Human defines the problem, AI solves it.
TDD 寫測試。SDD 寫規格。IDD 解 bug。 前兩個是手段,IDD 是目的。 完整論述見
MANIFESTO.md。
A Claude Code plugin that enforces issue-driven development as a complete methodology:
- Every change starts with an issue — the single source of truth
- Every issue is diagnosed before implementation — no guessing
- Every implementation is scope-controlled — no creep
- Every completion is independently verified — no "looks good enough"
- Every closure is documented — knowledge preserved
Claude Code is the canonical runtime for IDD skills. The shared skills/ tree is the canonical
source of workflow semantics; non-Claude runtimes are compatibility consumers of those same skills.
Other runtimes, including Codex, can load or interpret the shared skills when a platform shell or adapter is available, but shared skill loading does not imply full runtime parity. If a runtime cannot preserve a Claude-defined workflow gate, it must mark the fallback as degraded instead of silently weakening the IDD contract.
Runtime mapping references:
references/claude-code-tools.md— native Claude Code tool contractsreferences/codex-tools.md— Codex compatibility mapping and degraded fallback rules
The project does not maintain a skills-codex/ fork. Forked skill trees are a last resort and need
a documented incompatibility that cannot be handled through references, adapters, or plugin shell
metadata.
Each skill guards against a specific failure mode:
| Failure | Without this plugin | With this plugin |
|---|---|---|
| No documentation | Changes with no recorded reason | Every change traces to an issue |
| Surface-level fixes | Patch symptoms, root cause returns | Diagnosis required before implementation |
| Scope creep | Fix #42, refactor 3 unrelated files | Scope guardian flags unrelated changes |
| False confidence | "Should work" → ship broken code | Independent AI verification (Codex) |
| Lost knowledge | "What did we do?" 3 months later | Mandatory closing comment |
idd-issue → idd-diagnose → idd-implement → idd-verify → idd-close
① ② ③ ④ ⑤
上圖是最常見的 P-atomic 主幹。完整 36 條 sanctioned path 的選擇導航(何時走 Plan / Spectra / cluster / chain / batch…)見 wiki Path Map(流程圖 + 分類表;source =
docs/workflows.md,generator 防 drift,#276)。
| Skill | Purpose |
|---|---|
idd-issue |
Create well-documented GitHub Issue with original quotes and images |
idd-diagnose |
Find root cause (bug) or analyze requirements (feature/refactor) |
idd-plan |
Approval gate before TDD — presents the Implementation Plan via EnterPlanMode and waits for your OK; for a code issue that changes behaviour, idd-implement then proposes the light spec change (#480; approving the proposal itself in plan mode is #497). Runs after idd-diagnose, never instead of it: you reach it when the diagnosis routes the issue to Plan tier (the middle ground between Simple direct-implement and the full Spectra spec path). Invoking it on an undiagnosed issue aborts — run /idd-diagnose #N first (v2.36.0+) |
idd-implement |
Scope-disciplined implementation with TDD; leaves a spec change for every issue that changes behaviour (light Spectra propose → apply → archive on the branch), or declares no behaviour change (#480). It cannot write a spec change from a git linked worktree, where Spectra writes into the main tree: it stops there; use the main tree or a standalone clone |
idd-verify |
Independent verification — 5 Claude reviewer agents + cross-model Codex leg (gpt-5.x; executable from pai, governance from codex-pro per #264). --profile code/prose/academic + --file/--dir for non-code deliverables (#258, v2.97+) |
idd-close |
Closing comment documenting problem, root cause, solution, verification; refuses while the declared spec change is not archived on the default branch (#480) |
idd-clarify |
Terminology / Semantic accuracy axis (v2.72.0+, #135) — surfacing-only primitive that scans issue body for terminology mismatches / ambiguity / missing-context gaps, annotates via ### Clarity Surface block. Standalone runnable for retroactive audit; delegated by idd-issue Step 4.6; gated by idd-diagnose Step 0.5. Third axis alongside IC_R010 Confidence + IC_R007 Verbatim. |
idd-comment / idd-edit |
Add or amend issue comments with template guidance (decision / note / question / … / reply — reply is the human-facing point-by-point type, v2.100.0+) |
idd-list / idd-update / idd-report |
List open issues by phase, sync issue body, generate progress reports (idd-report --rollup = cross-issue human triage view over dashboard comments, #134 v2.97+) |
idd-find |
Semantic lookup (v2.97+, #139) — surfacing-only search over the open+closed corpus (GitHub relevance + phase/PR overlay); answers "have we solved something like X, and where" |
idd-discuss |
Capture selected human/AI conversations as append-only Discussion snapshots: provenance, stable IDs, retry reconciliation, and the existing privacy gate. Draft by default; explicit publication only. |
idd-ask |
Grounded QA over issues, PR evidence and Discussions (including comments/replies). Combined top-N ≤10, exact source links, explicit partial coverage; --corpus issues retains the original retrieval scope. Workflow state is evidence to interpret, not automatic truth. |
idd-reorganize |
Re-baseline (#200) — when an upstream artifact (issue framing / diagnosis / spec) turns out wrong, list every downstream artifact built on it and rule each one still valid / redo / void, leaving a re-baseline record |
idd-config |
Manage .claude/issue-driven-dev.local.json lifecycle: show / init / validate / which (v2.31.0) |
idd-all |
Orchestrator that drives the full pipeline (issue → close) end-to-end (v2.26.0; v2.28.0 unattended SDD chain; merges or pushes and runs /idd-close after a clean verify, #392) |
idd-all-chain |
Chain-solve mode (v2.55.0+ single-root, v2.60.0+ multi-root + DFS/BFS) — root issue(s) + auto-emergent spawned issues (sister bug / verify follow-up / mid-plan tangential / sister concern) through ONE cluster branch + ONE review PR. Thin recursive shell over idd-all using --in-chain flag (4th mode tuple (direct-commit, unattended)). Multi-root invocation #A #B #C with --bfs opt-in (default DFS). Hard caps: per-root depth=3, global max-issues=10 (v2.60.0+, was 2/5). Eligibility: same-file OR same-skill OR sister-bug. Verify FAIL = per-root halt (other root subtrees continue). STOPs at verified — never auto-close |
Not sure which skill / flag to use for your situation? See references/usecase-routing.md — a 24-row reference mapping common scenarios (single / batch / cluster-PR / external-agent / Plan / Spectra) to the exact skill chain + flags + contract docs.
When implement is delegated to an external agent (Codex, Copilot Workspace, remote claw on a sibling machine), idd-verify supports three input sources beyond the default local-diff:
idd-verify #98 --pr 123— verify a PR opened by the external agent. PR is the master comment location; ref'd issues get pointer comments. Issue↔PR correspondence is gate-checked (PR body mustRefs #Nmatching scope, else abort).idd-verify #98 --commits 3— when the external agent commits to your current working treeidd-verify #98 --branch <name>— when changes live on a branch but no PR yetidd-verify --pr 123(no issue) — auto-discover ref'd issues from PR body
Auto-detect: invoking idd-verify #98 with no input flag counts unpushed commits ref'ing #98 and queries open PRs ref'ing #98, then AskUserQuestion to pick between local diff vs PR. Catches the common forgotten-flag case. Full contract: references/external-agent-delegation.md.
Seven skills accept multiple #NNN arguments and dispatch to one of two modes:
- Batch mode (
idd-diagnose/idd-update/idd-comment/idd-edit) — independent per-issue ops looped sequentially. Each issue gets its own comment, its own auto-update phase, its own audit trail. Failure on one doesn't roll back the rest. - Cluster-PR mode (
idd-implement/idd-verify/idd-close) — multi-issue work sharing one feature branch + one PR. Branchidd/cluster-{slug}, commits tagRefs #N(multiple OK), verify report partitions per issue, close writes per-issue summary. Designed around the "7 issues → 2 themed PRs" workflow pattern.
Selector syntax in v1: explicit list (#34 #36 #38). idd-issue / idd-list / idd-report / idd-config / idd-all are out of scope. Single-issue invocation is unchanged. Full contract: references/batch-and-cluster.md.
Any IDD skill that posts @xxx to GitHub follows a mandatory 5-step protocol so the wrong person never gets pinged:
- Detect intent —
--mention <login>[,<login>...]flag (onidd-issue/idd-comment) or natural-language ("tag X" / "ping X" / "通知 X") - Fetch real list —
gh api repos/$REPO/collaborators(+ org members for org repos); training-data / chat-history / git-log handles are forbidden - Resolve — fuzzy match against
login+namefield; unique match → use, otherwise fallback - AskUserQuestion fallback — 0 or 2+ matches → menu populated from the real collaborator list, never guessed
- Verify pre-post — grep
@\w+from body, every token must be in the verified set, otherwise abort
GitHub mentions are an irreversible side effect; the rule is mandatory not advisory. See rules/tagging-collaborators.md (in-plugin) for the full protocol with examples.
When spectra-discuss is interrupted mid-flow to invoke an IDD skill (e.g. "let me capture this finding to the issue"), the bridge protocol preserves and resumes context:
- Step 0.7 Detect —
--resume-spectra="<topic>"flag,--sourcecontainsspectra-discuss, or.claude/.idd/state/bridge.jsonexists withactive_spectra_session: true(legacy fallback:.claude/state/idd-bridge.json) →SPECTRA_BRIDGE_ACTIVE=1. Open Spectra changes in the repository are not a signal (#502) - Step N-1 Bookmark —
.claude/.idd/state/bridge.jsonwritten with verbatimspectra_topic+issue_url+open_questions[]+idd_action_url+next_step_hint - Step N Resume Prompt — final output prints a clearly-delimited
↩ Resume spectra-discussblock with a copy-pasteable/spectra-discuss <topic>...prompt
Hard rules: never auto-invoke /spectra-discuss (user controls pacing); never paraphrase spectra_topic; resume prompt is the actual recovery — bookmark file is convenience.
idd-comment is the first skill with full implementation. idd-issue and idd-edit will gain it next; the rule defines the contract for all skills. See rules/spectra-bridge.md (in-plugin) for the full schema.
idd-implement accepts two flags that turn it from a single-purpose TDD loop into a dispatcher for other skills:
--with-skill <name>— GREEN phase callsSkill(skill=<name>)instead of directEdit(e.g.,--with-skill perspective-writerfor prose deliverables,--with-skill spectra-applyfor SDD-warranted)--extra '<text>'— free-text additional constraint, written into Implementation Plan's### Extra Requirementssection so checklist semantic gate can verify it
Auto-detection: idd-implement Step 1.5 also scans the diagnosis Strategy for 透過 X / via X patterns and resolves the with_skill target without requiring explicit flags. Compose order: explicit flag > diagnosis hint > none.
idd-implement now explicitly resolves whether work flows through a PR path (feature branch + push + gh pr create) or a direct-commit path (current branch, no PR), instead of implicitly following whatever branch the user happens to be on.
Resolution priority (highest first):
--pr/--no-prflag (per-invocation)- Fork detection (
gh repo view --json isFork) → forced PR path (forks have no upstream push permission) pr_policyconfig field:"always"/"never"/"ask"(default"ask")
idd-close adds a Step 1.5 PR Gate Check that refuses to close an issue when its PR is unmerged. idd-all (orchestrator) explicitly enforces --pr.
Full contract in references/pr-flow.md (in-plugin).
For monorepos and coordinated cross-repo issues, every IDD skill accepts --target owner/repo (or --target group:<label>) so a single workspace can drive issues across multiple GitHub repos:
- Fork-aware (v2.21.0) —
idd-issueresolves the upstream repo from the fork'sorigin - JSON config (v2.22.0, breaking) — per-repo settings move from
.local.mdto.local.json - Six-mechanism resolution (v2.25.0) — flag →
ask_each_timemenu → predicates → cascading walk-up → git remote fallback → orthogonal groups; supportscandidates[]withwhenpredicates andgroups[]for primary + tracking issue pairs
See references/config-protocol.md (in-plugin) for the full algorithm.
| Version | Date | Changes |
|---|---|---|
| v4.2.0 | 2026-11-16 | spec-change.py links 也讀 submodule 裡的標記(#548)— 標記可以在已取出的 submodule 的已追蹤檔案裡(含巢狀),讀的是主 repo 記錄的 commit。找不到標記、又有 submodule 搜尋不了時回 error,點名那個 Example 與最多五個路徑,不再回 missing。只在符合條件的 submodule 目錄執行 git(條件列在 CHANGELOG.md),執行時關掉 lazy fetch、拿掉六個繼承的 git 位置變數(GIT_DIR、GIT_WORK_TREE 等);搜尋 submodule 需要 git 2.45 以上。開了 submodule.recurse、標記又只在沒有工作樹的 submodule 時,現在回 error,取出後才回 ok。詳見 CHANGELOG.md。 |
| v4.1.0 | 2026-11-15 | 刻意沒有 remote 的程式 repo 可以用本機 branch 結案(#547)— IDD 設定新增 spec_archive 欄位:寫在 issue 所在 repo 的設定檔,以程式 repo 的路徑為 key,只在程式 repo 沒有任何 remote 時接受。idd-close Step 1.57 改讀設定的本機 branch、Step 1.5 略過 branch 查詢(open PR 照樣在 issue repo 查),spec-wrapper 第 5 步的連結檢查從這張 issue 的 baseline tag 或第一個 commit 的父 commit 開始;新增 spec-change.py archive-ref、link-base。沒有 remote 又沒有條目的 repo 改成停下、不 fetch;spec-change.py 不再去掉 top level 結尾的空白。另含 #508、#512、#502 的修正,詳見 CHANGELOG.md。 |
| v4.0.1 | 2026-11-11 | #480 的 verify 留下的四項說明與檢查缺口(#499)— idd-plan 用固定的 bash 區塊找這張 issue 開著的 change(issue 號不是數字時停下),spec-wrapper-contract 把區塊抽出來執行,並與契約第 3 步的查詢逐行比對;Aborted 列與「跳過 Step 2」的說法改成與契約一致;idd-implement Step 2 不再改寫在 plan mode 核准過的 Implementation Plan,change 的 tasks 另貼一則 ## Spec tasks 留言,選留言時只算有寫入權的帳號貼的。 |
| v4.0.0 | 2026-11-11 | 每一張實作的 issue 都留下規格改動或宣告過的豁免(#480,對流程是 BREAKING)— Simple 與 Plan 的 issue 由 idd-implement 包成輕量形式的 Spectra change(propose、apply,在 branch 上、verify 之前封存);attended 的 Plan tier 由 idd-plan 核准 Implementation Plan 後交給 idd-implement;Spectra tier 由 idd-all 照同一份契約收尾(references/spec-wrapper.md)。Implementation Complete 帶一行 **Spec**:(change、no behaviour change 或 third-party clone);新增或改到的 requirement 各附帶測試標記的 Example,由 scripts/spec-change.py 檢查;idd-verify 拿規格對照行為;idd-close Step 1.57 在宣告的 change 還沒封存進預設 branch 時不結案。限制:linked worktree 與沒有 Spectra 的 repo 會停下。 |
| v3.7.0 | 2026-11-10 | /idd-all verify 乾淨就自己合併並結案(#392,BREAKING)— PR mode squash-merge、direct-commit 在預設 branch 上 push,再跑 /idd-close,回報 Done:/Undo:;條件由 scripts/close-stage.py 在同一個 process 裡判定(意圖確認過、沒有別人插手、verify gate 乾淨、CI 通過、合併時不會自動關掉 issue);SKILL 新增〈權限〉一節,開 PR 前由 pr-check 判定;--review 停在 verified。討論中持續更新 issue(#424)— 新規則 rules/issue-write-back.md:你在 session 裡定案的改變寫回該在的地方(問題 → 正文段落、診斷 → Diagnosis 段落、新資訊 → 新 comment);BREAKING:/idd-edit --correct 在有寫入權的 repo 不再確認。 |
| v3.6.0 | 2026-11-03 | 本機 CI 閘門(#465)— idd-all 在 verify 之後跑 scripts/idd-ci-gate.sh;沒設 ci.command 的 repo 會停下來,要沿用遠端 CI 設 ci.mode: remote。 |
| v3.5.1 | 2026-11-03 | Squash merge 不再被 merge-completeness gate 誤報(#399)— idd-close Step 1.55 改問「分支的內容有沒有落地」:從 merge-base 往前,取最早一個包含分支淨改動的 commit,也就是分支落地的地方,再對它逐 commit 檢查分支 head 仍保留的效果。main 的 tip 已經有 head 的內容就直接回 0。兩步都用雙向 diff 逐行比對位置(檔案依路徑讀,不追蹤搬移),不做合併:所以 main 自己在別處加、刪、改的行,即使就在下一行,也不會誤報或藏住,只有那一行與 commit 的行文字相同時例外;在落地處找到的每一處缺漏,副本數也對得上才報;需要 git 2.30 以上(3.5.0 已是如此)。無法對應到落地處的 commit(整個分支沒落地,或已落地前綴之後的 commit)對 tip 比對新增的行在、刪掉的行不在。分支內被改寫過的行、合併後又改到同一行的收尾 commit 不再誤報(CHANGELOG 列出的幾種會退回 tip 的形狀除外);沒進 main 的 commit 照樣報出,只刪不加的 commit、只落地一部分效果的 commit 也一樣;只有一部分落地時(已落地的前綴在分支 head 之前的 20 個 commit 之內,淨改動為空的分支也一樣找),已落地的那部分不會混進清單。先合併一個 PR、按 Update branch 繼續做、再合併第二個 PR 時,第一個 PR 的 commit 不再因 main 之後改了那幾行而誤報(中間多推了 commit、第一個 PR 只收了一部分,或 main 在 Update branch 前改了那幾行、merge 保留分支這邊,也一樣)。git 記成搬移、且碰到分支檔案的搬移(含分支 merge 裡的搬移)不判斷(exit 4,idd-close 印一行 note 後繼續),除非 tip 在所有相關路徑上已經和 head 相同,或搬移是 main 在落地之後自己做的、而且沒有 commit 要在 tip 檢查。複製、搬走後在舊名放 shim、搬到既有名字、目錄改名照路徑讀,是已知限制,可能漏報或誤報。merge 之後 main 上的 revert 或覆蓋不在範圍內;其他已知限制見 CHANGELOG。 |
| v3.5.0 | 2026-10-20 | /idd-edit --correct 就地更正錯誤內容(#351,Spectra idd-edit-in-place-correction)— 只更正你自己(目前登入的 gh 使用者)寫的 comment 或開的 issue 正文(整則 comment 或一個 --section;issue 正文用 issue:<N> --section),預覽、確認後取代;舊版歸檔在 .claude/.idd/archive/edits/(不進 git、權限 0000、PreToolUse hook 擋 AI 讀取),可用 restore 還原。idd-verify/idd-diagnose/idd-reorganize 只提議 ### Proposed Corrections,不自己改。每種 /idd-edit mode 改由 helper 分步執行(run-start → prepare → apply/discard → finish)。BREAKING:--append/--replace/--prepend-note 拒絕預覽 exit 9、寫入失敗 4、讀回不符 7(原本都是 0),備份移到使用者暫存目錄的 idd-edit-backup/;/idd-comment --type errata 對你自己的 comment 改為委派 /idd-edit --correct 就地更正,不再另貼 errata(更正被拒或寫入失敗時仍會改貼)。另修 2.81.0–3.4.0 的 parse-args 輸出被 eval 當成 shell 指令執行的問題。詳見 CHANGELOG.md。 |
| v3.4.0 | 2026-10-15 | meeting issue 不再被 actionability gate 擋下(#383)— gate 新增 Diagnosis type 輸入(idd_diagnosis_type、--diagnosis-type),type 只取最新受信任 Diagnosis 的 ### Type;meeting 只略過 Complexity 訊號,parking-lot 與 ### Blocking 照評。idd-all/idd-plan/idd-implement/idd-list 依 TYPE=meeting 分派,idd-update 認得 ## Meeting Plan;idd-diagnose 只採信受信任作者在本文宣告的 meeting。詳見 CHANGELOG.md。 |
| v3.3.0 | 2026-10-07 | 既有工作偵測(#366)— scripts/check-existing-work.sh 在 diagnose / implement 之前一次查完所有 issue 是否已有 open PR 或既有分支在處理,回傳 blocked / resume / unknown / clear;idd-all、idd-all-chain、idd-diagnose、idd-implement、idd-close 接上這個判斷,不再在已有 PR 的 issue 上重做一遍。詳見 CHANGELOG.md。 |
| v3.2.0 | 2026-10-07 | idd-discuss + Discussion-aware idd-ask(#331,PR #334)— 把選定的人機對話存成 append-only 的 GitHub Discussion snapshot(來源、穩定 ID、重試對帳);idd-ask 預設納入 Discussions(`--corpus issues |
| v3.1.1 | 2026-10-07 | --blocked-by 原生依賴修正(#353,PR #358)— 改呼叫 GitHub 實際存在的 addBlockedBy mutation,失敗時原樣印出 GitHub 的錯誤。詳見 CHANGELOG.md。 |
| v3.1.0 | 2026-10-07 | Actionability gate(#316,PR #318,Spectra add-actionability-gate,七輪 verify)— 共用 helper scripts/lib/actionability.sh:### Complexity 以開頭 tier 字取值、延期語彙另給 reason;### Blocking 逐 bullet 讀;idd-list / idd-all / idd-implement / idd-plan 在任何 branch 或 comment 之前呼叫 gate 並印出兩段式判定;idd-list 新增 Parked 與 Needs diagnosis 組。已知與 markdown-it / CommonMark 的分歧(D1–D12)與環境依賴(locale、平台 C library、awk)以實測方向列出並釘住;結構規則凍結,交 #336 決定。詳見 CHANGELOG.md。 |
| v2.102.2 | 2026-08-01 | Deep Research 輕整合(#277,裁決 (b),PR #283)— idd-diagnose 非綁定 pointer(#111 形狀)+ 回填慣例(note 摘要+連結)+ corpus/外部世界邊界表 + 兩個實際使用者誤解的 canonical 澄清。詳見 CHANGELOG.md。 |
| v2.102.1 | 2026-08-01 | Reopen / resume path(#278,PR #282)— closed 之後事情還要做的合法回程:判準表 + 接點規則 + summary append-only 三步;usecase-routing 情境 31。詳見 CHANGELOG.md。 |
| v2.102.0 | 2026-07-27 | 三線發版:skill description contract + Path Map(#276,PR #279/#280 — wiki 導航圖 36 路徑全覆蓋 + generator 防 drift);egress 資料安全 cluster(#275 empty-body guard exit 15、#273 edit-comment verb — surgery 通道入網 + batch band 消費,PR #281)。42 suites 0 fail。詳見 CHANGELOG.md。 |
| v2.99.1 | 2026-07-18 | Staleness sweep + guard 擴網(#267,PR #268)— README 3 根 stale pin + docs catalog 回填至 v2.99;model-generation-sync 網擴及 README/docs、新 docs-catalog-sync suite(#122 根因 test-detectable)。38 suites 0 fail。詳見 CHANGELOG.md。 |
| v2.99.0 | 2026-07-18 | /idd-ask — issue 知識庫 grounded 問答(#72,PR #266)— surfacing family 第 4 員,鏡像 /spectra-ask:decide-to-search gate(bug 貌問題不觸發 diagnose)→ retrieval delegate idd-find backend → top-N 全文(5/上限 10)→ grounded 合成(blockquote 原問題、claim 必附引用、source priority、分歧 surface、### Referenced Issues、查無誠實)。#140 增員程序首戰(Q3 弱命中 judgment 入 family canonical)。新 capability spec idd-ask;37 suites 0 fail。詳見 CHANGELOG.md。 |
| v2.98.0 | 2026-07-18 | codex 通道完全依賴化(#264,PR #265)— vendored bin/codex-call 刪除(落後 pai 四個修正,切換即淨升級);executable 歸 pai(MIN_PAI 2.19.0,codexModel/codexEffort 契約)、治理歸 codex-pro(MIN_CODEX_PRO 0.7.0,defaults.json + profile.yaml 兩層);IDD 樹內零 model pin,換代只動 codex-pro 一處;superpowers 形狀依賴接線(fail-fast、無 soft fallback)。上游先行:codex-pro 0.7.0 + pai 2.19.0。詳見 CHANGELOG.md。 |
| v2.97.0 | 2026-07-17 | 9-issue drain:5 cluster PR(#259–#263)。Added: verify profiles(#258)— --profile code/prose/academic + --file/--dir 輸入源 + file SHA-256 freshness gate(#228 非 git 等價物)、config verify_profiles 自訂;/idd-find(#139)— open+closed 全語料 relevance 查找,surfacing-only family 第 3 員;dashboard comment 契約(#133)+ idd-report --rollup(#134)— 人類敘事快照(只綁 phase 轉換,anti-#116)與四分組 attention 視圖(pull-only);sdd_bias 開關(#252)— 硬閘出口 high 升 Spectra、default 逐 byte 不變;Layer V deferred-record(#120);surfacing-primitives family 文件(D12 軸)(#140)。Changed: model 世代同步(#251)— codex-call default 升 gpt-5.6-sol(live probe 驗證)、全樹世代中立、idd-route candidate 更名 codex-xhigh;docs path catalog 補完(#122)。5 新 drift-guard suites,聚合器 36 suites 0 fail。詳見 CHANGELOG.md。 |
| v2.96.0 | 2026-07-13 | gh-egress 硬化 cluster + idd-edit batch 語意。#227 exit-code 帶 ≥10(wrapper 絕不自行 exit <10 — $? 即可分流 gate-refusal vs gh-failure);#225 python3 統一 content-net 掃描(消滅 jq/no-jq 分歧;sensitive-key taxonomy;fail-closed 寬網);#226 Phase 2 rollout(6 skill 的 comment/edit egress 全接線 gh-egress — #117 mention net 機械生效於 comment channel);#158 idd-edit batch × R5 per-comment refuse + 繼續(outcome report + ✗ REFUSED 顯著行,任一 refused → exit 4)。(#227 #225 #226 PR #256;#158 PR #257)詳見 CHANGELOG.md。 |
| v2.95.0 | 2026-07-11 | Discussions intake bridge(#221,PR #255,Spectra discussions-intake-bridge)— idd-list --discussions(opt-in surface:probe→Q&A/Ideas+未答 filter→dedup→actionable 區塊)+ idd-issue --from-discussion(Provenance verbatim blockquote + draft-and-confirm 回文,unattended 絕不 post);絕不 auto-file。idd-verify diff-freshness gate(#228,PR #253)— FROZEN_SHA 比對,stale-snapshot aggregate 拒絕 + in-flight 不 commit 紀律。IDD_CALLER registry(#161,PR #254)— 動態樹掃描 drift-guard。詳見 CHANGELOG.md。 |
| v2.94.0 | 2026-07-10 | Selective git auto-tag(#85,PR #250)。Added: idd-issue 打 idd-{N}-baseline(main HEAD = rollback anchor),idd-verify 在 Aggregate PASS 打 idd-{N}-verified(review snapshot,cluster-aware)。只這兩個 milestone(無 diagnose/plan/implement tag),tag namespace 保持乾淨;phase-level tag 刻意留手動。新 config auto_tag(enabled 預設 ON、opt-out via enabled:false、baseline_format/verified_format/push_remote),tag 含 git push(repo-wide side effect,已明文標註);idempotent(tag 已存在→skip)+ graceful-skip on push fail(never aborts)。新增 auto-tag drift-guard(27 assertions)。詳見 CHANGELOG.md。 |
| v2.93.1 | 2026-07-09 | Collaborator identity registry in idd-config(#86,PR #247)。Added: optional collaborators[] config field — 把人的 alias / email / display-name → GitHub @login 不用猜(github_login 必填;email 是 PII,只放 private/gitignored config)。tagging-collaborators.md Step 2.5 先查 registry(accelerator,非 authority — 命中仍經 gh api users/<login> 存在性驗證,未命中 fall through 到 API fuzzy-match);idd-config validate 查 login charset + aliases 全域唯一 + PII 提醒。新增 collaborators-schema drift-guard(12 assertions)。Also bundled: discussion-metadata helper(#141 + #142,PR #248)— idd-issue Step 3.5 surface 討論型 issue 的 assignee / label / milestone(--discussion flag + 預設 off 的 body heuristic;surface-only,unattended 絕不 auto-set;blocked-by deferred);#142 label existence-check 防 blind add。discussion-metadata drift-guard(14 assertions)。詳見 CHANGELOG.md。 |
| v2.93.0 | 2026-07-08 | Reshape Plan / pre-implementation tier — Cluster C(#129 / #57 / #111,reshape-plan-preimpl-tier Spectra change, PR #249)。Added: first-class meeting issue type(meeting-first routing — deterministic type=meeting 在 Layer 1 / Layer V / Spectra / 硬閘 / Layer P 之前分流;Phase A/B/C deliberation Strategy;idd-plan skip-chain;self-contained close gate = decision→action mapping,無 /idd-verify,#57);complexity hard gate(≥5 檔互依概念 OR shared abstraction → MUST-trigger Plan,escalate-only、保留 Simple default,#129)。Changed: superpowers pre-implementation hand-off(README ↔ superpowers 階段對照表 + 非綁定 brainstorming pointer,不自建 staging skill,#111)。3 新 specs 促成(complexity-hard-gate / meeting-issue-type / superpowers-integration);93 drift-guard assertions;5 輪 6-AI ensemble 驗證。詳見 CHANGELOG.md。 |
| v2.92.0 | 2026-07-06 | /idd-all batch-drain release — 23 issues verified+closed via 16 merged PRs(#223, #229–#243),本 plugin 最大規模 self-dogfood。Added: unattended-contract(state-file 訊號 + TTL、TTY heuristic 廢除、idd-all/chain 依賴 early gates,#123/#222/#211);gh-egress unconditional @-mention net(escape-or-attest、--mention-attested,#117)+ 6 項 mechanical-net 精修(#203);idd-close Step 6.3 doc-sync sweep(#220);測試聚合器 + GitHub Actions CI、21 suites(#217);idd-list blocked-state 分組 + banner(#84);config Mechanism 3.5 submodule 路由(#162);enabled-state 偵測(#212);monorepo host 消歧(#68)。Changed: pai 升 install-time dependency + vendored fork 刪除、Tier 鏈 3→2(#219);idd-verify DA sequenced-spawn(#130);--force-linked-issue 分離(#172);worktree 公約統一(#169)。Audits: #210/#215。Follow-ups: #225–#228。詳見 CHANGELOG.md。 |
| v2.73.0 – v2.91.0 | 2026-05-22 – 2026-07-04 | Reason-pattern registry(#137/#150)、worktree isolation(#167)、retroactive close(#176)、conflict-class taxonomy + tracker phase(#179/#182)、tree-lock(#183)、merge-completeness gate(#184)、privacy-scrubbing gate 家族(#202)、agent-model dispatch 契約(#205)、pai canonical backend(#207)、superpowers hard dependency(#209)、rules layering + SessionStart hook(#214)等 — 完整逐版記錄見 CHANGELOG.md。 |
| v2.72.0 | 2026-05-22 | /idd-clarify skill — Terminology / Semantic accuracy quality axis (#135, add-idd-clarify-skill Spectra change). NEW composable primitive surfacing-only that scans existing issue body for terminology mismatches (per references/terminology-canonical.md library) / ambiguity / missing-context gaps; annotates via ### Clarity Surface block with status (surfaced / dismissed / resolved). Standalone runnable for retroactive audit + delegated by idd-issue Step 4.6 (between Step 4.5 Milestone and Step 4.7 Sister Sweep, skip in --multi-finding mode) + gated by idd-diagnose Step 0.5 (hard refuse on unresolved rows, per idd-close PR Gate Check + idd-all-chain #119 fail-fast precedents). Library initial seed 6 rows (K-means 特徵值→分群變數 / PCA-K-means / regression-ANOVA / accuracy-RMSE / Bayesian-frequentist / unsupervised-classification) per #804 incident + 通用 ML/統計 misuse patterns. Rule-of-three promotion threshold for new rows + open PR contribution. Triggering case: QEF #804 K-means「特徵值」(eigenvalue) source term actually meant「分群變數 / distinguishing variable」(老師 LINE clarify);AI 沿用 source verbatim, spec 階段意外修對是運氣不是 systematic safeguard。Recursive evidence: 本 session AI 連 4 次 under-verification (D06 collision #815 / Step 4.8 collision / customer_attributes 來源未指定 / 三輪 design over-abstract) — codify 後 surface 給 user dismiss/resolve 不繼續犯。Sister concerns filed: #136 idd-edit/idd-update symmetry P3 / #137 unattended Clarity Surface contract P2 / #138 ralph-loop → /goal substitution P3. |
| v2.56.0 – v2.71.0 | 2026-05-10 – 2026-05-21 | Multi-finding source mode, chain-solve hardening, auto-close-trap defence, cluster mode override, Simple-cluster docs batch, NSQL conformance, acceptance-review doctrine, NSQL doctrine follow-up family, multi-finding spec hardening from #48 verify, Phase 0.4 detection precision sweep from #53 verify, DA sentinel regex + CRLF normalization, spectra-archive sync + cluster glob + pasted-image cache. v2.56.0: idd-issue multi-finding source mode — auto-trigger on ≥2 paragraph-level findings from docx/pdf/Telegram/Mail/Notes adapters, 4-stage extract→pick→preview→dispatch pipeline. v2.57.0: idd-close Step 6.5 Distribution Sync chain — detection-driven 3-option AskUserQuestion chaining to plugin-update / mcp-deploy / cli-deploy at close moment. v2.58.0: idd-issue Stage 4.5 jsonl gitignore pre-flight gate. v2.59.0: idd-verify orchestration playbook — Step 2 switches TeamCreate → 5 parallel Agent calls + Codex background; NEW Step 2.5 Recovery Protocol. v2.60.0: idd-all-chain multi-root + DFS/BFS traversal + per-root halt + spawn-manifest schema v2 (#46). v2.60.1: cluster fix — PR-body auto-close trap (#87 + #74); 4 PR-body templates reworded digit-free + NEW idd-verify Step 0.8 preventive gate. v2.61.0: idd-verify Step 0.8 squash-commit-body auto-close trap fix (#97) — 2-source scan (PR body closingIssuesReferences + per-commit headline/body trap regex) + ### 引用 trap pattern 作反例的寫作紀律 writing discipline in CLAUDE.md. v2.62.0: cluster mode override (#96) — NEW pr-flow.md ### Cluster mode override (cluster mode = idd-implement path-resolution precondition, forces PR; verify/close are cluster-aware but consume the path) + idd-implement Step 0.5 bash cluster detection with explicit override notice mirroring fork detection; resolves a 3-file doc contradiction. v2.63.0: #96-backlog Simple cluster — 6 docs/reference follow-ups shipped via cluster-PR #101 (#60 #62 #63 #78 #90 #91); Simple-tier subset of an 18-issue /idd-diagnose batch over the #96-backlog cleanup. v2.64.0: IDD human-in-the-loop reconciled to the NSQL confirmation protocol (#103, PR #104). NSQL (kiki830621/NSQL v4.1.0) is the Human-AI Confirmation Protocol registered as a reference project in CLAUDE.md. F1 — idd-issue Step 5 echoes the AI-rendered ## Type / ## Expected / ## Actual + plain-language interpretation (NSQL run → report, not a confirm gate; creating an issue is reversible per v4.1.0). F2 — idd-diagnose Layer V clarify now renders candidate interpretations for the user to pick (NSQL P1 Read-Only for Humans), free-text named as fallback. F3 — Diagnosis report template gains ### Residue section + explanatory paragraph (NSQL §4.6: non-operationalizable intent, distinct from Layer V vagueness; (none) empty-state required). F4 — Layer P risk-sensitive boundary adds irreversible side effects to its enumerated vocabulary (catches Telegram-send / Stripe-charge / webhook-fire / file-delete cases). Plan tier: Spectra → /spectra-discuss re-evaluated under v4.1.0 traceability gate, re-routed Plan via EnterPlanMode. 6-AI verify PASS (0 blocking, in-PR polish merged). 3 follow-ups filed: #105 idd-close Residue acknowledgement, #106 plugin.json description chain via plugin-update, #107 idd-issue Step 5 echo CI/loop warning. v2.65.0: IDD acceptance-review reconciled as a MANIFESTO doctrine + thin --review flag, NOT a new mode-tuple axis (#102, PR #109). /spectra-discuss killed Direction A (3-tuple (path, interaction, acceptance)) and most of Direction B (merge-time gate); re-routed Spectra → Plan. NEW MANIFESTO section ## Human-in-the-loop: IDD 即 NSQL Confirmation Protocol formalises that IDD's human-in-the-loop is an NSQL Confirmation Protocol instance — human confirmation loop closes BEFORE execution (at issue + idd-diagnose); idd-verify is execution-fidelity, NOT a confirmation loop; verify-gated is the named, sanctioned terminal default disposition (one clean 6/6 verify PASS is sufficient). Verify-as-review reframe: 5 specialized adversarial agents + Codex on correctness exceed a single human merge reviewer's thoroughness; "AI verify PASS = no review" is a backwards read. --review is opt-in to re-open the confirmation loop (NOT a quality gate; per-invocation flag, NOT a standing config field). auto-merge legitimate under verify-gated PASS, justified by "verify is the gate" (not "merges are reversible"); guardrails mandatory; auto-merge ≠ auto-close; autopilot mechanics belong to #37; doctrine explicitly forbids /loop / external CI callers from interpreting Phase 6 verify-gated PASS as gh pr merge authorization. idd-all + idd-all-chain Phase 6/4 terminal messaging dispatches on $REVIEW_FLAG: default Verify: verify-gated PASS + Next: merge, then /idd-close; --review Verify: verify-gated PASS — awaiting human acceptance (re-opened confirmation loop per --review). Cluster PR body checklist conditional. --review orthogonal to --pr/--no-pr/--in-chain/--bfs/--cwd, messaging-only at orchestrator scope. 6-AI verify on PR #109 caught a HIGH blocking bug: idd-all-chain PR-body checklist initially used ${REVIEW_FLAG:+A}${REVIEW_FLAG:-B} thinking it was a mutex — it is NOT (${var:-word} returns $var when set, not the alternative branch). 3 of 6 reviewers convergent (regression / DA / Codex); 3 missed it including logic which claimed bash-correct without empirically testing. Fixed in-PR (dc61ffb) with explicit if/else building $REVIEW_CHECKLIST_LINE before the heredoc. The catch itself dogfoods the doctrine being shipped: the adversarial + cross-model ensemble is the falsifiable check. Sister consistency family follow-up #108 extended to 5-template + 3 satellite items (PR-body wording sync across idd-implement Step 5.5 + idd-all Phase 5 + pr-flow.md + chain-flow.md + Phase 4 stdout + Trace 1 example + DA3 wording). v2.66.0: NSQL doctrine follow-up family from #102 / #103 close (#105, #107, #108, PR #110). #105 — idd-close Step 3.6 Residue Acknowledgement closes the ### Residue write-only loop from v2.64.0 (#103 PR #104 DA finding D2 "latent capacity to drift into ritual filler with no consumer pressure"). NEW step reads latest Diagnosis comment's ### Residue section, silent-skips on (none) / missing / pre-v2.64.0; non-empty fires 3-option AskUserQuestion (acknowledge / file follow-up / skip); audit-trail PATCH appends ### Residue Acknowledgement to in-memory closing summary draft BEFORE Step 4 publish; SHOULD-tier non-blocking; IDD_LOW_BAR_RESIDUE_ACK=false rollback. #107 — idd-issue Step 5 NEW ⚠ paragraph warns CI / /loop callers about the v2.64.0 echo expansion (metadata-only → rendered ## Type / ## Expected / ## Actual + plain-language interpretation); near-verbatim per issue body with skill-internal #107 self-reference + precedent parenthetical. #108 — 5-template + 3-satellite consistency family closed. 4 PR-body templates synced to Verify-gated default wording across idd-implement Step 5.5:503 + idd-all Phase 5:755 + pr-flow.md:135 + chain-flow.md:254 (cluster variant); section heading ## Pending review → ## Review status. F3 satellite — idd-all-chain Phase 4 stdout dispatches via explicit if/else before heredoc (avoids ${VAR:-word} mutex from PR #109 F1). Trace 1 example refresh with parallel --review variant. DA3 wording precision messaging-only → orchestrator-scope messaging-only in 3 sites (idd-all + idd-all-chain Phase 0 + MANIFESTO --review paragraph). Per Option A: idd-implement does NOT accept --review. Live spec openspec/specs/idd-all-chain/spec.md:161 updated to match emission. 6-AI verify ensemble PASS post-fix (regression H1 spec drift + L1 off-by-2 line-number + L3 heading dissonance + Codex CHANGELOG accuracy nit all addressed in fix commit 6bae8e6). Master verify: PR #110 #issuecomment-4494037985. Squash-merged as cb19802. v2.67.0: multi-finding spec hardening family from #48's 6-AI verify (#75 #76 #77 #79 #80, PR #113). 5 sister bugs all same-file (skills/idd-issue/SKILL.md), shipped as one chain. #75 security — NEW ### Content sanitization contract (dual-track jsonl-verbatim-per-IC_R007 + GitHub-body-sanitized via Python Unicode-aware filter covering C0/DEL/C1 + bidi-override U+202A-U+202E + U+2066-U+2069 Trojan-Source CVE-2021-42574; sanitize_source_label() refuse-on-@token cross-references rules/tagging-collaborators.md; jq --arg/--argjson mandate; CAUTION banner above schema). #76 bug — run_id second→ms-precision via dispatch chain handling macOS BSD date hazard (GNU %3N → Python datetime fallback → .000Z); TOCTOU symlink check fail-closed; JSONL_WRITE_GUARD 15-bit nonce-retry invoked at run-start (not materialize) to lock canonical path before Stage 4 footer composition. Pre-v2.67.0 second-precision was the irreversible-side-effect failure mode added to Layer P vocabulary in v2.64.0 #103 F4. #77 enhancement — 7 corner-case spec contract gaps (flag-conflict layering, partner_eligible_set formal definition, Stage 3 Edit-row soft cap >5, [Back to top-3] in Stage 2 Other, Stage 1 path canonicalization, agent-crash known gap, Stage 4.5 unattended fallback). #79 enhancement — abort-path writes minimal aborted: true jsonl with partial timestamps (footer URLs in already-dispatched bodies remain valid); footer > **Action** line; "srt" source_type enum. #80 enhancement — Stage 1 anchor heuristics for MAY merge/split clauses (default preserve / split if ≥3 distinct topics / merge if <200 chars same-topic); max_possible_score formula explicit; N<3 degenerate-case picker shape table. 5/6-AI verify PASS post-fix caught 4 HIGH blocking dispatched in-PR per feedback_verify_fix_same_pr (F75-1 CJK UTF-8 byte-level corruption / F76-1 BSD date %3N silent literal / F76-2 JSONL_WRITE_GUARD dead code + RANDOM 15-bit drift / #79 abort spec/impl contradiction). Codex 6th reviewer hung at 12+ min — process gap recorded per v2.59.0+ convention. Squash-merged as 14bc930. v2.68.0: Phase 0.4 diagnosis-detection precision sweep from #53's verify follow-up family (#59 #64 #65, PR #114). 3 sister fixes. #59 — idd-all 2 Python substring sites (line 450 + 533) swapped to line-anchored re.search(r'(?m)^## Diagnosis', c['body']) matching check-diagnosis-readiness.sh canonical convention from #53 / PR #58. Scope narrowed during implementation: cited idd-list / idd-update sites are narrative prose; idd-close uses startswith() (already line-1-anchored). #64 — helper script regex widened from ^## Diagnosis to ^[ ]{0,3}## Diagnosis for CommonMark 1-3 space ATX heading indent. Col-0 still matches; 4+ space / tab / Unicode whitespace stay excluded (verified by DA's 31-fixture battery). #65 Approach A — NEW comment block documenting line-based detection's fenced-code false-positive limitation. Approach B (full markdown parser) deferred per feedback_lead_minimal. Mitigation: chain Phase 0.4 AskUserQuestion user override. 3-AI verify PASS (lean ensemble for 108-line diff). 0 blocking; 3 FYI observations. Not in scope: #61 test fixture infra — Plan-tier with framework-choice surface deferred. Squash-merged as a15a6a5. v2.71.0: /idd-all-chain Phase 0.4.5 fail-fast cap-exceeded preflight + anti-pattern A3 cite (#119, commits cb96579 + 7aa11ea). Pre-v2.71 silent-truncate ("filed only, not chained") behavior — when user passed N>cap roots, Phase 2 loop processed first cap then broke; remaining roots silently dropped — replaced with fail-fast refuse at Phase 0.4.5 before any cluster branch / manifest creation. Refuse message cites docs/workflows.md Anti-pattern A3 (P-chain-from-root 多 root 用 batch 跑) + suggests batch /idd-diagnose #N #M ... + (P-atomic / P-cluster-pr) per cluster. Existing Phase 2 silent-truncate warnings (line 368, 468) get (per docs/workflows.md anti-pattern A3) cite suffix for defense-in-depth (Phase 2 fires on spawn-driven overflow, orthogonal to Phase 0.4.5 input-cap). Configuration cap doc reframed "global cap" → "hard cap as ripple-chain guardrail" — cap is ripple subtree safety bound NOT batch knob; multi-root batch should use /idd-all #N #M ... --pr cluster-pr path. Failure mode table adds new row distinguishing Phase 0 root-count refuse from Phase 2 spawn-driven truncate. Ship dogfood: Round 1 verify (cb96579) FAIL — 3 independent reviewers (Logic + Regression + Codex) + DA reinforcement caught Phase 0.4.5 was placebo gate: F1 (${ROOTS[@]} undefined; Phase 0.1 uses ROOT_ISSUES_SORTED / N_ROOTS) + F2 ($CHAIN_MAX_ISSUES set at line 377 Phase 1, hoist needed for line 267 Phase 0.4.5) + F3 (exit 1 vs abort convention). Fix commit 7aa11ea resolves all 3; Round 2 verify 5 independent voices + 1 coordinator self-review (DA Agent crashed 2x with API socket closed — process gap surfaced #130) = 6 PASS, 0 BLOCKING. Spectra→Plan tier reframing per docs/workflows.md path-catalog approach (commit 92e82c2): skill behavior change collapsed from architectural Spectra (cap config-driven / new semantic detection) to Plan minimal MVP (fail-fast + cite A3 + doc-level cover). Companion docs introduce docs/workflows.md (8 sanctioned path categories + 6 anti-patterns) + docs/skill-dimensions.md (D1-D11 design axes) for IDD plugin architecture navigation. 6 IDD design issues migrated from psychquant-claude-plugins marketplace to architectural home (commit e827bc7): #89→#122 (strict-separation proposal, supersede by reframing), #69→#123, #29→#124, #27→#125, #33→#126, #7→#127. Follow-ups filed: #130 DA Agent socket crash, #131 cognitive-load proxy metric (residue), #132 Phase 2 truncate backward-compat policy (residue, likely moot). |
| v2.55.0 | 2026-05-10 | /idd-all-chain skill — chain-solve mode (#44, add-idd-all-chain-skill Spectra change). NEW thin recursive shell over /idd-all that drives root issue + auto-emergent spawn (sub-skill sister sweep / verify follow-up / mid-plan tangential / sister concern) through ONE cluster branch (idd/chain-<N>-<slug>) + ONE review PR (title prefix chain:, collapsed <details> per issue). NEW --in-chain flag on /idd-all derives 4th mode tuple (direct-commit, unattended) — sub-/idd-all skips Phase 0.5 PR-mode branch + Phase 5.5 PR open + sub-skill UNATTENDED MODE; mutex with --pr/--no-pr. NEW spawn manifest cross-skill contract at .claude/.idd/state/chain-spawned-issues.json (schema_version=1, atomic temp-file rename); 4 sub-skills (idd-implement / idd-verify / idd-plan / idd-diagnose) all conformantly write entries with spawn_kind + same_file_as_root + same_skill_as_root. Helper scripts/manifest-append.sh. Hard caps: depth=2, max-issues=5; over-cap spawns still file as follow-up issues but NOT enqueued. Chain-eligible heuristic: same_file OR same_skill OR sister-bug; ineligible spawns stay filed-only. Failure mode: chained verify FAIL halts queue + preserves partial commits (no rebase / revert) + abort report cites 4 recovery paths. Stops at verified — never auto-close, never auto-merge (per IDD discipline). MODIFIED capability idd-orchestrator-modes adds 4th tuple (direct-commit, unattended) for chain context; existing 3 tuples unchanged. NEW reference docs: references/spawn-manifest.md (schema canonical contract) + references/chain-flow.md (chain shell algorithm canonical contract). Backward compat: /idd-all #N (no flag) is byte-equivalent to v2.53.0 baseline. |
| v2.53.0 | 2026-05-07 | ralph-loop dependency declaration + runtime detect (#28, add-ralph-loop-dependency Spectra change). README ### Required for specific modes matrix declares ralph-loop optional dependency for idd-verify --loop. NEW scripts/check-ralph-loop.sh detector (version-glob path lookup with HOME-spaces tolerance via nullglob array; escape hatch: IDD_SKIP_RALPH_LOOP_CHECK=1). NEW idd-verify --loop Step 0a fail-fast gate (abort if ralph-loop missing, point user to install). NEW idd-all Phase 0.6 graceful-degrade gate (when --loop style is requested but ralph-loop missing, unwind feature branch + fall back to single-shot verify with audit trail; Trace 4 example added). Forward-compat plugins[].dependencies field experimentally added then removed per F1 verify finding (schema validate-fail). Address F1-F5 verify findings before merge. |
| v2.52.0 | 2026-05-05 | idd-issue ordered/unordered bundle flags (#21). NEW --parent <N> flag PATCHes parent body task list with new child entry, idempotent via #N reference scan + fallback ## Children anchor when no list exists. NEW --blocked-by <M>[,<M2>...] flag applies three-layer fallback chain: Layer 1 GraphQL addBlockedByDependency mutation attempt (graceful failure → warning + continue, no abort) + Layer 2 unconditional body blockquote > Blocked by #M (always readable in any markdown viewer) + Layer 3 parent task list annotation (blocked by #M) when --parent co-used. NEW --bundle-mode <ordered|unordered> flag orchestrates bundle creation in single invocation: builds 1 epic parent + N children with auto-applied --parent <epic>, ordered mode adds strict child[i] blocked by child[i-1] chain, unordered keeps task list only. Pre-flight gates: cross-repo refuse (parent in different repo → abort + redirect to groups), bundle-mode and group-mode mutual exclusion. Step 3.B inserted between 3.A (single repo) and 3.G (group), reusing 3.A flow. Orthogonal with Step 4.5 milestone, Step 4.7 sister sweep. NEW canonical reference doc references/bundle-flags.md. NEW ## Ordered Bundle Pattern section after Step 5 in idd-issue SKILL.md (3-mode comparison + 3 usage scenarios + design rationale for not creating separate /idd-bundle skill). NEW capability spec idd-issue-bundle (4 Requirements, frozen). Spectra change add-bundle-flags-to-idd-issue archived. No breaking changes — all flags additive. |
| v2.51.0 | 2026-05-04 | idd-list shows open PR info per issue + cluster detection (#13). NEW Step 2.5 batch fetches all open PRs once (gh pr list --state open --limit 100); NEW Step 3.5 client-side regex #(\d+)\b scans PR bodies and builds reverse issue→PR index plus cluster map (PRs ref'ing 2+ issues). Step 4 Format Output adds └─ PR #N (status, mergeable) sub-line per issue with PR ref; cluster leaders (lowest issue number) show — cluster: #X #Y #Z listing all members; cluster members show → see PR #N (cluster member) redirect. Direct-commit issues (no PR refs) display unchanged from v2.50 — fully backward compatible. Footer adds second line summarizing N issues bundled in M cluster(s); P solo PR(s); Q direct-commit. Step 5 Suggest Next extended to phase × PR state matrix (10+ rows): implemented + draft → gh pr ready N → /idd-verify --pr N; implemented + ready MERGEABLE → /idd-verify --pr N; verified + ready MERGEABLE → gh pr review N → gh pr merge N → /idd-close #N; verified + merged catch-up → /idd-close #N; CONFLICTING → gh pr checkout N → resolve; cluster member → see leader's next action. Sister concerns filed as P3 follow-ups: #14 (markdown-aware PR body parser) + #15 (cluster_leader config option). |
| v2.50.0 | 2026-05-04 | Layer V Vagueness Pre-check for complexity routing (#12). NEW Step 3.4 in idd-diagnose between Layer 1 disqualifier and Layer 2 Spectra evaluation. AI scores V1 (vague WHAT) + V4 (vague ACCEPTANCE) on Likert 6-point scale (no neutral midpoint), trigger threshold per-axis ≥ 4. Triggered cases fire Hybrid 3-option AskUserQuestion (clarify now / proceed anyway / escalate to Plan) with default option score-driven (V=4 → proceed, V=5 → clarify, V=6 → escalate). escalate force-sets verdict = Plan via Layer V and skips Layer 2/3/P. Routing parsers in idd-implement Step 2.5 + idd-all Phase 3 strip via X suffix to extract canonical tier (backward compat: bare verdicts unchanged). NEW project rule .claude/rules/attribute-assessment.md codifies meta-principle "attribute scoring SHALL use Likert scale, not keyword matching" — applies session-wide via root CLAUDE.md @import, scope beyond Layer V. MANIFESTO.md 5-axis bug-fix model expanded to 6-axis adding "Alignment quality" (TDD ❌ / SDD ❌ / IDD ✅), evidence = Layer V. idd-all unattended mode auto-applies proceed anyway + audit trail (same pattern as Plan tier under unattended). Backward compat: pre-v2.50 diagnoses NOT retroactively re-evaluated; existing Simple / Plan / Spectra / SDD-warranted verdicts remain valid. No --ignore-vagueness flag (option B proceed anyway covers it). Spectra change add-vagueness-layer-routing archived. New main spec openspec/specs/routing-vagueness-layer/spec.md (9 normative requirements with scenarios + example tables). |
| v2.36.0 – v2.49.0 | 2026-04-30 – 2026-05-03 | Spectra rename + Plan tier + cross-skill IC_R011 alignment. v2.36.0: 3-tier Complexity routing (Simple / Plan / Spectra; SDD-warranted alias preserved); NEW idd-plan skill with EnterPlanMode approval gate. v2.37.0: External-agent / PR mode for idd-verify (--pr <N> / --commits N / --branch <name> flags + auto-detect); references/external-agent-delegation.md + references/usecase-routing.md. v2.38.0: idd-route integration (data-driven agent routing recommendation; binary at PsychQuant/idd-route-swift); marketplace migration to PsychQuant/issue-driven-development. v2.39.0 / v2.40.0: --cwd flag for cross-repo orchestration (sub-skills inherit). v2.41.0: idd-close Step 0 supersession (Implementation Complete > Checklist treated as canonical when all [x]). v2.42.0 – v2.48.0: IC_R011 systematic alignment — canonical references/ic-r011-checkpoint.md + 5 sub-issues across idd-plan / idd-implement / idd-close / idd-diagnose / idd-issue (each gets a 3-option AskUserQuestion + audit trail at its deliberation moment). v2.49.0: IC_R011 Third-Party Skill Alignment for /spectra-discuss + /spectra-propose. |
| v2.34.0 | 2026-04-29 | Multi-issue invocation across 7 skills via two distinct modes. Batch mode (idd-diagnose / idd-update / idd-comment / idd-edit) — sequential per-issue ops, each issue gets its own comment / auto-update phase / audit trail; failure on one doesn't stop the rest; pure idempotent. Cluster-PR mode (idd-implement / idd-verify / idd-close) — multi-issue work shares one feature branch + one PR; commits tag Refs #N (multiple OK); branch named idd/cluster-{slug}; verify report partitions per-issue; close enforces per-issue checklist gate + writes per-issue summary (no batched fake summary). Designed around the "7 issues → 2 themed PRs" workflow pattern (e.g., 04/27 Docs+Sanitizer split). Selector syntax in v1: explicit list (#34 #36 #38) only; --label / --milestone selectors deferred. Out of scope: idd-issue (input phase), idd-list / idd-report (already multi by design), idd-config (N/A), idd-all (single-issue orchestrator). Single-issue invocation unchanged — backward compatible. New canonical contract: references/batch-and-cluster.md. |
| v2.33.0 | 2026-04-28 | NEW MANIFESTO.md — formalizes the methodology thesis. Decomposes "bug-solving capability" into 5 sub-capabilities (TDD covers 1.5/5, SDD 0/5, IDD 5/5). Adds two formal claims: (1) IDD opens a second axis orthogonal to TDD/SDD's verification axis (closure / DONE definition); (2) IDD ⊋ TDD ∪ SDD on falsifiability surface. Empirical case study: che-word-mcp #56 cluster — 30 findings, 5 sub-stack rounds, 6 patch releases, zero zombies. README + CLAUDE.md link to MANIFESTO. No code change. |
| v2.32.0 | 2026-04-28 | TWO new rules closing real-world workflow gaps observed in PsychQuant/contact-book#96. (1) NEW rules/tagging-collaborators.md — mandatory 5-step protocol when any IDD skill posts @-mention to GitHub: detect intent (--mention <login> flag or natural-language) → fetch real list (gh api repos/$REPO/collaborators + org members; training-memory / chat-history / git-log handles forbidden) → fuzzy match → AskUserQuestion fallback for 0/2+ matches → grep + verify pre-post (abort on unverified token). idd-issue + idd-comment gain --mention <login>[,<login>...] flag; idd-diagnose / idd-implement / idd-verify / idd-close enforce via prose detection. (2) NEW rules/spectra-bridge.md — bridge contract for IDD skills called mid-spectra-discuss: detection signals trigger SPECTRA_BRIDGE_ACTIVE, bookmark schema preserves verbatim spectra_topic + issue_url + open_questions[] + next_step_hint, final ↩ Resume spectra-discuss prompt block printed at skill exit. idd-comment is first skill with full implementation (Step 0.7 detect, Step 7 bookmark + resume); idd-issue / idd-edit will gain it next. No breaking changes — new flags opt-in. |
| v2.31.0 | 2026-04-27 | NEW idd-config skill — independent entry for .claude/issue-driven-dev.local.json lifecycle. Four subcommands: show (default; resolved target + cwd-aware predicate trace), init (interactive first-time setup, equivalent to idd-issue Step 0.5.E fork-aware detection without forcing an issue creation), validate (JSON schema + repo existence + predicate-key sanity), which (dry-run resolution at cwd, optional --title / --label to evaluate content predicates). Closes the gap where setup / inspection / monorepo predicate debugging was only available as a side effect of idd-issue. |
| v2.30.0 | 2026-04-26 | (1) idd-issue 資料保留鐵律 — all source attachments uploaded to attachments release by default without asking. New Source Type Adapter table covers .docx / .pdf / Telegram / Apple Mail / Notes / pasted text + Telegram fallback flow. Step 4 renamed 附加圖片(如果有) → 附加所有原始素材(鐵律:預設全保留) with violation checklist. (2) idd-implement --with-skill <name> + --extra '<text>' flags + new Step 1.5 Resolve Extra Requirements; GREEN phase calls Skill(skill=…) instead of Edit when with_skill set. First-class formalization of idd-implement × perspective-writer integration. |
| v2.29.0 | 2026-04-26 | Two-tier checklist gate in idd-close — Step 0 structural gate (existing) refuses close on unticked - [ ]; new Step 1.6 semantic gate does keyword extraction on each - [x] and verifies test/spec/file mentions correspond to real artifacts. Warn-only with three-way AskUserQuestion (proceed / investigate / edit). |
| v2.28.0 | 2026-04-26 | idd-all SDD path is now an unattended chain — spectra-discuss → spectra-propose → spectra-apply with explicit unattended hints in args. Orchestrator overrides sub-skill attended-by-default contracts via args, not by modifying sub-skills. |
| v2.27.0 | 2026-04-26 | PR vs direct-commit path routing in idd-implement (--pr / --no-pr flag, fork-aware default, new pr_policy config field). idd-close Step 1.5 PR Gate Check refuses close on unmerged PR. idd-all explicitly enforces PR path. New references/pr-flow.md as canonical contract. |
| v2.26.0 | 2026-04-26 | Add idd-all orchestrator skill that drives the full pipeline (issue → diagnose → implement → verify → close) end-to-end. |
| v2.25.0 | 2026-04-26 | Monorepo + multi-repo support via config-protocol — six-mechanism resolution. New candidates[] (path/git predicates), groups[] (primary + tracking with bidirectional cross-link comments), ask_each_time. |
| v2.22.x | 2026-04-22 | JSON config (breaking — .local.md → .local.json); fork-aware target repo selection; codex pinning. |
| v2.18.0 – v2.20.0 | 2026-04-14 – 2026-04-16 | Mandatory Step 0 Bootstrap Stage Task List for every IDD stage skill; idd-verify auto-triages follow-up findings into new issues. |
| v2.12.0 – v2.17.x | 2026-04-07 – 2026-04-14 | SDD as special case of IDD; checklist gate on close; idd-list / idd-comment / idd-edit skills; ban Closes/Fixes/Resolves trailers (they bypass idd-close gate). |
# Install (v2.38.0+ marketplace)
/plugin marketplace add https://github.com/PsychQuant/issue-driven-development
/plugin install issue-driven-dev@issue-driven-development
# Use (auto-completion: type "idd-" to see all skills)
/idd-issue "upload button doesn't work on mobile"
/idd-diagnose #42
/idd-implement #42
/idd-verify #42
/idd-close #42On first use, creates .claude/issue-driven-dev.local.md:
---
github_repo: "owner/repo"
github_owner: "owner"
attachments_release: "attachments"
---GitHub Discussions(Q&A / Ideas)是 Issues 之外的真實 intake channel — 本 plugin 現在能看到它:idd-list --discussions(opt-in)surface actionable 的 Discussions(未答 answerChosenAt: null、未被任何 issue 引用、Discussions 停用時優雅 no-op),idd-issue --from-discussion <url|number> 把一則 Discussion seed 成 issue(## Provenance + 原文 verbatim blockquote),建案後的 back-reference 回文 draft-and-confirm(unattended 絕不 auto-post)。鐵律:絕不 auto-file — bridge 只 surface,人來判斷。契約 + GraphQL 樣板:references/discussions-intake.md。
IDD tags two milestones so you can jump back to a stable checkpoint with git checkout instead of hunting commit hashes: idd-{N}-baseline (placed by idd-issue on main's HEAD when the issue opens — the rollback anchor) and idd-{N}-verified (placed by idd-verify on the Aggregate-PASS snapshot — review-ready). Only these two points are tagged (no diagnose/plan/implement tags), so the tag namespace stays clean.
Default-ON, opt-out via config. Tagging includes a git push (a repo-wide side effect), is idempotent (existing tag → skip), and graceful-skips on any push failure (never aborts the workflow). Turn it off with:
{ "auto_tag": { "enabled": false } }Full schema (enabled / baseline_format / verified_format / push_remote) in references/config-protocol.md → auto_tag field.
IDD 遵循「深度整合 >> hard-coded」原則(#209/#214):生態系已有 canonical 套件時依賴它、不在內部複製等價邏輯。每個整合的綁定形狀與缺席行為:
| Package | 綁定形狀 | 角色 | 缺席行為 |
|---|---|---|---|
parallel-ai-agents(psychquant-claude-plugins) |
Hard dependency — plugin.json dependencies 安裝時自動安裝(#219;cross-marketplace,root marketplace 需 allowCrossMarketplaceDependenciesOn 增列) |
idd-verify 的 canonical ensemble 引擎(4 lens + DA + Codex;≥2.18.0 STABLE contract) |
缺席/過舊 → 印一步安裝指令 + fall back manual fan-out(品質等價、較慢);vendored fork 已刪(#207 Residue 成熟) |
superpowers(claude-plugins-official) |
Hard dependency — plugin.json dependencies 安裝時自動安裝(#209) |
TDD 執行(idd-implement)、系統性除錯(idd-diagnose bug RCA)、完成前驗證的 canonical process 紀律 |
Delegation 點 fail-fast abort + 一步安裝指令;絕不 silent degrade |
parallel-ai-agents(pai) |
Canonical engine + 版本閘門 ≥ 2.19.0(#207/#264,STABLE external-consumer contract 含 codexModel/codexEffort) | idd-verify 的 6-AI ensemble 引擎(4 lenses + Devil's Advocate + Codex 編排)+ bin/codex-call executable(#264 起 IDD 不再 vendor) |
兩層 degrade:canonical → manual fan-out(每層印 notice;vendored fork 已刪 #219/#264) |
ralph-loop(claude-plugins-official) |
Optional mode 依賴(#28) | idd-verify --loop / idd-all (PR, unattended) 的 verify-fix loop driver |
--loop fail-fast + 安裝指引;idd-all graceful degrade 到 (direct-commit, attended) |
codex-pro(codex-pro marketplace) |
Hard dependency(#264)— codex 治理 single source(defaults.json + profile.yaml 契約,MIN 0.7.0) |
Verify 跨模型盲驗 lens 的 model/effort/max-time 治理(gpt-5.x,世代 pin 只在 codex-pro 一處) | 缺席 fail-fast abort + 一步安裝指令;絕不靜默 fallback 舊 model |
che-word-mcp / che-telegram-mcp / che-apple-mail-mcp / che-apple-notes-mcp |
Optional per source type(#27) | idd-issue 來源 adapter(docx / Telegram / Mail / Notes) |
該 source type fail-fast + 結構化安裝指引(見下方 Optional 表) |
ghCLI authenticated with GitHub- OpenAI Codex CLI installed (for
idd-verify) - ChatGPT Pro account (for the Codex cross-model lens, gpt-5.x — model governed by codex-pro's defaults.json)
superpowersplugin(claude-plugins-official)— hard dependency(v2.90.0+ #209)。安裝本 plugin 時 Claude Code 會經plugin.jsondependencies自動安裝(需 Claude Code v2.1.110+;遞移 enable 需 v2.1.143+)。idd-implement的 TDD 執行與完成前驗證、idd-diagnose的 bug RCA 執行框架 delegate 給它;缺席時該路徑 fail-fast abort + 一步安裝指令,不做 fallback- SessionStart hook(v2.91.0+ #214):plugin 啟用後每次 SessionStart 事件(startup / resume / clear / compact — 刻意含 compact,讓規則在 context 壓縮後存活)注入 ≤5 行 IDD 紀律:commit issue-reference(canonical:
rules/commit-issue-reference.md)與討論寫回(canonical:rules/issue-write-back.md)— 防 GitHub auto-close trap 於 skill 外的手動 commit - PreToolUse hook
archive-read-guard(v3.5.0+ #351):擋下 Claude 讀取 correction archive(.claude/.idd/archive/)的工具呼叫,/idd-edit --correct歸檔的舊版因此不會被讀回 context。fail-closed:任何指令文字只要提到這個路徑就擋,commit 訊息、issue 內文也一樣——改用檔案傳(git commit -F、gh … --body-file)或不寫出路徑;指向archive/之下的相對路徑(例如cd .claude/.idd && cat archive/…)在歸檔附近、或指名的.claude/.idd確實有歸檔時也擋,判斷以文字為主;在歸檔附近,同一個指令裡換目錄(cd、pushd、Set-Location…)又改權限就擋,在.claude/.idd裡任何改權限都擋(任何cd,連cd /tmp也算;cd scripts && chmod +x run.sh請改寫成chmod +x scripts/run.sh,或把 chmod 拆成另一個呼叫);檢查超過 5 秒即拒絕;遞迴改權限(chmod -R等)只在歸檔附近檢查:工作目錄在含歸檔的 repo 內、其上一層資料夾、或無法列出時,指令同時含權限指令與遞迴標記就擋,不看目標(grep -R chmod .、寫 script 的 heredoc 也擋——分開執行、改用 Write/Grep 工具或不加-R);其他目錄不檢查;已知限制(程式選項換目錄如env -C、root、Windows 等,完整清單見 CHANGELOG 3.5.0 第 26 輪);人要看舊版時在自己的終端機開;要暫時解除,建立~/.cache/idd/archive-read-guard.disabled(見hooks/archive-read-guard.py),用完刪掉。
/idd-issue 支援多種來源類型。文字直接貼進 prompt 不需要任何 plugin;若要從以下來源讀取,需另裝對應 MCP plugin:
| Source type | MCP plugin | Why | Install |
|---|---|---|---|
.docx / .doc |
che-word-mcp |
idd-issue Step 1 用 mcp__che-word-mcp__get_document_text / list_images / export_image 讀文字 + 抽圖 |
claude plugin install che-word-mcp@<your-marketplace> |
| Telegram chat range | che-telegram-mcp (telegram-all server) |
讀 chat history (get_chat_history) |
claude plugin install che-telegram-mcp@<your-marketplace> |
| Apple Mail message | che-apple-mail-mcp |
get_email + list_attachments + save_attachment |
claude plugin install che-apple-mail-mcp@<your-marketplace> |
| Apple Notes | che-apple-notes-mcp |
get_note + 抽 inline 圖 |
claude plugin install che-apple-notes-mcp@<your-marketplace> |
替換 <your-marketplace> 為實際的 marketplace 名稱。沒裝對應 plugin 不會壞 IDD 核心流程,但該 source type 會被迫 fallback 到「使用者手動處理」模式 (#27 追蹤改為明確 fail-fast 報錯)。
不知道
<your-marketplace>該填什麼?
查已加的 marketplaces:
claude plugin marketplace list加新 marketplace:
claude plugin marketplace add <owner>/<repo>常見對照(以 PsychQuant 維護的 plugins 為例,如有變動以
claude plugin marketplace list為準):
Plugin pattern Marketplace Add command che-*(incl.che-word-mcp/che-telegram-mcp/che-apple-mail-mcp/che-apple-notes-mcp)psychquant-claude-pluginsclaude plugin marketplace add PsychQuant/psychquant-claude-pluginsidd-*(本 plugin /idd-route)issue-driven-developmentclaude plugin marketplace add PsychQuant/issue-driven-developmentralph-loop(Anthropic 官方)claude-plugins-officialclaude plugin marketplace add anthropics/claude-plugins-official
某些 mode / flag 需要額外 plugin。未裝對應 plugin 時的行為見最右欄:
| Mode | Required plugin | Marketplace | Why | Behavior if missing |
|---|---|---|---|---|
idd-implement(TDD / 完成前驗證)+ idd-diagnose(bug RCA) |
superpowers |
claude-plugins-official |
process-discipline delegation(#209 hard dependency;正常情況由 dependencies 自動安裝,不會缺席) |
fail-fast abort + install hint |
idd-verify --loop |
ralph-loop |
claude-plugins-official |
outer driver for verify-fix loop | fail-fast abort + install hint |
idd-all (PR, unattended) (default) |
ralph-loop |
同上 | unattended 在 verify findings 後需 driver 觸發下一輪 | graceful degrade to (direct-commit, attended) + warning(保護 v2.40.0 caller backward compat) |
idd-all --no-pr (direct-commit, attended) |
(none) | n/a | user 在 keyboard 自然推進 | n/a |
Single-skill calls (no --loop) |
(none) | n/a | atomic skill 自完成 | n/a |
Install:claude plugin marketplace add anthropics/claude-plugins-official 然後 claude plugin install ralph-loop@claude-plugins-official。
Native alternative —
/goal(v2.1.139+,#138):以上「outer driver for verify-fix loop」的角色,也可改用 Claude Code 內建的/goal—— 設一個 completion condition(例如「verify 通過、blocking findings = 0」),它每個 turn 後用 small fast model 檢查、未達成就自動再跑下一個 turn,直到條件滿足。/goal是 native(無需安裝第三方 plugin),ralph-loop是claude-plugins-official的 plugin —— 兩者都能驅動 loop,擇一即可。目前idd-all/idd-verify --loop的 gate 偵測的是ralph-loopplugin;若偏好 native,可手動以/goal驅動(是否把 gate 改為支援/goal為更大設計題,見 #138 residue)。
IDD pipeline stage 對照 superpowers counterpart。pre-implementation staging(brainstorm → written plan)委派 superpowers,IDD 不自建對等 skill(依 deep-integration-over-hardcode:superpowers 既是 hard dependency,hand-off pointer 必然可解析,自建 = vendored fork)。verify ensemble 與 close audit trail 標為 IDD 獨有 — 無 superpowers 對應。
| IDD stage | superpowers counterpart | 說明 |
|---|---|---|
idd-issue(問題框定) |
(無 — issue framing 是 IDD 入口) | well-documented issue |
idd-diagnose(bug RCA) |
superpowers:systematic-debugging |
RCA 執行框架(delegate,#209) |
| pre-implementation staging(brainstorm) | superpowers:brainstorming |
design-heavy issue 的非綁定 hand-off pointer(idd-issue summary step + idd-diagnose 印出;使用者自選是否跟進,IDD 自身不 invoke) |
| written plan | superpowers:writing-plans |
同上 staging;IDD 的 idd-plan 是 approval-gate 版、不 delegate |
idd-implement(TDD) |
superpowers:test-driven-development |
TDD 執行紀律(delegate,#209) |
idd-verify |
IDD 獨有 — 無 superpowers 對應 | 6-AI ensemble cross-verify |
idd-close |
IDD 獨有 — 無 superpowers 對應 | closing summary audit trail + two-tier gate |
hand-off pointer ≠ delegation(#111 契約):上表的
superpowers:brainstorming/writing-plans是 hand-off pointer,不是 delegation。界線:
- delegation(IDD flow 內部呼叫 superpowers skill):僅
idd-implementTDD 與idd-diagnoseRCA 的 process-discipline 走 delegation(#209)。planning discipline(idd-plan/ Spectra,含對應的brainstorming/writing-plans)禁止 delegation。- hand-off pointer(idd-issue / idd-diagnose 印給使用者的非綁定建議,允許):IDD 自身不 invoke 該 skill,使用者自選是否跟進。
機械檢查(drift-guard 硬鎖,見
scripts/tests/superpowers-staging/):grep -rn 'superpowers:' plugins/issue-driven-dev/skills/idd-plan/零命中 ——idd-plan內不得出現任何 superpowers 呼叫;pointer 只活在 idd-issue / idd-diagnose 的 prose 建議裡。
idd-route— Data-driven agent routing (Codex / Claude Opus / Sonnet / Haiku) per IDD issue。idd-diagnoseStep 3.7 偵測到idd-route已裝就會自動呼叫做 enrichment;沒裝則 silently skip。非必要,純粹增強 routing 建議品質。Install:claude plugin install idd-route@issue-driven-development
scripts/check-plugin-presence.sh (#34) 用「filesystem 存在」作為 plugin 安裝判準 — 檢查 ~/.claude/plugins/cache/<marketplace>/<plugin>/<version>/.claude-plugin/plugin.json 即視為 installed,無 signature / hash / integrity 驗證。
| Environment | Trust model fit |
|---|---|
| Single-user macOS / Linux local dev | ✅ 設計預設場景,trust model 適用 |
Shared dev container / CI runner with shared $HOME |
plugin.json 騙過 detect → IDD 跑 --loop / source-type adapter as if plugin present。Caller 自負其責 |
| Multi-user shared workstation |
Hardening path (若部署到 untrusted 環境):
- Pre-validate plugin install via
claude plugin list(source-of-truth from Claude Code itself, not filesystem) - 設
IDD_SKIP_PLUGIN_CHECK=1env var 跳過 detect,接受 bypass cost (--loop跑了但無真實 driver = silent fail) - 部署時 lock
~/.claude/plugins/cache/permissions (chmod 700) 防 hostile write
Future hardening (#41 reopen criteria): 若 multi-user / CI deployment 真實成本上升,evaluate hash verification step against marketplace-published plugin manifest。
Use /idd-discuss to draft a record of a selected conversation topic. Explicitly request publication
to create a Discussion, or continuation to append a new source batch. Initial bodies and existing
comments are never overwritten; each later snapshot states the current understanding and cites its
selected original messages. Unknown authors/models/times stay unknown.
The capture contract documents source JSON, stable topic/source IDs, local serialization, remote reconciliation and conservative handling of uncertain writes. Different state directories or devices must be serialized; this is not a global exactly-once service. The repository must enable Discussions and an actual category must be selected before publication. The helper does not enable Discussions or publish test messages automatically.
/idd-ask --corpus all searches both issue and Discussion evidence; issues and discussions select
a single corpus. Replies have exact citation URLs, and API failures or bounded reads are disclosed.
This improves traceability, not the truth of every AI statement. Runtime helpers use Python and gh. The shared egress security gate additionally requires the
pinned Markdown parser; install it with:
python3 -m pip install -r plugins/issue-driven-dev/scripts/requirements-egress.txtUse the same Python executable as the helper. Missing/incompatible parser versions refuse all egress. The skill YAML contract test also uses PyYAML 6.0.2; CI installs both dependencies.