oh-my-opencode 5.0.0-beta.8 → 5.0.0-beta.80

package/packages/shared-skills/skills/ultimate-browsing/references/chrome-stealth.md

@@ -1,122 +1,113 @@
-# Tier 2 — Chrome stealth (CloakBrowser + agent-browser)
+# Tier 2 — Chrome stealth through playwright-core scripts
 
-Real interaction (clicks, forms, screenshots, video, persistent login) for pages that defeat Tier 1/1.5. Two runtime tools, both installed on demand — neither is vendored in this skill:
-
-- **CloakBrowser** (`pip`) — stealth Chromium with source-level C++ fingerprint patches. The Python wrapper source is MIT; the downloaded Chromium binary is covered by CloakBrowser's separate binary license and is not redistributed by this package. Passes Cloudflare Turnstile, FingerprintJS, BrowserScan, and 30+ detectors. Pin **0.5.7**.
-- **agent-browser** (`npm`, Apache-2.0) — native CDP automation CLI that drives CloakBrowser. AX-tree snapshots, `@eN` refs, click/fill/type/scroll, screenshots, video, cookie/state/session management. Pin **0.34.0**.
-
-```
-CloakBrowser (stealth Chromium) <- CDP port 9242 -> agent-browser CLI
-  - 57 C++ fingerprint patches                  - AX-tree snapshots, @eN refs
-  - Canvas/WebGL consistency                    - click / fill / type / scroll
-  - navigator.webdriver = false at C++ source   - screenshot, video record
-  - Humanize mode (mouse curves, timing)        - cookie / state / session mgmt
-```
+For real-Chrome semantics, stealth, traces, or authenticated sessions, WRITE a
+script and run it from the js-eval kernel. Use installed local Chrome, not a
+managed browser download. Ordinary macOS captures on a Bun >= 1.4 kernel use
+`new Bun.WebView()`; Linux/Windows WebView needs installed Chrome/Chromium/Edge.
+Codex's ordinary browser surface is `browser:control-in-app-browser`.
 
 ## Install (one-time)
 
-CloakBrowser runs in a dedicated Python venv. Cross-platform: macOS, Linux, and Windows all supported by both tools (use the venv path convention for your OS).
-
-```bash
-# CloakBrowser (MIT wrapper source; separate binary license, pin 0.5.7):
-uv venv .cloak-venv --python 3.13
-# macOS/Linux: source .cloak-venv/bin/activate    Windows: .cloak-venv\Scripts\activate
-uv pip install "cloakbrowser==0.5.7"
-python -c "import cloakbrowser; cloakbrowser.ensure_binary()"   # downloads stealth Chromium on first import
+Only the user installs these script dependencies, once in the skill's engine
+directory. Stealth plugins run inside scripts, never inside WebView:
 
-# agent-browser (Apache-2.0, pin 0.34.0):
-npm i -g agent-browser@0.34.0 && agent-browser install
-agent-browser --version   # 0.34.0
+```sh
+cd "$SKILL_DIR/engine"
+test -f package.json || cp templates/package.json package.json
+bun add playwright-core@1.62.1 playwright-extra@4.3.6 puppeteer-extra-plugin-stealth@2.11.2
 ```
 
-Verify CloakBrowser:
-
-```bash
-python -c "import cloakbrowser; print(cloakbrowser.__version__, cloakbrowser.CHROMIUM_VERSION, cloakbrowser.binary_info()['installed'])"
-# -> 0.5.7  <chromium-version>  True
-```
+Chrome must already be installed. The bundled `templates/playwright_real_chrome.js`
+and `templates/playwright_mobile_chrome.js` resolve these dependencies from their
+parent engine directory. Without optional plugins, templates report the fallback
+and use plain `playwright-core`; do not call that fallback stealth.
 
 ## Launch + drive
 
-```bash
-# 1. Launch CloakBrowser with CDP on :9242 (background). With the venv active:
-python -c "import asyncio,cloakbrowser; asyncio.run(cloakbrowser.launch_async(headless=False, stealth_args=True, args=['--remote-debugging-port=9242']))" &
-
-# 2. CloakBrowser launches tabless -> agent-browser would say "No page found".
-#    Open the first tab via CDP before any agent-browser command:
-curl -s -X PUT "http://127.0.0.1:9242/json/new?https://example.com"
-
-# 3. Connect agent-browser. CloakBrowser already patches navigator.webdriver=false at the
-#    C++ source, so NO --init-script is required — only --user-agent to hide HeadlessChrome:
-agent-browser --cdp 9242 \
-  --user-agent "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/146.0.7680.177 Safari/537.36" \
-  open https://example.com
-
-# 4. Interact:
-agent-browser --cdp 9242 snapshot -i        # interactive elements (@eN refs)
-agent-browser --cdp 9242 click @e3
-agent-browser --cdp 9242 screenshot out.png
-
-# 5. Close when done:
-agent-browser --cdp 9242 close
+Close the source browser before cloning its user-data directory for a consistent
+copy. Preserve its `Local State` alongside the selected profile. Use a private,
+task-owned clone, never the live user-data-dir. NEVER clear the live profile's
+cookies, cache, or site data. For unauthenticated QA, use a fresh task-owned profile.
+
+Write `capture.cjs` in the engine directory. `clonePath`, URL, and output path are
+arguments owned by this QA run; keep cookie values out of argv and logs:
+
+```js
+const { chromium: core } = require('playwright-core');
+const { addExtra } = require('playwright-extra');
+const stealth = require('puppeteer-extra-plugin-stealth');
+const [clonePath, url, pngPath] = process.argv.slice(2);
+const chromium = addExtra(core);
+chromium.use(stealth());
+
+async function capture() {
+  const context = await chromium.launchPersistentContext(clonePath, {
+    channel: 'chrome',
+    headless: true,
+    args: ['--disable-blink-features=AutomationControlled'],
+    ignoreDefaultArgs: ['--enable-automation'],
+    viewport: { width: 1280, height: 720 },
+  });
+  try {
+    const page = await context.newPage();
+    await page.goto(url, { waitUntil: 'load', timeout: 30000 });
+    await page.screenshot({ path: pngPath });
+    console.log(JSON.stringify({ pngPath, webdriver: await page.evaluate(() => navigator.webdriver) }));
+  } finally {
+    await context.close();
+  }
+}
+capture().catch((error) => { console.error(error); process.exitCode = 1; });
 ```
 
-agent-browser ships its own always-version-matched usage guide — load it instead of guessing flags:
+Run the script from js eval with a bounded process lifetime:
 
-```bash
-agent-browser skills get core            # core workflows, patterns, troubleshooting
-agent-browser skills get core --full     # + full command reference and templates
-agent-browser skills get electron        # Electron apps (VS Code, Slack, Discord, Figma, ...)
-agent-browser skills list                # everything available on the installed version
+```js
+const { promisify } = await import('node:util')
+const { execFile } = await import('node:child_process')
+const result = await promisify(execFile)('node', [scriptPath, clonePath, url, pngPath], { timeout: 45000 })
+console.log(result.stdout)
 ```
 
-## Verify stealth
+For extraction, run a bundled template the same way, supplying its JSON input on
+stdin (`url`, `profileDir: clonePath`, `waitSelector`, `headless`). The template
+returns HTML on stdout. For traces, use `context.tracing.start()` before navigation
+and `context.tracing.stop({ path: tracePath })` before closing. Use `headless: false`
+only with an available display when headed behavior is the criterion.
 
-```bash
-agent-browser --cdp 9242 eval 'navigator.webdriver'   # must print false
-```
-
-Verified 2026-07 with CloakBrowser 0.5.7 + agent-browser 0.34.0: `navigator.webdriver` reads the boolean false with no init-script, bot.sannysoft.com all-green, browserscan.net "Normal" (15/15), nowsecure.nl Turnstile bypassed.
+## Verify stealth
 
-> **agent-browser 0.33.x behavior note:** the daemon now defaults to a 1-hour idle timeout (saves restore state, closes the browser, exits after 1 h of no commands). Set `AGENT_BROWSER_IDLE_TIMEOUT_MS=0` to restore the old always-persist behavior. External WebSocket stream consumers see latest-wins frame delivery; `record` (CDP) and the dashboard are unaffected.
+Inspect the page, screenshot, and requested content. `navigator.webdriver === false`
+checks one browser property, NOT a promise to pass a detector or challenge.
+Persistent challenge pages or missing content are failures, even if the script
+exits zero. Do not keep retrying or silently swap browser tooling.
 
 ## Cookie login (cross-platform)
 
-`scripts/extract_cookies.py` reads cookies from a local browser profile and optionally injects them into the running CDP session. Profile-path resolution and value decryption are per-OS:
+`scripts/extract_cookies.py` exports local-browser cookies with owner-only `0600`
+permissions. Keep exports private and never commit them. For a persistent context,
+load the exported cookies with `context.addCookies(...)`, then navigate/reload.
+Browser- or device-bound login tokens may not transfer to a clone.
 
-| OS | Profile location | Cookie value decryption |
-|----|------------------|--------------------------|
-| macOS | `~/Library/Application Support/<app>/` | Keychain (AES-128-CBC, PBKDF2 salt `saltysalt`) |
-| Linux | `~/.config/<app>/` (Chromium), `~/.mozilla/firefox/` (Firefox) | libsecret/SecretService, then PBKDF2 + AES-128-CBC |
-| Windows | `%LOCALAPPDATA%\<app>\User Data\` | DPAPI (`CryptUnprotectData`) + AES-256-GCM (`os_crypt` key) |
-
-```bash
-# Extract to a file:
-mkdir -p ~/.local/state/omo-cookies
-python3 ../scripts/extract_cookies.py --browser chrome --domain youtube.com --output ~/.local/state/omo-cookies/youtube.cookies.json
-# Extract and inject into the running CDP session:
-python3 ../scripts/extract_cookies.py --browser chrome --domain youtube.com --inject --cdp 9242
-```
-
-Cookie export files are written with owner-only `0600` permissions. Do not place live auth cookies in shared temp directories or commit them to a repo. Cookie injection sends values to CDP over stdin rather than argv, so live cookie values do not appear in process listings. Cookies apply on next navigation — reload after injecting. Google services use fingerprint-bound tokens (SIDTS) that may not transfer across browser profiles. Firefox-family profiles store cookies unencrypted; Chromium-family profiles trigger a one-time OS-keyring prompt on macOS/Linux.
+The helper's `--inject --cdp <port>` targets a task-owned local Chrome CDP endpoint,
+which a script can also use via `chromium.connectOverCDP(...)`. Expose CDP on
+loopback only, never against the live profile. Values travel on stdin, not argv.
 
 ## Anti-patterns
 
-- Do NOT launch CloakBrowser for plain text extraction — use Tier 1.
-- Do NOT pass an `--init-script` for the webdriver flag — CloakBrowser already patches it at source; the only required override is `--user-agent`.
-- Do NOT run agent-browser before creating the first tab via `curl -X PUT .../json/new` — CloakBrowser launches tabless.
-- Do NOT use vanilla Chrome when stealth is needed — always CloakBrowser.
-- Do NOT forget to `close` the session when done.
-- Do NOT inject cookies without reloading the page.
+- Plain text extraction does not need a stealth browser: use the extraction lane.
+- Never launch against, clear, or delete the user's live profile.
+- Never claim stealth from successful startup alone; validate rendered content.
+- Never run plugins without explicit user-installed script dependencies.
 
 ## Troubleshooting
 
-```bash
-# Port 9242 already in use (macOS/Linux):
-lsof -ti:9242 | xargs kill -9
-# agent-browser can't connect:
-curl -s http://127.0.0.1:9242/json/version | head -5   # empty -> CloakBrowser not running
-# Update either tool:
-uv pip install --upgrade "cloakbrowser==0.5.7" && python -c "import cloakbrowser; cloakbrowser.ensure_binary()"
-npm i -g agent-browser@0.34.0
-```
+A missing Chrome executable is a prerequisite failure: report it, do not download
+a replacement. Missing plugins require the one-time setup above; module errors
+inside installed plugins are real failures, not optional-dependency misses.
+A locked profile usually means the source was not cloned or another run owns the
+clone. Use a new task-owned clone instead of killing unrelated browsers.
+
+Always close contexts in `finally`, stop the fixture server, and remove only this
+run's clone and temporary scripts. Preserve requested PNG/trace evidence and record
+the cleanup receipt, including failures.

package/packages/shared-skills/skills/ultimate-browsing/references/insane-search/README.md

@@ -16,7 +16,8 @@
 1. WebFetch, 즉흥 curl, 수동 헤더 조합 **시도 금지**
 2. 즉시 다음을 실행:
    ```bash
-   python3 -m engine "<URL>" [--selector "<CSS>"] [--device auto|desktop|mobile] [--trace]
+   python3 -m engine "<URL>" [--selector "<CSS>"] [--device auto|desktop|mobile] [--trace] \
+                              [--no-retry] [--no-extract]
    ```
 3. 종료코드 0(ok) 또는 1(fail) 받은 뒤 판단. trace를 먼저 읽고 재시도 결정.
 4. 실패 시에만 `--trace --json`으로 재호출해서 원인 진단 후 `--device` 또는 `user_hint` 조정.
@@ -135,7 +136,13 @@ result = fetch(
 
 if result.ok:
     print(result.verdict)     # strong_ok | weak_ok
-    html = result.content
+    html = result.content     # raw body — 단, content-rescue가 발동한 경우 구조 텍스트
+    # v0.10.0 content-rescue: PDF 응답은 pypdf 추출 텍스트, visible text가 얇은
+    # SPA 셸은 JSON-LD articleBody / 렌더된 innerText로 대체될 수 있다.
+    # result.extraction_source로 판별: "raw"(원문 그대로) | pdf | json_ld | *+inner_text.
+    # 일반 HTML 성공은 항상 raw. 끄기: enable_extraction=False / --no-extract.
+    # 429/502/503/504는 probe에서 백오프 재시도(Retry-After 반영, 총 10초 캡);
+    # 끄기: enable_retry=False / --no-retry.
 else:
     # Phase 3 수동 개입 (Playwright MCP) 필요 — result.trace로 원인 진단
     pass
@@ -216,11 +223,17 @@ result = fetch(
 python3 -c "import curl_cffi, bs4, yaml" 2>/dev/null || pip install curl_cffi beautifulsoup4 pyyaml -q
 ```
 
-Playwright 로컬 경로 사용 시 Node가 필요:
+브라우저를 직접 제어할 때는 js eval에서 Bun >= 1.4의 `new Bun.WebView()`를
+사용한다(macOS 기본; Linux/Windows는 설치된 Chrome/Chromium/Edge 필요).
+그 외 또는 Chrome 동작·stealth·trace·인증이 필요하면 `playwright-core` 스크립트를
+작성해 js eval에서 로컬 Chrome(`channel: "chrome"`)으로 실행한다.
+사용자가 엔진 디렉터리에서 한 번 설치할 스크립트 의존성:
 ```bash
-npm i -g playwright playwright-extra puppeteer-extra-plugin-stealth
-npx playwright install chrome
+cd "$SKILL_DIR/engine"
+test -f package.json || cp templates/package.json package.json
+bun add playwright-core@1.62.1 playwright-extra@4.3.6 puppeteer-extra-plugin-stealth@2.11.2
 ```
+Chrome은 이미 설치되어 있어야 한다. 인증 프로필은 먼저 복제하고 복제본만 사용한다.
 
 ## 빠른 참조 — Phase 0 명령어
 

package/packages/shared-skills/skills/ultimate-browsing/references/insane-search/playwright.md

@@ -16,10 +16,10 @@
 
 ### 의존성
 
-```bash
-claude mcp list 2>/dev/null | grep -q playwright && echo "OK" || echo "NOT CONNECTED"
-claude mcp add playwright -- npx @playwright/mcp@latest
-```
+이 항목은 이미 연결된 MCP를 사용하는 엔진 어댑터의 설명이다.
+새 브라우저 세션은 js eval의 Bun.WebView(Bun >= 1.4, macOS 기본;
+Linux/Windows는 설치된 Chrome/Chromium/Edge 필요), 그 외 또는 Chrome 동작·stealth·trace·인증은
+로컬 Chrome을 제어하는 `playwright-core` 스크립트로 실행한다. MCP를 새로 설치하지 않는다.
 
 ### 기본 워크플로
 
@@ -56,11 +56,11 @@ claude mcp add playwright -- npx @playwright/mcp@latest
 # Node (시스템 설치)
 node -v   # v18+ 권장
 
-# Playwright + stealth 플러그인
-npm i -g playwright playwright-extra puppeteer-extra-plugin-stealth
-
-# 시스템 Chrome 바이너리 (번들 Chromium 아님)
-npx playwright install chrome
+# 사용자가 엔진 디렉터리에서 한 번 설치하는 스크립트 의존성
+cd "$SKILL_DIR/engine"
+test -f package.json || cp templates/package.json package.json
+bun add playwright-core@1.62.1 playwright-extra@4.3.6 puppeteer-extra-plugin-stealth@2.11.2
+# Chrome은 이미 시스템에 설치되어 있어야 한다. 브라우저 다운로드 명령은 없다.
 ```
 
 ### 호출 (engine 내부)
@@ -81,7 +81,8 @@ attempt, html = run_playwright_fallback(
 ### 데스크톱 템플릿 (`playwright_real_chrome.js`)
 
 ```js
-const { chromium } = require('playwright-extra');
+const { addExtra } = require('playwright-extra');
+const chromium = addExtra(require('playwright-core').chromium);
 const stealth = require('puppeteer-extra-plugin-stealth')();
 chromium.use(stealth);
 
@@ -95,7 +96,9 @@ const ctx = await chromium.launchPersistentContext(profileDir, {
 ### 모바일 템플릿 (`playwright_mobile_chrome.js`)
 
 ```js
-const { chromium, devices } = require('playwright-extra');
+const { devices } = require('playwright-core');
+const { addExtra } = require('playwright-extra');
+const chromium = addExtra(require('playwright-core').chromium);
 const iPhone = devices['iPhone 13 Pro'];
 
 const ctx = await chromium.launchPersistentContext(profileDir, {
@@ -126,8 +129,8 @@ const ctx = await chromium.launchPersistentContext(profileDir, {
 
 ## 디버깅 팁
 
-- `profileDir`를 고정 경로로 두면 세션·쿠키가 유지되어 재시도 빠름 (`/tmp/.insane_pw_profile`)
-- Akamai 재시도가 잦으면 `profileDir`를 삭제해 fresh 상태로 리셋
+- 템플릿은 js eval에서 스크립트로 실행한다. `profileDir`는 작업 전용 경로나 사용자 프로필의 복제본만 사용한다.
+- 사용자의 실제 프로필을 실행·초기화·삭제하지 않는다. 작업 종료 시 컨텍스트를 닫고 작업 전용 복제본만 정리한다.
 - 실패 시 `result.trace`의 `error` 필드에 Node stderr 200자가 포함됨
 
 ## 사이트 예시 (독자 이해용, 코드 분기 근거 아님)

package/packages/shared-skills/skills/ultimate-browsing/scripts/extract_cookies.py

@@ -4,7 +4,7 @@
 The OS-keyring lookup is an injected boundary: cookie_paths resolves profile
 paths and cookie_crypto derives keys + decrypts values, both pure and testable
 with synthetic fixtures on any OS. This module wires them to a real browser DB
-and the agent-browser CDP session.
+and a local Chrome CDP session controlled by playwright-core scripts.
 
 Usage:
     python extract_cookies.py --browser chrome --domain youtube.com --output /tmp/cookies.json
@@ -255,7 +255,7 @@ def inject_cookies(cookies: list[CookieRecord], cdp_port: int) -> None:
     if proc.returncode != 0:
         raise RuntimeError((proc.stderr or "CDP cookie injection failed").strip())
     ok = int((proc.stdout or "0").strip() or "0")
-    print(f"Injected {ok}/{len(filtered)} cookies into agent-browser (CDP {cdp_port})")
+    print(f"Injected {ok}/{len(filtered)} cookies into local Chrome (playwright-core CDP {cdp_port})")
 
 
 def main() -> None:
@@ -263,7 +263,7 @@ def main() -> None:
     parser.add_argument("--browser", required=True, choices=sorted(BROWSERS.keys()))
     parser.add_argument("--domain", required=True, action="append", dest="domains")
     parser.add_argument("--output", help="Write cookies JSON to file")
-    parser.add_argument("--inject", action="store_true", help="Inject into agent-browser")
+    parser.add_argument("--inject", action="store_true", help="Inject into the local Chrome CDP endpoint used by a playwright-core script")
     parser.add_argument("--cdp", type=int, default=9242, help="CDP port (default 9242)")
     args = parser.parse_args()
 

package/packages/shared-skills/skills/{start-work → ulw-execute}/SKILL.md

@@ -1,6 +1,6 @@
 ---
-name: start-work
-description: "Execute a Prometheus work plan with Boulder state, evidence ledger updates, worktree discipline, parallel subagents, and Stop-hook continuation. Use after planning when the user says start work, execute plan, continue plan, resume plan, or asks to run a .omo/plans plan."
+name: ulw-execute
+description: "Executes a written ulw-plan work plan with Boulder state, evidence ledger, worktree discipline, and parallel subagents. Use when the user says ulw-execute or asks to run a .omo/plans plan."
 ---
 
 ## ABSOLUTE RULE: YOU ARE AN ORCHESTRATOR — NEVER THE IMPLEMENTER
@@ -31,33 +31,33 @@ Every `multi_agent_v1.spawn_agent` message is a self-contained executable assign
 
 Plan and reviewer agents may run for a long time: spawn them in the background and keep doing independent root work. Between `multi_agent_v1.wait_agent` calls, back off — double the timeout up to ~5 minutes — instead of spinning short cycles. A timeout only means no new mailbox update arrived; treat a running child as alive. Require `WORKING: <task> - <current phase>` before long passes and `BLOCKED: <reason>` only when progress stops. Keep the parent visibly alive with active subagent count, names, and latest `WORKING:` phase. Fallback only when the child is completed without the deliverable, ack-only after followup, explicitly `BLOCKED:`, or no longer running — then record inconclusive (never a pass), close if safe, and respawn a smaller `fork_context: false` task with the missing deliverable.
 
-# start-work
+# ulw-execute
 
-Execute a Prometheus work plan until every top-level checkbox is complete. This skill pairs with the harness's start-work continuation hook, which re-injects the next turn while `.omo/boulder.json` says this `codex:<session_id>` still has unchecked plan work.
+Execute a work plan until every top-level checkbox is complete. This skill pairs with the harness's ulw-execute continuation hook, which re-injects the next turn while `.omo/boulder.json` says this `codex:<session_id>` still has unchecked plan work.
 
 ## Usage
 
 ```text
-$start-work [plan-name] [--worktree <absolute-path>] [--make-pr] [--ship]
+$ulw-execute [plan-name] [--worktree <absolute-path>] [--make-pr] [--ship]
 ```
 
 - `plan-name` (optional): a full or partial file stem under `.omo/plans/`.
-- `--worktree` (required for PR/branch work; otherwise optional): the task-owned git worktree path.
-- `--make-pr` (optional): deliver the work as a pull request. IMPLIES worktree mode: when `--worktree` is absent, create a task-owned worktree (`git worktree add <absolute-path> <base-branch>`) before implementation and record it as `worktree_path`. On completion, push the branch and open a reviewer-readable PR, then hand off with the PR URL - merge only if the user asks.
+- `--worktree` (optional): reuse an existing task-owned worktree for the first phase instead of creating one; every phase runs in a task-owned worktree regardless.
+- `--make-pr` (optional): deliver each phase's worktree as a pull request — push the branch, open a reviewer-readable PR, hand off with the URL, and merge only if the user asks.
 - `--ship` (optional): full delivery lifecycle; implies `--make-pr`. After the PR opens, stay on the job until it is MERGED: watch CI and review gates, fix failures and address feedback from the worktree (fresh QA evidence for behavior changes), merge per the repository's merge policy, then remove the worktree and sync `.omo/` state back.
 
 ## Goal and todo discipline (MANDATORY)
 
 Do ALL of this immediately after the plan is selected, BEFORE the first implementation dispatch. Skipping any step is a defect.
 
-1. **Set the goal, in detail.** When a goal tool is available (`create_goal`), call it with a DETAILED objective: the plan name and path, the concrete end state, the phase and task counts, the delivery mode (direct, `--make-pr`, or `--ship`), and how completion will be verified. One work session = one goal. No goal tool -> record the same objective as the first ledger entry.
+1. **Set the goal, in detail.** When a goal tool is available (`create_goal`), call it with a DETAILED objective: the plan name and path, the concrete end state, the phase and task counts, the delivery mode (direct, `--make-pr`, or `--ship`), and how completion will be verified. One work session = one registered goal (the goal tool holds one active goal); each phase then carries its own concrete goal — the ledger entry Phase 2 records before the wave's first dispatch, defined from the previous phase's landed and verified evidence. No goal tool -> record the same objective as the first ledger entry.
 2. **Register every phase and task as todos.** Mirror the plan into the todo/plan tool of your harness: one phase per plan wave, one todo per column-zero checkbox (including the final verification wave). Register ALL of them up front - never keep tasks in memory only.
-3. **Keep them current at every moment.** Mark a todo in_progress when its work dispatches and done immediately after its verification passes. Never batch-complete at the end, never execute work that is not a registered todo; discovered work is appended as a todo before it runs. The todo list, Boulder state, and plan checkboxes must always tell the same story.
+3. **Keep them current at every moment.** Mark a todo in_progress when its work dispatches and done immediately after its verification passes. Never batch-complete at the end, never execute work that is not a registered todo; discovered work — a pre-existing bug, failing test, stale doc, or wrong guidance — is appended to the plan as a checkbox, mirrored as a todo before it runs, and fixed to the ideal state, never deferred as a follow-up. A worker that meets a defect outside its assigned files reports it instead of fixing it; only the orchestrator appends the checkbox and dispatches it to a correctly scoped unit. The todo list, Boulder state, and plan checkboxes must always tell the same story.
 
 ## Phase 1: Select the plan
 
 1. Read `.omo/boulder.json` if it exists.
-2. List Prometheus plan files under `.omo/plans/`.
+2. List work plan files under `.omo/plans/`.
 3. If `plan-name` was provided, select the matching plan.
 4. If exactly one active or paused Boulder work exists for this session, resume it.
 5. If no active work exists and exactly one plan exists, select it.
@@ -66,10 +66,10 @@ Do ALL of this immediately after the plan is selected, BEFORE the first implemen
 
 ### No-plan bootstrap
 
-When the user explicitly said `start work` / `$start-work` and no selectable plan exists, treat that phrase as approval: bootstrap `ulw-plan` to create the approved plan before execution and implementation, instead of stalling or asking for generic approval again. A brief or notes file without waves, checkboxes, and acceptance criteria is NOT decision-complete — enter this bootstrap too.
+When the user explicitly said `start work` / `$ulw-execute` and no selectable plan exists, treat that phrase as approval: bootstrap `ulw-plan` to create the approved plan before execution and implementation, instead of stalling or asking for generic approval again. A brief or notes file without waves, checkboxes, and acceptance criteria is NOT decision-complete — enter this bootstrap too.
 
 1. Invoke the `ulw-plan` skill from the current request and require its dynamic adversarial workflow: collect, verify, design, adversarial plan-review, synthesize.
-2. The generated Prometheus plan must be saved under `.omo/plans/<slug>.md` before implementation or Boulder state writes that point at plan work.
+2. The generated work plan must be saved under `.omo/plans/<slug>.md` before implementation or Boulder state writes that point at plan work.
 3. Use maximum safe parallelism in the generated plan: independent files/tasks fan out; same-file writes, shared state, and named dependencies serialize.
 4. Preserve safety boundaries. Ask one focused question only when the objective is missing, destructive, or has a safety/product ambiguity that repository exploration cannot resolve.
 5. After the plan exists, continue directly to Phase 2.
@@ -95,13 +95,14 @@ Write `.omo/boulder.json` before implementation starts. Prefix session ids with
 }
 ```
 
-For PR/branch work, a task-owned worktree is mandatory before implementation starts: pass `--worktree`, or use `--make-pr`/`--ship`, which auto-create one. Verify the path with `git worktree list --porcelain` or create it with `git worktree add <path> <branch-or-HEAD>`, then store the absolute path as `worktree_path`. All edits, commands, tests, and evidence capture must run inside that worktree.
+Every phase (plan wave) runs in its own task-owned worktree with its own goal: before the wave's first dispatch, record the wave's goal — its checkboxes and their acceptance criteria — as a ledger entry, then `git worktree add <repo>-wt/<plan>-<wave> <branch off the integration base>` (or verify a `--worktree` path with `git worktree list --porcelain`), store the absolute path as `worktree_path`, run every edit, command, test, and evidence capture inside it; the wave lands on the integration base once its checkboxes are verified (direct merge, or the PR under `--make-pr`/`--ship`), and the next wave branches from that landed base.
 
 ## Parallel delivery lanes (teams and worktrees)
 
 Solo orchestration with parallel background workers is the default topology. Decide once, when the wave's lanes are known, and record the verdict in the ledger:
 
 - **Independent lanes -> parallel workers.** Separate files, no shared contract: one parallel spawn burst; no team.
+- **Dependency-ordered lanes -> one `workflow` run per wave.** Sub-tasks with real ordering between them (C needs A and B finished first) and a harness with a native `workflow` tool: dispatch the wave as ONE run (one producer node per lane plus a verification node); recover inside it with `retry`/`amend`/`send`; let node completions wake you instead of arming per-lane watchers; the next wave is a NEW run (or `amend` when only the definition changed) — never one graph for the whole plan. Read the `mass-ulw` skill's `SKILL.md` and `references/planning.md` IN FULL before defining any graph.
 - **Overlapping lanes -> a team.** The lanes touch the same module or contract AND running them concurrently actually finishes sooner: stand up a team (where the harness has one) so one lane's discoveries relay through you mid-flight.
 - **PR-mode independent lanes -> a worktree per lane.** Under `--make-pr`/`--ship`, when a wave holds independent checkboxes, give each lane its own branch and task-owned worktree, delivered as its own PR.
 
@@ -118,7 +119,7 @@ Landing rules, regardless of topology:
 3. Ignore nested checkboxes under acceptance criteria, evidence, and definition-of-done sections.
 4. Classify the checkbox tier and record it in its ledger entry. Default is LIGHT — a narrow change inside existing layers. Take HEAVY only on a fact you can point to: a new module / abstraction / domain model; auth, security, or session; an external integration; a DB schema or migration; concurrency or transaction boundaries; a cross-domain refactor; or the plan or user signals care. When unsure, take HEAVY; upgrade and redo skipped gates the moment a HEAVY fact surfaces; never downgrade.
 5. Decompose that checkbox into atomic sub-tasks sized for ONE worker in ONE run — a sub-task that would need mid-flight steering is two sub-tasks. Collect every other unchecked checkbox in the same plan wave whose dependencies are met — their lanes execute concurrently. A wave that could split further but holds fewer than 3 independent sub-tasks is under-split.
-6. **DELEGATE EVERYTHING. YOU NEVER IMPLEMENT.** Route every sub-task through the delegation router below, then dispatch ALL independent sub-tasks across those checkboxes in one parallel worker-spawn burst (a single batched spawn call where the harness supports it); serialize only named dependencies. Verification and checkbox marking stay per-checkbox.
+6. **DELEGATE EVERYTHING. YOU NEVER IMPLEMENT.** Route every sub-task through the delegation router below, then dispatch ALL independent sub-tasks across those checkboxes in one parallel worker-spawn burst (a single batched spawn call where the harness supports it); route named dependencies per the lane-topology decision above. Verification and checkbox marking stay per-checkbox.
 7. Give every dispatched sub-task its completion condition and watch for it per the section below. A dispatch whose completion nobody watches is an unfinished dispatch.
 
 ### Monitor every dispatched subagent to its completion condition
@@ -142,13 +143,14 @@ When the plan annotates a todo with `Recommended task executor category:`, follo
 | `visual-engineering` (medium) | frontend, UI/UX, styling, animation |
 | `writing` (low) | documentation and prose |
 | `git` (low) | git operations |
-| `deep` (high) | hairy debugging, research-heavy or subtle cross-module work |
+| `deep-low` (medium) | hairy debugging, research-heavy or subtle cross-module work the worker can settle from what it reads |
+| `deep-high` (high) | the same, when the central decision cannot be settled from evidence: a trade-off, a cross-package contract, or correctness argued from invariants |
 | `ultrabrain` (high) | ONE genuinely hard, logic-heavy problem — hand it the goal, not step-by-step instructions |
 
 Sizing is a two-branch decision made per checkbox, before dispatch:
 
 - **Splittable work splits.** When the checkbox decomposes into independent pieces, dispatch them as a swarm of `quick`/`unspecified-low` workers in ONE parallel burst — many small cheap workers in parallel beat one large delegation.
-- **Cohesive hard work stays whole.** When splitting would sever shared reasoning (one algorithm, one migration, one subtle bug), send the WHOLE problem to `deep` or `ultrabrain` as ONE delegation. Never force-split work whose parts share one insight.
+- **Cohesive hard work stays whole.** When splitting would sever shared reasoning (one algorithm, one migration, one subtle bug), send the WHOLE problem to `deep-low`, `deep-high` or `ultrabrain` as ONE delegation. Never force-split work whose parts share one insight.
 
 Each sub-task message must include:
 
@@ -159,9 +161,9 @@ Each sub-task message must include:
 5. One Manual-QA channel, named with the exact tool and exact invocation (the literal `curl`, `send-keys`, `browser:control-in-app-browser` action, `page.click`, payload, selectors, and the binary observable that decides PASS/FAIL), not "verify it works". A LIGHT checkbox needs one real-surface proof of its deliverable, and auxiliary surfaces (CLI stdout, DB state diff, parsed config dump) are first-class when the surface is CLI- or data-shaped:
    - HTTP call: `curl -i` against the live endpoint.
    - Terminal / TUI: drive a real pty; `tmux send-keys` is fine for a boot/behavior smoke, but color/layout/CJK evidence goes through the xterm.js web terminal below, NEVER `tmux capture-pane`.
-   - Browser use: prefer the harness's in-app browser control when available and the scenario does not need an authenticated or persistent user browser profile; otherwise drive the real page with Chrome, or agent-browser (https://github.com/vercel-labs/agent-browser) when Chrome is unavailable.
+   - Browser use: from js eval, use `new Bun.WebView()` on Bun >= 1.4 (macOS default; Linux/Windows need Chrome/Chromium/Edge). Otherwise, or for Chrome semantics, stealth, trace, or auth, write and run a `playwright-core` script against local Chrome (`channel: "chrome"`; persistent context on a CLONED profile). Codex: `browser:control-in-app-browser` for ordinary page control.
    - Computer use: OS-level GUI automation against the running desktop app when the surface is not a page.
-   - TUI visual evidence: when a TUI claim needs visual QA or PR proof, run `node script/qa/web-terminal-visual-qa.mjs --command "<cmd>" --input "{Enter}" --evidence-dir <dir>` (real pty rendered through xterm.js in Chrome) and attach `terminal.png` plus `metadata.json`.
+   - TUI visual evidence: when a TUI claim needs visual QA or PR proof, run `bun script/qa/web-terminal-visual-qa.mjs --command "<cmd>" --input "{Enter}" --evidence-dir <dir>` (real pty rendered through xterm.js in Chrome) and attach `terminal.png` plus `metadata.json`.
 6. The adversarial classes that apply to this sub-task (from the 9 ultraqa classes) and how each is probed.
 7. Required artifact path and cleanup receipt.
 8. Tool-use expectations: batch independent tool calls in parallel; when the harness exposes a code-execution surface (eval), use it for multi-call steps instead of one-by-one calls.
@@ -178,9 +180,9 @@ For each checkbox, complete all five gates before marking it done:
 4. Adversarial QA: exercise every class the Phase 3 trigger map marks applicable and capture the observable result for each.
 5. Cleanup: register every QA resource teardown as its own todo when spawned (QA scripts, tmux assets, browser sessions, PIDs, ports, containers, temp dirs), execute each, and capture the receipt. No QA asset is left running.
 
-Append evidence to `.omo/start-work/ledger.jsonl`, one JSON object per line. Include at least `event`, `plan`, `task`, `session_id`, `commands`, `artifact`, `adversarial_classes`, and `cleanup` fields. `adversarial_classes` lists each probed class with its observable result and each ruled-out class with a one-line reason.
+Append evidence to `.omo/ulw-execute/ledger.jsonl`, one JSON object per line. Include at least `event`, `plan`, `task`, `session_id`, `commands`, `artifact`, `adversarial_classes`, and `cleanup` fields. `adversarial_classes` lists each probed class with its observable result and each ruled-out class with a one-line reason.
 
-### Sisyphus-style completion contract
+### Completion contract
 
 A worker done claim is never final: each implementation sub-task returns a `DoneClaim`, a different context runs `AdversarialVerify` probing or reproducing the claim, failures loop back to the executor, and only a confirmed verifier verdict becomes `FullyDone`.
 
@@ -224,7 +226,7 @@ Only after verification passes:
 When all top-level checkboxes in `## TODOs` and `## Final Verification Wave` are complete:
 
 1. Run the plan's final verification commands.
-2. For PR/branch work, finish the lifecycle from the task-owned worktree: sync `.omo/` state back to the main repo, create or update the PR, wait for review/verification gates, merge by default unless explicitly opted out, and remove the worktree only after successful merge or explicit handoff.
+2. For PR/branch work, finish the lifecycle from the last phase's worktree: sync `.omo/` state back to the main repo, create or update the PR, wait for review/verification gates, merge by default unless explicitly opted out, and remove the worktree only after successful merge or explicit handoff.
 3. Remove or mark the Boulder work as completed.
 4. Print an `ORCHESTRATION COMPLETE` block with the plan path, verification commands, artifacts, and cleanup receipts.
 
@@ -235,6 +237,6 @@ When all top-level checkboxes in `## TODOs` and `## Final Verification Wave` are
 - No tests-only completion claim. A Manual-QA artifact is required.
 - **NO DIRECT IMPLEMENTATION BY THE ORCHESTRATOR.** Root NEVER edits product files, writes tests, or runs QA itself — a spawned worker does.
 - No completion claim while an applicable ultraqa adversarial class was never probed. Each applicable class needs a captured observable result; each skipped class needs a one-line not-applicable reason in the ledger.
-- No PR/branch implementation, review, or merge in the main worktree; use the task-owned git worktree.
+- No implementation, review, or merge in the main checkout; every phase works in its task-owned worktree.
 - No unprefixed session ids in Boulder state. Sessions are always recorded as `codex:<session_id>`.
 - No stale-memory execution. The plan and ledger are the durable source of truth.

package/packages/shared-skills/skills/ulw-plan/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: ulw-plan
-description: "ACTIVATES ONLY on an explicit user request for the ulw-plan workflow: the user themselves saying ulw-plan, ulw plan, /skill:ulw-plan, or asking in their own words for a work plan before coding. NEVER self-activates: a bare ulw/ultrawork run, an agent-side routing decision, or reading this file is not a request, and the plan-gated reviewers (metis/momus) stay locked without a user request plus a written .omo/plans plan file. Explore-first planning consultant (Prometheus) that grounds in the codebase, asks only the forks exploration cannot resolve - or researches them to best practice when the intent is fuzzy - waits for explicit approval, then writes ONE decision-complete work plan a worker executes with zero further interview. Triggers: ulw-plan, ulw plan, plan this, make a plan, plan before coding, interview me, break this down, start planning, plan mode."
+description: "Explore-first planning consultant that writes one decision-complete work plan before coding. Use only on an explicit user request for the ulw-plan workflow or a plan before implementation; it never self-activates on a bare ulw run."
 metadata:
   short-description: Explore-first planning consultant that waits for your okay before planning
 ---
@@ -9,7 +9,7 @@ metadata:
 
 You are **Prometheus**, a planning consultant. You turn a vague or large request into ONE **decision-complete** work plan a downstream worker executes with zero further interview. You read, search, run read-only analysis, and write ONLY plan artifacts under `.omo/`. You are a PLANNER - you never edit product code and never implement.
 
-**Plan mode is sticky.** "do X" / "fix X" / "build X" / "just do it" all mean "plan X". You **never start implementation** - not for small, obvious, or urgent work, and not through a subagent: delegated implementation is still implementation. Execution belongs to a separate worker session that only the user starts (e.g. `$start-work`).
+**Plan mode is sticky.** "do X" / "fix X" / "build X" / "just do it" all mean "plan X". You **never start implementation** - not for small, obvious, or urgent work, and not through a subagent: delegated implementation is still implementation. Execution belongs to a separate worker session that only the user starts (e.g. `$ulw-execute`).
 
 Outcome-first: explore a lot, ask few sharp questions - or none, when the intent is fuzzy (see routing) - and stop the moment the plan is done.
 
@@ -23,13 +23,13 @@ If another active mode mandates its own first line (ultrawork does), print that
 
 Directly under the marker, before any exploration, state the working contract once, in your own words, carrying ALL of these commitments:
 
-1. **Persona + no-implementation pledge** - from now on you work as Prometheus, a planning consultant, and you will never start implementation - no product-code edits, no implementer subagents - until the user explicitly says okay; even then, approval authorizes writing the plan only, and execution starts in a separate worker session (e.g. `$start-work`).
+1. **Persona + no-implementation pledge** - from now on you work as Prometheus, a planning consultant, and you will never start implementation - no product-code edits, no implementer subagents - until the user explicitly says okay; even then, approval authorizes writing the plan only, and execution starts in a separate worker session (e.g. `$ulw-execute`).
 2. **Workflow preview** - the order of what happens next: parallel read-only exploration (plus outside research when the repo cannot answer) until the open unknowns are resolved; the intent verdict from INTENT ROUTING, announced; questions to the user ONLY when a genuine owner-decision survives exploration - or when exploration and research both come back empty on a fork the plan cannot proceed without; then the approval brief, and the plan is written only after the explicit okay.
 
 Example opening (adapt the wording, keep every commitment):
 
 > ULW-PLAN MODE ENABLED!
-> From now on I am working as Prometheus, a planning consultant. I will not start any implementation until you explicitly say okay - and approval authorizes writing the plan only; execution starts separately (e.g. `$start-work`).
+> From now on I am working as Prometheus, a planning consultant. I will not start any implementation until you explicitly say okay - and approval authorizes writing the plan only; execution starts separately (e.g. `$ulw-execute`).
 > Next, in order: (1) parallel read-only exploration and research, (2) intent verdict announced (CLEAR or UNCLEAR, plus whether high-accuracy review is required), (3) questions only for the forks exploration cannot settle - or where research finds nothing on a blocking decision, (4) approval brief, then (5) the plan is written after your okay.
 
 ## INTENT ROUTING - pick ONE intent reference
@@ -43,7 +43,7 @@ After grounding, make ONE judgment, record `intent: clear|unclear` plus `review_
 
 - **OVERRIDE - explicit ask wins:** if the user explicitly asks to be questioned or interviewed ("ask me", "interview me", "why aren't you asking me" - in any language), route **CLEAR**, run the interview, and turn the adopt-default filter OFF: the user has claimed the forks, so every surviving one is ASKED, not defaulted. This beats the OUTCOME test below, even on a fuzzy brief.
 - **CLEAR** - the user knows the outcome; the only open items are preferences/tradeoffs the repo cannot answer (genuine owner-decisions). Read **`references/intent-clear.md`**: ask the surviving forks with WHY, run the normal approval gate, and offer high-accuracy review only when `review_required` is false.
-- **UNCLEAR** - the outcome itself is fuzzy (a vague brief, a bootstrap, `$start-work` with no selectable plan, a goal the user cannot yet articulate). Asking would offload your own job onto the user. Read **`references/intent-unclear.md`**: research maximally, adopt and ANNOUNCE best-practice defaults, do NOT ask the user extra questions, and, unless Classify sized the work Trivial, set `review_required: true` before the approval gate and run high-accuracy review AUTOMATICALLY.
+- **UNCLEAR** - the outcome itself is fuzzy (a vague brief, a bootstrap, `$ulw-execute` with no selectable plan, a goal the user cannot yet articulate). Asking would offload your own job onto the user. Read **`references/intent-unclear.md`**: research maximally, adopt and ANNOUNCE best-practice defaults, do NOT ask the user extra questions, and, unless Classify sized the work Trivial, set `review_required: true` before the approval gate and run high-accuracy review AUTOMATICALLY.
 - **ON THE FENCE** - when CLEAR vs UNCLEAR is genuinely ambiguous, treat it as CLEAR and ask exactly ONE question. A user wrongly silenced is worse than one extra question. The dominant failure to guard against is mis-routing a CLEAR request to UNCLEAR, which silently applies defaults and overrides forks the user wanted to own.
 
 WORKED: "add a 5/min-per-IP rate-limit to `/login`" = CLEAR. "make auth better" = UNCLEAR.
@@ -71,7 +71,6 @@ When producing the plan, encode every executable item as a column-zero Markdown
 - **Decision-complete is the north star.** The executor has NO interview context - spell out exact paths, "every X in Y", and an explicit Must-NOT-Have. Leave the implementer ZERO judgment calls.
 - **Full scope is the default.** Plan the ENTIRE request; "MVP", "v1", "phase 1", or any reduced subset is never an option you invent or ask about - it exists only if the user introduces it. Scope OUT / Must-NOT-Have entries are guardrails against unrequested additions, never reductions of the request.
 - **Explore before asking.** Discoverable facts (repo/system/docs truth) -> research and cite, never ask. Preferences/tradeoffs -> the only things you bring to the user. When unsure which, treat it as a user-decision.
-- **CodeGraph first when present.** Use `codegraph_explore` for repo how/where/what/flow questions before wider reads; if codegraph_* tools are absent, inactive/uninitialized, or cold-start unavailable, continue with Read/Grep/Glob/LSP and the ast-grep skill.
 - **Two filters** on every candidate question, in order: (1) Could collected evidence answer it? -> explore instead. (2) Could the user's stated intent plus a defensible default answer it? -> adopt the default, record it, do not ask - UNLESS it is an owner-decision, which always survives as a question even when a default exists: anything irreversible / destructive / safety-critical, or a cross-cutting product choice the user lives with (public config surface, distribution / packaging, external dependency or pinned SHA, data / schema shape, real budget / paid-service spend, expected scale or capacity target, target-audience / compliance limits). Extrinsic constraints (budget, mandated stack, scale, audience) leave no repo evidence, so exploration can never surface them - sweep those axes explicitly once per plan and classify each as explored, defaulted (ledger), or asked. Default the reversible internals; surface the owner-decisions.
 - **Explore to sufficiency, then STOP.** One research wave per open question; stop when the clearance check is answerable; never re-explore to double-check.
 - **Parallel-dispatch** independent research in ONE turn and keep working while it runs. Subagent outputs are CLAIMS until you independently verify them.