@kaydash9999/banker-plugins 0.26.1 → 0.27.0

package/skills/ultra-init/SKILL.md

@@ -38,7 +38,7 @@ So the effective order is `ralplan → ralph → QA gate`. This is intentional,
 <Loop_Authority_Rules>
 Several bundled skills are loop/persistence engines. To prevent competing loops, **only ONE loop authority is active at a time, and the phases run strictly sequentially** (a phase fully completes before the next begins):
 
-1. In **Phase 2, ralph is the sole loop authority** ("the boulder never stops"). Let its loop run to completion. Do **NOT** arm a Claude Code `/goal` Stop-hook loop beside it.
+1. In **Phase 2, ralph is the sole loop authority** ("the boulder never stops" in Claude Code; on Codex banker's `ralph` Stop hook keeps the turn going). Let its loop run to completion. Do **NOT** arm a Claude Code `/goal` Stop-hook loop beside it.
 2. **The QA gate runs only in Phase 3, AFTER ralph has finished**, as the terminal QA sub-gate — never concurrently with ralph. banker's `ralph-qa` owns that judge → fix loop on both runtimes; ultra-init owns no loop.
 3. **There is no ultrawork phase.** OMC 5.0.0 retired `ultrawork`; ralph's own parallel delegation covers it inside Phase 2.
 </Loop_Authority_Rules>
@@ -53,19 +53,20 @@ Several bundled skills are loop/persistence engines. To prevent competing loops,
   omc ralph verify --help 2>&1 | grep -q 'Usage: omc ralph verify' && echo "omc ok" || echo "omc CLI missing or too old"
   ```
   If the check fails, **guide installation or update first** via `/banker:setup` → oh-my-claudecode (or the `setup-omc` skill / `npm install -g oh-my-claude-sisyphus@latest`), then resume once it passes. There is no way around a missing or stale CLI: ralph itself stops without `omc ralph verify`, so do not start Phase 1.
-- **런타임 매핑 (Claude Code ↔ Codex).** `ralplan`·`ralph`·`cancel` 은 Claude Code에선 OMC, **Codex에선 oh-my-codex(OMX)가 제공하는 동명 스킬**이다(Codex는 `omx setup` 전제, 점검은 `omx --version`). OMX 의 `ralph` 는 `omc` CLI 를 쓰지 않으므로 위 `omc` 점검은 Claude Code 에만 해당한다. Phase 3 QA 게이트는 두 런타임 모두 banker `ralph-qa` 다: Claude Code `Skill("banker:ralph-qa")`, Codex npm 길 `$banker-ralph-qa`, 플러그인 길 `$banker:ralph-qa`. Codex 에는 Skill 도구가 없으니 그 `SKILL.md` 를 읽고 그대로 따른다(npm 길 `<S>/../banker-ralph-qa/SKILL.md`, 플러그인 길 `<S>/../ralph-qa/SKILL.md`, `<S>` 는 이 스킬 폴더). `<RQ>` 는 ralph-qa 의 `references` 폴더다(Claude Code 와 Codex 플러그인 길 `<S>/../ralph-qa/references`, Codex npm 길 `<S>/../banker-ralph-qa/references`). 아래 Phase들의 `Skill("oh-my-claudecode:<name>")` 표기는 Claude Code 기준 — **Codex 런타임에선 OMX의 해당 스킬을 적용**한다.
-- **Codex 의 tmux 밖 실행.** OMX 의 런타임 워크플로(`ralph`, `team` 등)는 OMX tmux 런타임을 요구한다. OMX 의 `AGENTS.md` 는 tmux 밖(Codex App, 일반 `codex`, `codex exec`)이면 그 사실을 알리고 가장 가까운 대체 경로로 계속하라고 정한다(oh-my-codex 0.20.2 `templates/AGENTS.md`). 그래서 tmux 밖이라는 사실만으로는 멈추지 않는다. 그 단계는 이 세션이 그 스킬의 `SKILL.md` 절차를 네이티브 하위 에이전트로 그대로 따르고, 보고에 `<단계>: tmux 밖 대체 실행(OMX 런타임 없음)` 한 줄을 적는다. 종결 낱말은 다른 멈춤 조건과 마지막 ralph-qa 판정으로만 정한다. 대체 실행도 할 수 없으면(하위 에이전트 기능이 꺼짐 등) `STOPPED(omx-runtime-unavailable)` 로 멈추고 tmux 안의 OMX 에서 다시 실행하라고 안내한다.
-- **Codex 의 계획과 상태 경로.** 이 문서의 `.omc/plans/`, `.omc/state/sessions/<id>/prd.json` 은 Claude Code(OMC) 경로다. Codex 에서는 OMX 경로를 쓴다: 계획은 `.omx/plans/`(ralplan 이 쓰는 `prd-*.md`, `test-spec-*.md`), ralph 의 진행 기록은 `.omx/state/<scope>/ralph-progress.json`. Codex 에서 `.omc/` 아래에 계획을 쓰지 않는다.
+- **런타임 매핑 (Claude Code ↔ Codex).** `ralplan`·`cancel` 은 Claude Code에선 OMC, **Codex에선 oh-my-codex(OMX)가 제공하는 동명 스킬**이다(Codex는 `omx setup` 전제, 점검은 `omx --version`). `ralph` 는 Claude Code에선 OMC, **Codex에선 banker 의 Codex 전용 `ralph`** 다(npm 길 `$banker-ralph`, 플러그인 길 `$banker:ralph`, `SKILL.md` 는 npm 길 `<S>/../banker-ralph/SKILL.md`, 플러그인 길 `<S>/../../codex/skills/ralph/SKILL.md`). OMX 0.21 이 `$ralph` 를 없앴으므로 OMX 의 `$ralph` 는 부르지 않는다. banker `ralph` 는 OMX 와 `omc` CLI 를 쓰지 않으므로 위 `omc` 점검은 Claude Code 에만 해당한다. Phase 3 QA 게이트는 두 런타임 모두 banker `ralph-qa` 다: Claude Code `Skill("banker:ralph-qa")`, Codex npm 길 `$banker-ralph-qa`, 플러그인 길 `$banker:ralph-qa`. Codex 에는 Skill 도구가 없으니 그 `SKILL.md` 를 읽고 그대로 따른다(npm 길 `<S>/../banker-ralph-qa/SKILL.md`, 플러그인 길 `<S>/../ralph-qa/SKILL.md`, `<S>` 는 이 스킬 폴더). `<RQ>` 는 ralph-qa 의 `references` 폴더다(Claude Code 와 Codex 플러그인 길 `<S>/../ralph-qa/references`, Codex npm 길 `<S>/../banker-ralph-qa/references`). 아래 Phase들의 `Skill("oh-my-claudecode:<name>")` 표기는 Claude Code 기준 — **Codex 런타임에선 OMX의 해당 스킬을 적용**한다(`ralph` 만 banker `ralph`).
+- **Codex 의 tmux 밖 실행.** banker `ralph` 는 tmux 가 필요 없다. OMX 의 런타임 워크플로(`team` 등)는 OMX tmux 런타임을 요구한다. OMX 의 `AGENTS.md` 는 tmux 밖(Codex App, 일반 `codex`, `codex exec`)이면 그 사실을 알리고 가장 가까운 대체 경로로 계속하라고 정한다(oh-my-codex 0.20.2 `templates/AGENTS.md`). 그래서 tmux 밖이라는 사실만으로는 멈추지 않는다. 그 단계는 이 세션이 그 스킬의 `SKILL.md` 절차를 네이티브 하위 에이전트로 그대로 따르고, 보고에 `<단계>: tmux 밖 대체 실행(OMX 런타임 없음)` 한 줄을 적는다. 종결 낱말은 다른 멈춤 조건과 마지막 ralph-qa 판정으로만 정한다. 대체 실행도 할 수 없으면(하위 에이전트 기능이 꺼짐 등) `STOPPED(omx-runtime-unavailable)` 로 멈추고 tmux 안의 OMX 에서 다시 실행하라고 안내한다.
+- **Codex 의 계획과 상태 경로.** 이 문서의 `.omc/plans/`, `.omc/state/sessions/<id>/prd.json` 은 Claude Code(OMC) 경로다. Codex 에서는 OMX 경로를 쓴다: 계획은 `.omx/plans/`(ralplan 이 쓰는 `prd-*.md`, `test-spec-*.md`), ralph 의 진행 기록은 저장소 밖 게이트 폴더의 `ralph.md`(banker `ralph` 의 `ralph-state.mjs`). Codex 에서 `.omc/` 아래에 계획을 쓰지 않는다.
 
 ### Phase 1 — PLAN: ralplan --deliberate (합의 계획)
 - Invoke `Skill("oh-my-claudecode:ralplan")` with `--deliberate <brief>` in **non-interactive** mode (so it produces the consensus-approved plan, marks it `pending approval`, and **stops without launching team/ralph itself** — ultra-init controls the hand-off).
+- **Codex: plan only.** OMX 0.21's `$ralplan` ends by moving on to `$ultragoal`. Tell it to write the plan and stop there: no `$ultragoal`, `$ralph`, `$team` or implementation. Phase 2 is banker's `ralph`.
 - ralplan runs **Planner → Architect → Critic** to consensus (max 5 iterations), and in deliberate mode adds a **pre-mortem (3 scenarios) + expanded test plan (unit/integration/e2e/observability)**.
 - Capture the final plan and its path in `.omc/plans/`: **ADR + ordered user stories + per-story testable acceptance criteria + the test plan**. This is the seed for Phases 2–3 and the durable record of the run.
 - Autonomy: treat the consensus-approved plan as approved-for-execution and continue. (`--gated` → pause here for user approval, then continue.)
 
 ### Before Phase 2 — record what already fails
 - Make the gate folder with `node '<RQ>/gate-evidence.mjs' init --cwd '<repo>'` (its `dir` is the gate folder `<g>`) and run each targeted QA dimension's command once on the unchanged tree, through ralph-qa's runner, as iteration 0: `node '<RQ>/gate-evidence.mjs' run --dir '<g>' --id <dimension> --cmd '<command>' --timeout <seconds> --iter 0 --cwd '<repo>'`. One command per dimension (no `&&` or `;` chains), a `--timeout` on each; `scenarios` has no baseline (its rows test the change).
-- This **pre-change record** (iteration 0 in `<g>/evidence.md`) is the QA gate's baseline: ralph-qa counts a failure with the same command and failure signature as not this change's (`baseline`). It covers checks outside ralph's feedback commands, and on Codex it is the only baseline (OMX `ralph` keeps none). On Codex the runner writes outside the workspace: ask for approval to run it outside the sandbox.
+- This **pre-change record** (iteration 0 in `<g>/evidence.md`) is the QA gate's baseline: ralph-qa counts a failure with the same command and failure signature as not this change's (`baseline`). It covers checks outside ralph's feedback commands. On Codex banker's `ralph` keeps a baseline of its own in its own gate folder; this one is still the QA gate's. On Codex the runner writes outside the workspace: ask for approval to run it outside the sandbox.
 - Never use `git stash` or `git reset` for this; if the tree already has changes, take the record from a clean copy of HEAD (`git worktree add --detach <tmp> HEAD`, run with `--cwd <tmp>`, then `git worktree remove <tmp>`). A fresh worktree has no installed dependencies (`node_modules`, a venv): link or install them there first, and confirm the commands really ran — a record made only of "command not found" errors is no baseline. The runner writes the folder a command ran in as `<repo>` in the failure signature, so a failure in the copy and the same failure in the repo match.
 - The record lives in `<g>`; if the session may compact or end before Phase 3, put the `<g>` path in the `ready-compact` note. The gate folder lives until Phase 3 has its end word: no cleanup's day rule takes it (ralph-qa's gate folder rule).
 - **Not a git repo** (common for a new project): if `init` (or `run`) stops with `not-a-git-tree` (exit 3, nothing made or run), first check that the folder the stop names (`checked` in its JSON, an absolute path) is the target repo; if not, call it again with `--cwd '<repo>'`. Only when the target really is no git repo: the gates cannot have evidence (the work tree fingerprint sees no change). Take no baseline, and in Phase 3 hand ralph-qa no `--gates`; it then writes `게이트: 선언 없음(git 저장소 아님)` in its report and seat instructions. Never run `git init` yourself: tell the user that `git init` and a first commit make the gates work next time. The same on both runtimes (Claude Code, and Codex on the npm and the plugin route).
@@ -75,7 +76,8 @@ Several bundled skills are loop/persistence engines. To prevent competing loops,
 - **No questions mid-run:** ralph 5.6.1 asks the user once for the PRD's `repoQualityClass` (`prototype`, `production` or `library`) when the task does not say. Infer it from repo signals (CI config, test depth, published package metadata) and put it in the handoff; let ralph ask only when the signals genuinely conflict.
 - ralph first records the feedback baseline (`omc ralph verify --write-baseline`: the failures that already existed), then iterates **story-by-story**, delegating to parallel agents where stories allow, until all stories `passes: true`, followed by **reviewer verification → mandatory ai-slop-cleaner deslop → post-deslop regression re-verify**. ralph is the loop authority for this phase. Note its session id (the `<id>` in its PRD path `.omc/state/sessions/<id>/prd.json`) for a resume, and the baseline failures it reports at its closeout for the Phase 4 summary.
 - **Feedback commands must print only failures.** OMC 5.6.1's `omc ralph verify` fingerprints every output line and normalizes only `<number><unit>` durations, timestamps, `/tmp` paths and hex ids, so a summary line that changes between runs or when tests are added reads as a NEW failure (node:test's default reporter: `ℹ duration_ms 151.4`, `ℹ tests 8` — red on every run with no change). Have ralph set the PRD `feedbackCommands` before it records the baseline to forms that are constant on success and name each failure (node:test: `FORCE_COLOR=0 NODE_NO_WARNINGS=1 node --test --test-reporter=dot 2>&1 | grep -Ev '^[.X]+$'` — the variables keep color codes from slipping past the filter and process-id warnings out of the output; adapt the form on Windows). If the gate still flags only such lines, fix the commands and re-record ralph's baseline on the current tree (`omc ralph verify --write-baseline`); the pre-change record still tells old failures from new ones. Never edit the baseline file by hand; never stash or reset.
-- Do not interrupt on "The boulder never stops" continuations — that signal means ralph's iteration continues. Let it run to completion.
+- **Codex: banker `ralph`.** Follow its `SKILL.md` (paths in Phase 0) instead of OMC `ralph`. Hand it the plan path with its stories and acceptance criteria, each targeted dimension's command as a gate (one command per gate id), `--no-deslop` if passed, and "no questions: write assumptions down". `--critic` has no Codex counterpart: ralph's fresh reviewer takes that place. It runs its own gate folder, apart from `<g>`. Go on to Phase 3 only after its `RALPH COMPLETE` line; on `RALPH STOPPED` stop and report its reason. The `repoQualityClass` and `omc ralph verify` bullets above are Claude Code only. For a resume, note its gate folder (`ralph-state.mjs find --cwd '<repo>'` lists it).
+- Do not interrupt on "The boulder never stops" continuations (Claude Code) or banker ralph's Stop-hook continuations (Codex) — that signal means ralph's iteration continues. Let it run to completion.
 
 ### Phase 3 — QA GATE: banker ralph-qa (QA 순환)
 - After ralph completes, invoke ralph-qa once with `--gates=<dimensions> --lens=both`: every dimension the project supports (default `typecheck,lint,build,tests`, or the `--qa` list; `scenarios` adds adversarial scenario rows). One run covers every dimension: its gate conjunct holds only when each one does.
@@ -87,7 +89,7 @@ Several bundled skills are loop/persistence engines. To prevent competing loops,
 - This is a belt-and-suspenders gate over ralph's own regression: a final all-green check under ultra-init's authority (ralph has already finished, so there is no competing loop).
 
 ### Phase 4 — COMPLETE (완료 / 정리)
-- `Skill("oh-my-claudecode:cancel")` → clear ralph / mode state cleanly (don't leave stale state files).
+- `Skill("oh-my-claudecode:cancel")` → clear ralph / mode state cleanly (don't leave stale state files). On Codex banker's `ralph` sets no OMX mode: there is nothing to cancel for it.
 - Emit a **concise, evidence-backed summary**: what was planned (plan path), stories delivered, files touched, reviewer verdict, deslop result, ralph-qa's verdict line with its gate block (per dimension), and the baseline failures that predate the change (the `baseline` gates and ralph's closeout list), listed apart.
 - **Resuming an interrupted run:** the plan file in `.omc/plans/` and ralph's PRD (`.omc/state/sessions/<id>/prd.json`, with its stories' `passes` flags) are the durable record. Hand both to a new `ralph` run; for a new session, prepare it with `ready-compact --hand-off`, which must also carry the `<g>` path (the pre-change record) and the first ralph session id. On resume, Phase 3 judges against the iteration-0 records in `<g>`; never run iteration 0 again on the half-changed tree, and do not judge by the baseline a new ralph records there: either would excuse regressions made before the interruption. The same goes for a resume in the same session.
 

package/skills/ultraqa/references/gate-evidence.mjs

@@ -1,9 +1,9 @@
 #!/usr/bin/env node
 /**
  * gate-evidence: a gate command run by this program, not by the model, and what it saw, as the evidence ralph-qa
- * (--gates), ultraqa, all-in-one, ultra-init and ultra-chain judge by. The source of truth is
- * skills/ralph-qa/references/gate-evidence.mjs; skills/ultraqa/references holds a byte-identical copy
- * (scripts/sync-adapter.js).
+ * (--gates), ultraqa, the Codex ralph, all-in-one, ultra-init and ultra-chain judge by. The source of truth is
+ * skills/ralph-qa/references/gate-evidence.mjs; skills/ultraqa/references and codex/skills/ralph/scripts hold
+ * byte-identical copies (scripts/sync-adapter.js).
  *
  *   node gate-evidence.mjs init [--cwd <repo>]   the gate folder <g> of one ralph-qa run (ralph-qa-gates.<letters and digits>): made
  *        once before the baseline, kept across the iterations, apart from the payload folder seat-run.mjs makes anew

package/skills/ultraqa/references/run-folder.mjs

@@ -1,8 +1,8 @@
 #!/usr/bin/env node
 /**
- * run-folder: the private run folders of ralph-qa and ultraqa, and the work tree snapshot their gate evidence is judged
- * against. The source of truth is skills/ralph-qa/references/run-folder.mjs; skills/ultraqa/references holds a
- * byte-identical copy (scripts/sync-adapter.js).
+ * run-folder: the private run folders of ralph-qa, ultraqa and the Codex ralph, and the work tree snapshot their gate
+ * evidence is judged against. The source of truth is skills/ralph-qa/references/run-folder.mjs;
+ * skills/ultraqa/references and codex/skills/ralph/scripts hold byte-identical copies (scripts/sync-adapter.js).
  *
  *   bases(ctx)                        { base, known }: this run's base, and every base a run of any version used
  *   sharedAbove(p)                    the nearest folder above p every account may write in, or null (POSIX)

package/skills/visual-ralph/SKILL.md

@@ -15,11 +15,16 @@ description: "프론트엔드 UI 를 레퍼런스(생성 이미지, 정적 이
 
 | 단계 | Claude Code (OMC) | Codex (OMX) |
 |---|---|---|
-| 구현 루프 | `Skill("oh-my-claudecode:ralph")` | OMX `$ralph` |
+| 구현 루프 | `Skill("oh-my-claudecode:ralph")` | banker `$banker-ralph`(플러그인 길 `$banker:ralph`). OMX 0.21 이 `$ralph` 를 없애 OMX 것은 부르지 않음 |
 | 비주얼 판정 | `Skill("oh-my-claudecode:visual-verdict")` | OMX Visual Verdict |
 | 이미지 이해 | **Read 툴 네이티브**(이미지 직접 판독) | `vision` 에이전트 |
 | UI 레퍼런스 생성 | **Stitch**(`/banker:setup-stitch` MCP UI-gen); 없으면 **정적/라이브 URL 레퍼런스**로 폴백(생성 불요) | `$imagegen` + `omx imagegen continuation`(Stop-hook 큐) |
-| 아티팩트/상태 경로 | `.omc/artifacts/visual-ralph/<slug>/` | `.omx/artifacts/visual-ralph/<slug>/` · `.omx/state/sessions/` |
+| 아티팩트/상태 경로 | `.omc/artifacts/visual-ralph/<slug>/` | `.omx/artifacts/visual-ralph/<slug>/` |
+
+Codex 의 banker `ralph` 는 `SKILL.md` 를 읽고 따른다. npm 길 `<S>/../banker-ralph/SKILL.md`, 플러그인 길 `<S>/../../codex/skills/ralph/SKILL.md`(`<S>` 는 이 스킬 폴더).
+- 그 ralph 는 게이트 기록, 검토, 슬롭 점검이 지금 작업 트리와 맞아야 끝난다. 저장소 안에 새로 쓰는 스샷과 판정 파일도 작업 트리를 바꾼다
+- 그래서 반복마다 스샷과 Visual Verdict 를 ralph 의 슬롭 점검 기록 전에 끝낸다. 그 뒤에 새 스샷을 찍었으면 슬롭 점검 기록, 게이트, 검토를 다시 한다
+- Visual Verdict 점수 `>= 90` 은 ralph 검토자에게 수용 기준으로 넘긴다
 
 생성 레퍼런스 경로가 양쪽 다 없으면(imagegen 부재) → **정적 레퍼런스 이미지나 라이브 URL을 요구**한다
 (생성 없이도 루프는 성립한다). 라이브 URL 클론은 이 스킬이 흡수한다(구 `$web-clone` 미사용).
@@ -53,7 +58,7 @@ description: "프론트엔드 UI 를 레퍼런스(생성 이미지, 정적 이
 - **생성 레퍼런스**: 위 런타임 매핑의 imagegen 경로로 UI 서술에서 생성. 프롬프트 요건: `ui-mockup`
   분류, viewport/종횡비·대상 표면, 레이아웃·계층·타이포·컬러 무드·정확한 텍스트, 로고/워터마크 금지,
   불가능한 UI 디테일·판독불가 텍스트 회피.
-  - Codex 런타임에서 활성 Ralph 루프 중 내장 이미지 툴 호출 전엔 continuation 체크포인트를 큐잉:
+  - Codex 런타임에서 OMX 루프(OMX 0.20.x 의 ralph 모드 등)가 켜져 있을 때만, 내장 이미지 툴 호출 전에 continuation 체크포인트를 큐잉(banker `ralph` 는 OMX Stop 체크포인트를 쓰지 않으므로 이 단계가 없음. 이미지 생성은 구현 위임 전에 끝냄):
     `omx imagegen continuation <session> --artifact <slug> --generated-dir "$CODEX_HOME/generated_images/<session>" --work-dir ".omx/artifacts/visual-ralph/<slug>"`
     (내장 이미지 생성이 턴을 즉시 종료할 수 있어, 다음 Stop 체크포인트가 아티팩트 회수·QA를 재개.)
 - 승인된 레퍼런스는 **워크스페이스로 복사**한다(예: `.omc/artifacts/visual-ralph/<slug>/reference.png`).
@@ -109,7 +114,7 @@ Route/표면: <route 또는 컴포넌트>
 완료 전 build/lint/test 실행.
 명시 요청 없인 방향전환 금지."
 ```
-(Claude=`Skill("oh-my-claudecode:ralph")`, Codex=`$ralph` 에 위 지침 전달.)
+(Claude=`Skill("oh-my-claudecode:ralph")`, Codex=banker `$banker-ralph`(플러그인 길 `$banker:ralph`) 에 위 지침 전달.)
 
 ## 함정
 - **승인 게이트 스킵 금지** — 러프 이미지를 진실원본으로 굳히면 방향이 틀어진다.