" with two-axis resolution + attended-assumes-user + verify-is-terminal)
• 修改 idd-all/SKILL.md argument-hint — 98b7a71 (added [--pr | --no-pr] and --no-pr 觸發 HITL direct-commit + attended)
• 更新 references/pr-flow.md — 98b7a71 (new "idd-all path resolution" section) + cleanup in 5fc39fc (lines 26-27, 55, 175, 186 stale text)
• [~] 新增 references/usecase-routing.md — superseded: file already existed pre-idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1; instead added row 11 + new row 11b inline (5fc39fc)
• 更新 SKILL frontmatter description / argument-hint — 98b7a71 (also captured as task 7.1)
• 加至少 1 個 sample run trace (PR + HITL) 到 SKILL.md examples — 98b7a71 (Trace 1 + Trace 2); also added executable smoke test in f7230b8

Risks

風險	緩解
Sub-skill 在 attended mode 卡住等不到 user 回應	文件明示「attended assumes user in session」；不加 silent timeout
User 一個 repo 的 pr_policy: never 配合 cross-repo invocation 時 confused（other repo 的 config 不適用）	cross-repo --cwd 已 walk-up resolve config（cross-repo-cwd.md 既有契約），會解到 target repo 的 config，正確
Spectra path 在 attended mode 跑很久（spectra-discuss 多輪）— /loop 自動化呼叫到此會卡死	/loop 自動化必須顯式帶 --pr 進 unattended 模式；零 backward incompatibility
「混合 mode」（attended PR / unattended direct）使用者寫了 config 期望但被 v2.41 推遲	v2.41 第一階段不支援 explicit --attended/--unattended flag；只支援 derived from pr_policy。文件明說

Complexity

Spectra

觸發項:

• Layer 2 (Necessary) ✓ — idd-all 是 published skill，flag/config 是給 future callers (含 /loop 自動化) 的 frozen contract
• Layer 3 — modifies normative behavior of existing spec ✓ — pr-flow.md 既有契約「idd-all 強制 PR path」會被改
• Layer 3 — affects 2+ existing specs needing consistency ✓ — pr-flow.md + idd-all SKILL.md + idd-implement SKILL.md + cross-repo-cwd.md 都要協同
• Layer 3 — architectural decision with long-term maintenance implications ✓ — orchestrator path/interaction model 是會被 future skills 繼承的結構性決定

Recommended next step

走 /spectra-discuss（default，不 opt-out propose） — 7 個 open questions 雖然在這份 diagnosis 提出了我的推薦答案，但其中 Q1（flag 命名）、Q4（unattended hint 拆解的精準位置）、Q7（pr_policy 衍生 vs orthogonal flag）都需要與 user 對齊才有 single source of truth。spectra-discuss 把這些 assumption 列出讓 user 修正後，再進 spectra-propose。

[kiki830621]

Verify: #1

Engine

Agent Team (5 Claude reviewers) + Codex (gpt-5.5 xhigh)

• requirements / logic / security / regression / devils-advocate
• 獨立 Codex 跨模型盲驗

Diff verified: HEAD~1..HEAD (commit 98b7a71, 7 files, 646 insertions / 126 deletions)

Result

FAIL — 3 blocking findings (P0/P1), 5 contract gaps (P1), 9 follow-up items (P2/P3)

要求覆蓋率

AC	Status	主要 evidence
1. --no-pr 走 direct-commit + 允許 sub-skill AskUserQuestion	FULLY (設計層) / NOT (runtime — Finding 1 阻擋)	spec.md:18-22, SKILL.md:160-161, 227-234
2. idd-all #N --pr 行為不變	FULLY byte-equivalent (regression reviewer 確認)	SKILL.md:189-222, 524-555
3. Documentation 兩 mode 契約差異	FULLY (但 4 處 stale doc — Finding 5)	SKILL.md:24-34, 633-682, pr-flow.md:175-190
4. Sub-skill unattended hint opt-in	PARTIALLY (Finding 6 違反 SHALL NOT inject)	5 處 conditional guard ✓; spectra-apply attended branch 仍注入 prompt segment ✗
5. Test / dry-run trace 證明兩 mode	PARTIALLY	Documented trace ✓ (SKILL.md:635-682); actually executed test ✗
6. 每 mode 各跑通一個 sample run	NOT	tasks.md 8.1/8.2/8.3 全 deferred to user post-merge

Findings (合併去重 6 reviewers)

🔴 P0 — Blocking (must fix before close)

#	Severity	Finding	Source	Action
1	P0	Step 0.2 不 parse --pr/--no-pr — Phase 0.5 引用 $PR_FLAG 與 $CONFIG_PATH 兩變數，但 Step 0.2 只處理 --cwd。整條 mode resolution 無法執行。spec Requirement 1 的 "explicit --no-pr flag" 與 "backward-compatible default" scenario 都 unreachable。	devils:1, logic:1	Blocking
2	P0	AskUserQuestion 被當 shell command 用 — ANSWER=$(AskUserQuestion ...) 在 bash subshell 內執行；AskUserQuestion 是 Claude tool 不是 CLI。silently fail。pr_policy: ask (default) caller 整條 fallback 是 ghost path。直接違反 proposal "no backward incompat" — /loop 只帶 issue 號的 caller 會 hang。	devils:2, logic:2, codex:2	Blocking
3	P0	pr_policy: ask 預設行為文件矛盾 — proposal 說「第一次呼叫仍預設 PR path」(零 backward incompat)；spec.md:14 與 SKILL.md:512 說「ask = AskUserQuestion」。兩個語意不相容。/loop 既有 caller 沒帶 --pr 在 ask 配置下會卡死。	codex:risk-2, logic:2	Blocking

🟠 P1 — Contract violations

#	Severity	Finding	Source	Action
4	P1	--pr + --no-pr 並用未偵測 — precedence chain 第一條 wins，但 user 同時帶兩 flag 沒 explicit conflict abort。	logic:1, devils:1	Block / In-scope fix (≤5 行)
5	P1	4 處 stale doc 仍說 "always PR path" — pr-flow.md:5,55、usecase-routing.md:44、plugins/issue-driven-dev/CLAUDE.md:44；新 pr-flow.md 段標 "v2.44.0+" 應是 "v2.46.0"。	devils:3	Block / In-scope fix (sed-mass-replace)
6	P1	Attended 仍注入 directive 違反 spec — Phase 3b spectra-apply attended 分支仍加 "Native AskUserQuestion / continue-confirmation prompts may fire — user is in session"。Spec 4.94: "MUST NOT inject any UNATTENDED MODE directive" — 字面違反 (spec 寫的是 unattended directive 但 spirit 是任何 attended-branch directive)。	devils:5	Block / In-scope fix
7	P1	Attended SDD path 沒 fallback capture Conclusion: line — 該 format 只在 unattended hint block 內被強制；attended 沒 directive、sub-skill 不一定 emit。下游 propose chain 接到空字串。	devils:6	Block / In-scope fix
8	P1	Phase 6 stdout 文案 cosmetic drift — "Phase 5: Open PR" → "Push + Open PR"; "Verify: PASS (6-AI)" → "Verify: PASS"; "Next steps (manual): 1./2./3." → 單行。/loop 與其他 stdout-scraper caller 風險。	regression:1, devils:4	Follow-up (測 /loop 後再修)

🟡 P2 — Logic / safety gaps (build-stage fixes)

#	Severity	Finding	Source	Action
9	P2	direct-commit 跳過 clean-tree gate，HITL user 若有 staged .env/secret 會 commit。建議 passive warning (不 abort)	security:1, devils:8	Follow-up
10	P2	goto Phase 6 在 fenced bash block — bash 沒 goto，照貼會出錯	logic:3	Follow-up (改 if [ "$PATH_AXIS" = "PR" ]; then ... fi)
11	P2	PR-mode branch-already-exists handler 是 no-op (:) 後接必 fail 的 git checkout -b；Failure Modes 承諾的 AskUserQuestion 不存在	logic:5	Follow-up
12	P2	MAX_ROUND 出現在 user-facing abort 字串字面 (沒被 substituted)	logic:6, regression:3	Follow-up
13	P2	Failure Modes table 仍標 "Phase 0.3 abort"，但 dirty/branch gate 已搬到 Phase 0.5	logic:7, regression:2	Follow-up
14	P2	follow-up issue title sanitization — idd-issue 收 verify finding text 沒 length/control-char 過濾	security:5	Follow-up

🟢 P3 — Defensive / portability (file as separate issues)

#	Severity	Finding	Source	Action
15	P3	fork detection || echo "false" fail-open；gh repo view 失敗時違反 spec "fork forces PR"	security:2	Follow-up
16	P3	SLUG sed BSD/GNU | 行為差異	security:3	Follow-up
17	P3	PR body here-doc 非 quoted — 目前 placeholder 是註解，未來 maintainer 真填 ${TITLE} 會 RCE	security:4	Follow-up
18	P3	GITHUB_REPO 衍生值缺 post-derive shape assert	security:6	Follow-up
19	P3	direct-commit + main branch 缺顯眼 warning (commits 直 land 主 branch，沒 PR gate)	devils:8	Follow-up
20	P3	跨 repo example 兩段重複描述	regression:4	Follow-up

Scope Check

• 改動範圍:SKILL.md, pr-flow.md, openspec/changes/idd-all-hitl-mode/{proposal,specs,tasks} — within idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1 scope
• scope creep:direct-commit 容許 dirty working tree 是 acceptance criteria 沒明說的設計選擇 (codex:risk-3)。建議補進 spec 或加 warning

Devil's Advocate Verdict

兩個 P0 finding 合起來足以斷言「兩種模式按設計走」是不可達的。implementation 沒實際跑過 — markdown 把 $PR_FLAG 寫進 conditional 但變數從未賦值,而 pr_policy: ask 的 fallback 是 syntax error。Reviewer 不應將 #1 標 PASS。

Routing

依鐵律:有 blocking findings 就不 close。

Required for close:

1. 修 P0 Finding 1-3 (補 --pr/--no-pr/--cwd parser + 把 ask-policy 從 bash 拿出來變 agent-level prose)
2. 修 P1 Finding 4-7 (in-scope fix)
3. P1 Finding 8 stdout drift 視 /loop 測試結果決定

P2/P3 follow-ups:14 個 (Findings 9-20),建議批次開新 issue 不阻擋本 #1 close。

Verify trail

• Reviewer findings: /tmp/verify_1_findings_{requirements,logic,security,regression,devils_advocate}.md
• Codex output: /tmp/codex-verify-1.md
• Diff: /tmp/idd_verify_1_diff.patch (1038 lines)
• Commit: 98b7a71 ("feat: idd-all HITL mode — Phase 0.5 (path, interaction) tuple resolution (Refs idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1)")

[kiki830621]

Follow-Ups Filed (P2/P3 from verify FAIL)

12 P2/P3 findings from #issuecomment-4366336928 batch-filed as 9 focused issues:

Bucket	Issue	Severity	Findings consolidated
direct-commit safety warnings	#2	P2/P3	9, 19
SKILL.md pseudocode quality	#3	P2	10, 11, 13
Phase 4 ${MAX_ROUND} substitution	#4	P2	12
Fork detection fail-closed	#5	P3	15
SLUG sed BSD/GNU portability	#6	P3	16
PR body heredoc quoting	#7	P3	17
GITHUB_REPO shape assert	#8	P3	18
idd-issue title sanitization	#9	P2 (cross-skill)	14
Example dedup	#10	P3	20

All 9 ref Related: #1. None blocks #1 close — the 3 P0 + 5 P1 in the verify report are the actual blockers. P2/P3 are queued as separate work.

[kiki830621]

Verify (round 2): #1 — PASS

Engine

Focused single Devil's Advocate re-verify against the 8 specific blockers from round-1 (#issuecomment-4366336928).

Diff verified: 5fc39fc (this commit) on top of 98b7a71 — 5 files changed, +144/-39 lines.

Result

PASS — all 3 P0 + all 5 P1 blockers from round-1 resolved.

#	Blocker	Round-1	Round-2
P0 #1	$PR_FLAG/$CONFIG_PATH referenced but unset	FAIL	PASS — Step 0.2 expanded
P0 #2	AskUserQuestion as bash $(...)	FAIL	PASS — case label hands off to prose
P0 #3	pr_policy: ask vs absent semantic conflict	FAIL	PASS — split into rules 6 & 7 in spec + impl
P1 #4	--pr + --no-pr simultaneous undetected	FAIL	PASS — argv loop aborts
P1 #5	4 stale docs claim "always PR"	FAIL	PASS — all 4 + 2 residuals (pr-flow.md:55, 186, 26-27) cleaned
P1 #6	spectra-apply attended branch injected directive	FAIL	PASS — branch now no-ops
P1 #7	Attended Conclusion: capture would fail	FAIL	PASS — fallback chain (line → paragraph → AskUserQuestion)
P1 #8	Phase 6 stdout cosmetic drift	DEFERRED	DEFERRED (no new drift in round-2)

Spec changes (round-2 alignment)

• Resolution precedence: ask (or absent) split into 2 rules (explicit ask vs absent)
• 2 new scenarios:
  • "backward-compatible default (no flag, no config)" — covers /loop on fresh repo
  • "explicit ask requires user choice" — covers AskUserQuestion path
• Renamed existing "backward-compatible default" → "(explicit --pr)" for clarity

spectra validate --changes idd-all-hitl-mode PASS, spectra analyze Clean (0 findings).

Documented intentional divergence

pr-flow.md now explicitly notes that idd-implement and idd-all diverge on absent config by design:

• idd-implement absent → AskUserQuestion (interactive single-issue context)
• idd-all absent → (PR, unattended) (orchestrator with /loop callers; absent must not hang)

Routing

• ✅ All blocking findings closed
• ⏸️ P2/P3 follow-ups (idd-all: direct-commit mode 加 passive warning (dirty tree / on default branch) #2-idd-all Examples: 跨 repo invocation 段重複描述,合併 #10) remain open but don't block idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1 close
• ⏸️ AC idd-all Phase 0.5: fork detection 改 fail-closed (gh repo view 失敗時預設 PR mode) #5 (executed test trace) and AC idd-all SLUG derivation: BSD/GNU sed alternation (\| not portable) #6 (live sample run per mode) still deferred to user post-merge smoke testing — that's the only AC gap remaining

Recommendation: ready for /idd-close #1 once user confirms AC #5/#6 deferral is acceptable (or runs the smoke tests to flip them PASS).

Verify trail

• Round-1 master: idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1 (comment)
• Round-2 reviewer report: /tmp/verify_1_round2.md
• Diff: /tmp/idd_verify_1_round2_diff.patch (270 lines)
• Commits: 98b7a71 (initial) → 5fc39fc (fixes)