bug: /idd-verify 結束沒確認 teams 被關掉 — TeamDelete fails on idle-but-active reviewers + no cleanup SOP

[kiki830621]

Problem

User observation, 2026-05-11 session:
「idd-verify結束知後好像沒有確認teams是不是有被關掉」
— surfaced after this session's /idd-verify --pr 58 (verify-pr58 team)

/idd-verify SKILL.md Step 0 Bootstrap TaskList 包含 launch_parallel_reviewers (TeamCreate + 5 Agent + Codex) 和 wait_for_claude_agents (wait until idle + read findings file)，但 沒有 explicit cleanup step confirming the team is torn down after verify completes。

實際 session 觀察 (verify-pr58):

• 5 reviewer agents 全部 idle + 寫完 findings file ✓
• Codex CLI 完成 + master comment posted to PR ✓
• TeamDelete 嘗試 → FAIL with: Cannot cleanup team with 5 active member(s): requirements, logic, security, regression, devils-advocate. Use requestShutdown to gracefully terminate teammates first.
• SendMessage shutdown_request → 5 teammates ack (idle notifications) but don't send shutdown_response{approve:true} → team remained active
• Subsequent TeamDelete 重試持續失敗,team 卡在 active state 持續到 session 結束

Type

bug (process gap in /idd-verify SOP)

Expected

/idd-verify Step 0 Bootstrap TaskList 應該有 explicit cleanup task (e.g. cleanup_team) after post_master_and_pointers + restore_working_tree:

TaskCreate(name="cleanup_team", description="Step 5.5 (NEW): SendMessage shutdown_request to all 5 reviewers, wait for shutdown_response approve, then TeamDelete. Fallback: if teammates don't respond within timeout, document active-team residue in final report (non-fatal) so user can manually TeamDelete next session.")

Either:

1. Mechanical fix — SKILL.md adds Step 5.5 cleanup_team after restore_working_tree
2. Documented fallback — if team cleanup is fundamentally async / unreliable in current Claude Code teammates protocol, document the limitation + provide manual TeamDelete retry hint

Actual

無 cleanup step in SKILL.md。Teammates 在 idle 但 active 狀態下殘留,user 必須:

• 手動嘗試 TeamDelete (fail)
• 手動 SendMessage shutdown_request to each (5 calls)
• 手動 SendMessage 提示 teammate respond with shutdown_response (also no response)
• 接受 team 殘留 + 等 session 結束 cleanup

Impact

• Workflow friction: 跑完一次 /idd-verify 後 next TeamCreate (e.g. for /idd-verify on another PR) fails with "Already leading team 'verify-prXX'" — user must work around via subagent_type=general-purpose without TeamCreate (lost team coordination value)
• Resource leak: 5 active teammates per verify run not cleanup → multi-verify session has 10/15/20 active teammates idling
• SOP gap: any future skill author looking at /idd-verify SOP doesn't see cleanup pattern → likely repeated in other team-using skills

Related

• Session evidence: verify-pr58 team cleanup attempt 2026-05-10 ~17:30 PST → 18:00 PST (multiple TeamDelete retries)
• Similar pattern in /idd-verify --pr 67 / --pr 69 (current session) — same gap, but in those rounds Anthropic API limit blocked teammates anyway so issue was masked

Suggested investigation

1. 確認 Claude Code teammates protocol: does shutdown_request actually require teammate to send shutdown_response{approve:true} or just respond at all?
2. If yes → 為什麼 idle teammates didn't respond? (Are they actually receiving the request but not processing? Bug in agent loop?)
3. Mechanical workaround: add timeout-based force-cleanup option to TeamDelete? Or rely on session-end cleanup?
4. SKILL.md fix: add cleanup task to Step 0 Bootstrap regardless of underlying protocol limitations (best-effort + document fallback)

[kiki830621]

Verify (via PR #73)

Result: PASS — #70 structurally resolved (no TeamCreate → no TeamDelete failure surface in idd-verify).
Full report: #73 (comment)

This issue's findings: see "#70" section in the linked report. After PR #73 merges, recommend /idd-close #70 citing #52 Plan implementation as resolution.

[kiki830621]

Closing Summary (retroactive — auto-closed via Resolves trailer in PR #73 squash commit)

IDD discipline note: this issue was auto-closed by the Resolves #70 trailer in commit c31b097 (squashed merge of PR #73). This violated IDD's "no auto-close trailers" rule (CLAUDE.md). Posting retroactive Closing Summary to restore audit trail. Will not reopen → re-close (that's noise per IDD retroactive protocol).

Problem

/idd-verify Step 0 Bootstrap TaskList 沒有 explicit cleanup task confirming team teardown after verify completes。實際 session 觀察（this session's verify-pr58 / verify-pr67 / verify-pr69 / verify-pr71 cycles）：

• 5 reviewer teammates 全部 idle + 寫完 findings file ✓
• Codex CLI 完成 + master comment posted ✓
• TeamDelete 嘗試 → FAIL with: Cannot cleanup team with 5 active member(s)
• SendMessage shutdown_request → teammates ack but don't send shutdown_response{approve:true}
• Subsequent TeamDelete 重試持續失敗 → team 卡 active 殘留至 session end

Impact: workflow friction（next TeamCreate fails with "Already leading team"）+ resource leak（5 active teammates idling per verify run）+ SOP gap（其他 team-using skills 可能複製此 pattern）。

Root Cause

兩層：

1. Coordination protocol issue：Claude Code teammates 在 idle 狀態下不主動 shutdown_response{approve:true} despite receiving shutdown_request。可能是 agent loop bug 或 protocol mismatch。
2. idd-verify SKILL.md Step 0 缺 cleanup_team task：即使 protocol 正常，也沒有 explicit step 來嘗試 + log + fallback。

Solution — structural resolution via #52

/idd-close #52 (PR #73, merged as c31b097) structurally dissolves this bug surface：

idd-verify Step 2 從 TeamCreate (5 teammates) 重構成 5 standalone Agent(subagent_type=general-purpose) calls。

關鍵差異：

Aspect	TeamCreate (old)	5 Agent calls (new, #52)
Team lifecycle	Long-lived team, needs explicit TeamDelete	Each Agent call returns to coordinator after completion
Cleanup mechanism	TeamDelete (failed per this issue)	N/A — no team exists
Failure surface	TeamDelete fails on idle members	Structurally impossible — no team to clean up
Resource state	5 active teammates persist post-verify	All Agents return + are GC'd by harness

Result: 本 issue 報告的 cleanup gap surface 在 idd-verify 中結構性消失。

Verification

• PR feat: idd-verify orchestration playbook (5 Agents + Recovery Protocol) #73 codex verify 3 rounds (R1 3 P1 → R3 PASS) confirmed Step 2 spawn restructure correctness
• /idd-close #52 confirmed all Plan tier D1-D5 implemented + dispatch summary updated
• Master verify report: feat: idd-verify orchestration playbook (5 Agents + Recovery Protocol) #73 (comment)

Process Gap Caveat: First-real-use validation of the 5-Agent mechanism is deferred to next /idd-verify invocation in fresh session (Anthropic API rate-limit prevented exercising the new mechanism during PR #73 verify). If the new mechanism surfaces new bugs upon first-real-use, those will be filed as standard /idd-issue follow-ups (not regressions of #70 — #70's specific TeamDelete surface is genuinely gone).

Changes

• c31b097 feat: idd-verify orchestration playbook (5 Agents + Recovery Protocol) (feat: idd-verify orchestration playbook (5 Agents + Recovery Protocol) #73) — structurally resolves this issue
• 8ac8206 chore: bump issue-driven-dev v2.58.0 → v2.59.0 (Refs feature: idd-verify orchestration playbook (subagent_type guidance + recovery protocol) #52, resolves bug: /idd-verify 結束沒確認 teams 被關掉 — TeamDelete fails on idle-but-active reviewers + no cleanup SOP #70 structurally)

Files changed (for this issue's resolution):

• plugins/issue-driven-dev/skills/idd-verify/SKILL.md (per feature: idd-verify orchestration playbook (subagent_type guidance + recovery protocol) #52, replaces TeamCreate)
• plugins/issue-driven-dev/.claude-plugin/plugin.json v2.58.0 → v2.59.0
• .claude-plugin/marketplace.json synced

Cross-references

• Structural resolver: feature: idd-verify orchestration playbook (subagent_type guidance + recovery protocol) #52 (idd-verify orchestration playbook)
• Closing summary for resolver: feature: idd-verify orchestration playbook (subagent_type guidance + recovery protocol) #52 (comment)
• Master verify report: feat: idd-verify orchestration playbook (5 Agents + Recovery Protocol) #73 (comment)
• Originating session evidence: this session's verify-pr58/67/69/71 cycle 2026-05-10~12 (where TeamDelete repeatedly failed)

Closing Follow-ups Filed (v2.45.0+ #527)

(none — Step 3.5 keyword scan trigger phrases all bound to cross-link mentions of resolver #52; no orphan deferred work)

IDD discipline note (for future)

The Resolves #70 trailer in commit c31b097's body violated CLAUDE.md "No auto-close trailers" rule. Per IDD retroactive protocol, this Closing Summary serves as audit trail repair. Future commits should use Refs #70 — resolves structurally (free-form descriptive text, NOT a GitHub-recognized trailer) instead of Resolves #70 (which IS a recognized auto-close trailer).