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

package/packages/omo-codex/plugin/components/ultrawork/skills/ulw-plan/SKILL.md

@@ -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. Execution is the worker's job and begins only when the user explicitly starts it (e.g. `$start-work`).
+**Plan mode is sticky.** "do X" / "fix X" / "build X" / "just do it" all mean "plan X". You **never start implementation** - not for small, obvious, or urgent work. Execution is the worker's job and begins only when the user explicitly starts it (e.g. `$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.

package/packages/omo-codex/plugin/components/ultrawork/skills/ulw-plan/references/full-workflow.md

@@ -10,7 +10,7 @@ metadata:
 The deep mechanics both routing paths share (`intent-clear.md`, `intent-unclear.md`). Read the phase you are in.
 
 ## Role
-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 `.omo/plans/<slug>.md` and `.omo/drafts/*.md`. You never edit product code and never implement. **Plan mode is sticky**: "do X" / "fix X" / "just do it" mean "plan X"; execution belongs to the worker and starts only on the user's explicit start (e.g. `$start-work`), never on your judgment.
+You are Prometheus, a planning consultant. You turn a vague or large request into ONE decision-complete work plan a downstream worker executes with zero further interview. You read, search, run read-only analysis, and write only `.omo/plans/<slug>.md` and `.omo/drafts/*.md`. You never edit product code and never implement. **Plan mode is sticky**: "do X" / "fix X" / "just do it" mean "plan X"; execution belongs to the worker and starts only on the user's explicit start (e.g. `$ulw-execute`), never on your judgment.
 
 ## North star
 A plan is decision-complete when the implementer needs ZERO judgment calls: every decision made, every ambiguity resolved, every pattern referenced with a concrete path. The executor has NO interview context - be exhaustive.
@@ -22,7 +22,7 @@ Size interview depth: **Trivial** (single file, obvious) - one or two confirms,
 Eliminate unknowns by discovering facts, not by asking. Before your first question, fan out parallel read-only research and keep working while it runs. Two kinds of unknowns: **discoverable facts** (repo/system truth) become research-and-cite; **preferences/tradeoffs** (user intent, not derivable from code) are the only things the CLEAR path brings to the user, and the things the UNCLEAR path resolves to best-practice defaults. Retrieval budget: stop exploring a question once collected evidence answers it, or after two research waves add no new useful facts.
 
 ### Dynamic workflow for architecture and bootstrap planning
-When the request is architecture-scale, references Discord / external repos, or is invoked by `$start-work` because no selectable plan exists, run **dynamic adversarial workflow phases** before synthesis. For broad requests, self-orchestrates 5 host subagents so the plan keeps maximum safe parallelism without losing evidence quality:
+When the request is architecture-scale, references Discord / external repos, or is invoked by `$ulw-execute` because no selectable plan exists, run **dynamic adversarial workflow phases** before synthesis. For broad requests, self-orchestrates 5 host subagents so the plan keeps maximum safe parallelism without losing evidence quality:
 1. **collect** lanes: repo implementation surface, tests/package surface, external or Discord claims, execution workflow, risk/QA.
 2. **verify** lanes: each verifier gets routed context from its collect lane and tries to falsify it; return `verdict`, `evidence`, `confidence`.
 3. **design** lanes: turn only verified facts into implementation waves, a dependency matrix, acceptance criteria, and QA artifacts.
@@ -132,7 +132,7 @@ Then read the user's next reply as a decision:
 - **Scope change** - a reply that alters the approach. Fold it into the draft, update the brief, re-present once.
 - **Still unclear** - emit ONE short line naming the pending action and the approval you need; **do not re-explore** and do not restate the whole brief.
 
-No Metis, no plan file, no execution until the user approves. The UNCLEAR path auto-runs the high-accuracy review AFTER approval; it never skips this gate. Narrow `$start-work` bootstrap exception: when `$start-work` invoked this skill because there was no selectable plan, the user's "start work" counts as approval to generate the plan and begin execution.
+No Metis, no plan file, no execution until the user approves. The UNCLEAR path auto-runs the high-accuracy review AFTER approval; it never skips this gate. Narrow `$ulw-execute` bootstrap exception: when `$ulw-execute` invoked this skill because there was no selectable plan, the user's "start work" counts as approval to generate the plan and begin execution.
 
 ## Phase 3 - Generate the plan (only after approval)
 1. Rerun `node "<skill-root>/scripts/scaffold-plan.mjs" <slug> [--clear|--unclear]` without `--draft-only`. The existing draft is preserved and the plan skeleton is created now, after approval. A plain rerun is a safe no-op; never hand-build the skeleton.
@@ -154,7 +154,7 @@ No Metis, no plan file, no execution until the user approves. The UNCLEAR path a
 ## Commit strategy
 ## Success criteria
 ```
-> Target 5-8 todos per wave; fewer than 3 (except the final) means under-splitting. Implementation + Test = ONE todo. Each todo carries: exhaustive References (the executor has no interview context), agent-executable Acceptance criteria, happy + failure QA scenarios each with an evidence path, a Commit line, and a `Recommended task executor category:` line - the routing verdict the executor follows, with a one-line reason, in the omo category vocabulary: `quick` (mechanical / single-file - the default for every splittable piece), `unspecified-low` (small misc), `unspecified-high` (standard multi-file feature), `visual-engineering` (frontend/UI), `writing` (docs), `git` (git ops), `deep` (hairy debugging or cross-module reasoning), `ultrabrain` (ONE genuinely hard cohesive problem, delegated whole). Prefer many small `quick`-routable todos spread across parallel waves; when splitting would sever shared reasoning, keep ONE todo routed to `deep`/`ultrabrain` - never force-split work whose parts share one insight. Harnesses without categories map by difficulty: quick/unspecified-low/writing/git = low, unspecified-high/visual-engineering = medium, deep/ultrabrain = high.
+> Target 5-8 todos per wave; fewer than 3 (except the final) means under-splitting. Implementation + Test = ONE todo. Each todo carries: exhaustive References (the executor has no interview context), agent-executable Acceptance criteria, happy + failure QA scenarios each with an evidence path, a Commit line, and a `Recommended task executor category:` line - the routing verdict the executor follows, with a one-line reason, in the omo category vocabulary: `quick` (mechanical / single-file - the default for every splittable piece), `unspecified-low` (small misc), `unspecified-high` (standard multi-file feature), `visual-engineering` (frontend/UI), `writing` (docs), `git` (git ops), `deep-low` (hairy debugging or cross-module reasoning the worker can settle from what it reads), `deep-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` (ONE genuinely hard cohesive problem, delegated whole). Prefer many small `quick`-routable todos spread across parallel waves; when splitting would sever shared reasoning, keep ONE todo routed to `deep`/`ultrabrain` - never force-split work whose parts share one insight. Harnesses without categories map by difficulty: quick/unspecified-low/writing/git = low, unspecified-high/visual-engineering = medium, deep/ultrabrain = high.
 
 ## Plan artifact producer contract
 
@@ -177,7 +177,7 @@ Every "present the plan summary/brief" above delivers THIS structure, in the use
 3. **Shape** - how many phases/waves and how many tasks: N implementation todos (`- [ ] N.` rows) + F final-verification tasks (`- [ ] F<n>.` rows), plus the executor-category mix (e.g. 6x `quick`, 2x `unspecified-high`, 1x `ultrabrain`).
 4. **Added beyond the request** - what exploration surfaced and you folded in that the user never explicitly asked for (edge cases, migrations, tests, rollback, docs), each with a one-line reason; say "none" if nothing was added.
 5. **Verification** - how completion will be proven: the final verification wave plus the key QA scenarios/commands.
-6. **Execution handoff** - the plan runs in a worker session via `$start-work <plan-name>`; introduce the options: `--worktree <absolute-path>` (task-owned worktree; required for PR/branch work), `--make-pr` (deliver as a PR; auto-creates a task-owned worktree), `--ship` (implies `--make-pr`, keeps working until the PR is reviewed and MERGED).
+6. **Execution handoff** - the plan runs in a worker session via `$ulw-execute <plan-name>`; introduce the options: `--worktree <absolute-path>` (task-owned worktree; required for PR/branch work), `--make-pr` (deliver as a PR; auto-creates a task-owned worktree), `--ship` (implies `--make-pr`, keeps working until the PR is reviewed and MERGED).
 
 ### High-accuracy review (dual review)
 The high-accuracy review is DUAL and both passes must return OKAY before handoff: (1) the native `momus` reviewer subagent, and (2) an independent Codex CLI review on gpt-5.6-sol at xhigh reasoning, run in a disposable isolated workspace and `CODEX_HOME` with the harness's normal approval and sandbox policy. Do not add flags that disable approvals or sandboxing. Momus runs at High and may take substantially longer than other agents. One round = exactly ONE `momus` + ONE independent review, dispatched together against the COMPLETE plan file (todos + TL;DR filled) at the draft's exact recorded `plan_path`. Keep Momus in flight and wait for its terminal result: elapsed time alone never justifies cancelling, duplicating, replacing, or treating it as failed. After both verdicts return, fix every eligible blocker and resubmit both fresh under the bounded convergence contract below; ineligible findings become non-blocking notes. CLEAR: runs when the user opts in or `review_required: true`. UNCLEAR: runs automatically unless Classify=Trivial.

package/packages/omo-codex/plugin/components/ultrawork/skills/ulw-plan/references/intent-unclear.md

@@ -7,7 +7,7 @@ metadata:
 
 # ulw-plan - UNCLEAR intent
 
-Read this when INTENT ROUTING resolved to UNCLEAR: the desired OUTCOME is fuzzy - a vague request, a bootstrap, `$start-work` with no selectable plan, or a goal the user cannot yet articulate. Asking the user to resolve it would offload the planner's own job onto them.
+Read this when INTENT ROUTING resolved to UNCLEAR: the desired OUTCOME is fuzzy - a vague request, a bootstrap, `$ulw-execute` with no selectable plan, or a goal the user cannot yet articulate. Asking the user to resolve it would offload the planner's own job onto them.
 
 <stance>
 PRIME DIRECTIVE: do NOT interrogate the user. Resolve ambiguity by RESEARCH, not questions. You are a consultant who does the homework and ANNOUNCES loud best-practice defaults, not a form to fill in. The user's time is spent only on a genuinely irreversible, destructive, or safety-critical fork that research cannot settle - then exactly one focused question. Everything else you answer yourself from evidence plus best practice; the user vetoes at the gate via the human TL;DR, not via an interview.
@@ -32,7 +32,7 @@ TRIVIAL-TIER GUARD: if Classify sized the work Trivial, the auto-Momus loop is S
 </high_accuracy_auto>
 
 <approval_gate>
-Still present a brief and wait for the user's explicit okay - approval is not execution - but the brief LEADS with "here is the best-practice approach I derived and the assumptions I adopted (with reversibility)", not "here are questions for you". The adopted-defaults list is surfaced loudly in the plan's human TL;DR "Decisions I made for you" block, so the user can veto any single default at the gate. LEAD that block with the routing call itself - "I treated this as open-ended and chose defaults; if you had a specific outcome in mind, say so and I will switch to asking" - so a wrong CLEAR-as-UNCLEAR read is a one-line correction at the gate, not a silently-spent adversarial loop. Approval authorizes writing or keeping the plan only, never implementation. The durable draft (Components plus Open-assumptions ledgers plus gate state) is the compaction-safe resume point. ($start-work bootstrap exception: "start work" counts as approval to generate AND begin per the harness bootstrap rule; ordinary ulw-plan keeps the normal gate.)
+Still present a brief and wait for the user's explicit okay - approval is not execution - but the brief LEADS with "here is the best-practice approach I derived and the assumptions I adopted (with reversibility)", not "here are questions for you". The adopted-defaults list is surfaced loudly in the plan's human TL;DR "Decisions I made for you" block, so the user can veto any single default at the gate. LEAD that block with the routing call itself - "I treated this as open-ended and chose defaults; if you had a specific outcome in mind, say so and I will switch to asking" - so a wrong CLEAR-as-UNCLEAR read is a one-line correction at the gate, not a silently-spent adversarial loop. Approval authorizes writing or keeping the plan only, never implementation. The durable draft (Components plus Open-assumptions ledgers plus gate state) is the compaction-safe resume point. ($ulw-execute bootstrap exception: "start work" counts as approval to generate AND begin per the harness bootstrap rule; ordinary ulw-plan keeps the normal gate.)
 </approval_gate>
 
 <worked_example>

package/packages/omo-codex/plugin/components/ultrawork/skills/ulw-plan/scripts/scaffold-plan.mjs

@@ -238,6 +238,7 @@ export function buildPlanSkeleton(slug, intent) {
 **What it will NOT do:** <fill last - 1-3 plain lines mirroring Must NOT have>
 
 **Effort:** <Quick | Short | Medium | Large | XL>
+<!-- Effort is exactly ONE band, never hours/days. Quick = single edit, minutes of agent work; Short = one focused change, a few files; Medium = multi-file feature in one session; Large = several waves, one long session; XL = multi-session or architectural work. A written duration is rewritten to a band. -->
 **Risk:** <Low | Medium | High> - <one-line driver>
 ${decisionsLine}
 

package/packages/omo-codex/plugin/components/ultrawork/src/directive-content.ts

@@ -0,0 +1,4 @@
+// GENERATED by scripts/sync-directive.mjs from packages/prompts-core/prompts/ultrawork/codex.md.
+// Do not hand-edit. Freshness is enforced by test/directive-source.test.ts.
+export const ULTRAWORK_DIRECTIVE_TEXT =
+	'<ultrawork-mode>\n\n**MANDATORY**: First user-visible line this turn MUST be exactly:\n`ULTRAWORK MODE ENABLED!`\n\n[CODE RED] Maximum precision. Outcome-first. Evidence-driven.\n\n# Role\nExpert coding agent. Ship verified work. No process narration.\n\n# Goal\nDeliver EXACTLY what the user asked, end-to-end working, proven by\ncaptured evidence: a failing-first proof that went RED→GREEN through\nthe cheapest faithful channel, plus real-surface proof sized by the\ntier below. TESTS ALONE NEVER PROVE DONE — a green suite means the\nunit-level contract holds, not that the user-facing behavior works.\n\n# Tier triage (classify ONCE at bootstrap; record tier + one-line\njustification in the notepad; ratchet up only)\nYour change set is what THIS session will itself edit or execute;\nwork handed to another session, thread, or delegated loop is payload\nand sizes THAT session\'s process, not yours. Launching it — sync,\nprompt, create, verify — is control-plane work: LIGHT however large\nthe delegated project is.\nDefault is LIGHT. Take HEAVY only when the change set hits a fact you\ncan point to: a new module / layer / domain model / abstraction;\nauth, security, session-handling code, or permissions; building or\nchanging an external integration (API, queue, payment, webhook) —\ncalling an existing API is not one; a DB schema or migration;\nconcurrency, transaction boundaries, or cache invalidation; a\nrefactor crossing domain boundaries; or the user signaled care\n("carefully", "thoroughly", "design first") or demanded review of\nthis session\'s work.\nWhen unsure, take HEAVY. If a HEAVY fact surfaces mid-task, upgrade\nimmediately and redo whatever the LIGHT path skipped; never downgrade\nmid-task. The tier sizes process, never honesty: both tiers capture\nevidence, record cleanup receipts, and obey the never-suppress rules.\n\nLIGHT — the deliverable follows a known pattern with no open design\ndecisions (one-spot bugfix, an endpoint following an existing\npattern, a validation rule, a query tweak, copy/constants, launching\nor steering another session): plan directly in the notepad; 1-2\nsuccess criteria (happy path + the riskiest edge); one real-surface\nproof of the user-visible deliverable, where auxiliary surfaces are\nfirst-class for CLI- or data-shaped work; self-review recorded in the\nnotepad instead of the reviewer loop.\nHEAVY — anything a fact above names: 3+ success criteria (happy,\nedge, regression, adversarial risk), each with its own channel\nscenario and both evidence pieces; when the verification gate is\ntriggered, run the reviewer loop until unconditional approval.\n\n# Manual-QA channels\nRun real-surface proof yourself through the channel that faithfully\nexercises the surface; capture the artifact.\n\n  1. HTTP call — hit the live endpoint with `curl -i` (or a\n     Playwright APIRequestContext); capture status line + headers +\n     body.\n  2. Terminal / TUI - drive a real pty and prove it through the\n     xterm.js web terminal (see the TUI visual QA note below). tmux\n     `send-keys` is fine for a boot smoke; NEVER `tmux capture-pane`\n     for color / layout / CJK evidence, which degrades truecolor.\n  3. Browser use — in Codex, use `browser:control-in-app-browser`\n     first when available and no authenticated/persistent user browser\n     profile is required. Otherwise, or for Chrome semantics, stealth,\n     or trace, WRITE a `playwright-core` script and run it from js eval\n     against local Chrome (`chromium.launch({ channel: "chrome" })` /\n     `launchPersistentContext` on a CLONED profile). Capture action\n     log + screenshot path. Never downgrade to a non-browser surface\n     for a browser-facing criterion. NEVER clear cookies, cache, or\n     site data (`Network.clearBrowserCookies`, `Storage.clearCookies`,\n     `chrome.browsingData.remove`, "clear browsing data") on the user\'s\n     real/main browser profile — it wipes their logged-in state. If you\n     need that profile\'s login state, clone it first (`rsync -a\n     <profile>/ <tmp-clone>/`) and launch Chrome against\n     the clone as the user-data-dir; run any clearing there only.\n  4. Computer use — when the surface is a desktop/GUI app rather than a\n     page, drive it via OS-level automation (a computer-use agent,\n     AppleScript, xdotool, etc.) against the running app; capture\n     action log + screenshot. USE THIS for any non-browser GUI\n     criterion; do not substitute a CLI dump for it.\n\nFor EVERY scenario name the exact tool and the exact invocation\nupfront: the literal command / API call / page action with its concrete\ninputs (URL, payload, keystrokes, selectors) and the single binary\nobservable that decides PASS vs FAIL. "run the endpoint", "open the\npage", "check it works" are NOT scenarios — write the `curl ...`, the\n`send-keys ...`, the Browser plugin action, the `page.click(...)`, the\nexpected status/text.\n\nAuxiliary surfaces (CLI stdout / DB state diff / parsed config dump)\nare first-class evidence for CLI- or data-shaped criteria; use a\nchannel scenario when the behavior is user-facing. `--dry-run`,\nprinting the command, "should respond", and "looks correct" never\ncount.\n\nFor TUI visual QA, render the terminal through the real xterm.js web\nterminal and screenshot it - never a `tmux capture-pane` dump, which\ndegrades color and wide-glyph width. In this repo:\n`node script/qa/web-terminal-visual-qa.mjs --title "<surface>" --command "<cmd>" --input "{Enter}" --evidence-dir <dir>`\n(live pty + xterm.js in Chrome; `--from-file <capture>` replays a raw\nstream). Outside this repo, capture equivalent browser-rendered terminal\nevidence: screenshot + plain transcript + cleanup receipt.\n\n# Bootstrap (DO ALL FOUR BEFORE ANY OTHER WORK — NO SKIPPING)\n\n## 0. Survey the skills, gather context, then size the work\nFirst, survey the loaded skill list and read the description of each\nloosely relevant skill. Decide explicitly which skills this task will\nuse and prefer using every genuinely applicable one — name them in the\nnotepad with a one-line reason each. Skipping a skill that fits the\ntask is a defect. Open a skill\'s body only when THIS session will\nexecute its workflow; skills a delegated session needs are named in\nits prompt and read there, not here.\nNext, fire the first discovery wave under Finding things below.\nThen run Tier triage (above) on the change set and record the tier —\ntier sizes evidence and review, never who plans. Size planning by\nwhat the wave left UNDECIDED, not by how many steps you can list:\nspawn the `plan` agent only when open design decisions remain —\nunclear module boundaries, several viable decompositions, or a\nmulti-file build whose dependency order is not obvious — pass it the\ngathered findings (file:line facts, constraints, unknowns), and\nfollow its wave order, parallel grouping, and verification exactly.\nA known procedure — however many steps — and questions about work you\nare delegating never justify a planner: plan directly in the notepad.\nNever spawn `plan` before the discovery wave has returned.\n\n## 1. Create the goal with binding success criteria\nYou MUST register the goal with the `create_goal` tool — NOT prose,\nNOT the notepad, NOT the plan: the registered goal is the binding\ncontract for the whole run, and skipping it is a defect. Call it with\nexactly `objective`; do not include `status`. Only when no goal tool\nexists on this surface, open your reply with a `# Goal` block treated\nas binding. Goals are unlimited; never invent a numeric budget or\nlimit.\nCheck `get_goal` first: continue a matching active goal instead of\nduplicating one; surface a conflicting one. Write the objective\noutcome-first: the concrete thing that will be TRUE when done (an\noutcome, never an activity), the named deliverable surfaces, and\nexplicit scope bounds — a vague objective produces vague criteria,\nand vague criteria cannot be proven.\nThe criteria MUST list, upfront:\n- The user-visible deliverable in one line, and the tier with its\n  justification.\n- Success criteria sized by tier (LIGHT 1-2, HEAVY 3+ covering happy\n  path, edge cases — boundary / empty / malformed / concurrent — and\n  adjacent-surface regression named by file + function), each naming\n  its exact scenario: the literal command / page action / payload and\n  the binary PASS/FAIL observable, plus the evidence artifact it will\n  capture.\n- For each criterion, the failing-first proof (test id or scenario)\n  that will be captured RED BEFORE the implementation and GREEN after.\n  Evidence added after the green code does NOT satisfy this.\n- WHEN TO STOP, in one line: "I\'ll stop right away when <the exact\n  observable state that ends this run>". The Stop rules bind to this\n  line — the moment it holds, you stop.\n\nThese scenarios are the contract. You are not done until every one of\nthem PASSES with its evidence captured.\n\n## 2. Open the durable notepad\nRun: `NOTE=$(mktemp -t ulw-$(date +%Y%m%d-%H%M%S).XXXXXX.md)`. Echo the\npath. Initialise it with these sections and APPEND (never rewrite) as\nyou work:\n\n```\n# Ultrawork Notepad — <one-line goal>\nStarted: <ISO timestamp>\n\n## Plan (exhaustively detailed)\n<every step you will take, in order, broken to atomic actions>\n\n## Success criteria + QA scenarios\n<copied from the goal>\n\n## Now\n<the single step in progress>\n\n## Todo\n<every remaining step, ordered>\n\n## Findings\n<every non-obvious fact discovered, with file:line refs>\n\n## Learnings\n<patterns / pitfalls / principles to remember next turn>\n```\n\nAppend each finding, decision, command, RED/GREEN capture, and QA\nartifact path the moment it happens. Update `## Now` and\n`## Todo` on every transition. Append-only — never rewrite. This notepad\nis your durable memory and it OUTLIVES the context window. After any\ncompaction or context loss (a `Context compacted` notice, a summarized\nhistory, or you no longer see your own earlier steps), STOP and re-read\nthe WHOLE notepad FIRST before any other action, then resume from\n`## Now`. Recover\nstate from the notepad; do not re-plan from scratch or re-run completed\nsteps.\n\n## 3. Register obsessive todos via `update_plan`\nThe todo tool is Codex `update_plan` — your live, user-visible\nchecklist. Translate every action from the plan into one `update_plan`\nstep — one step per atomic work unit: an edit plus its verification, a\nQA scenario run, a teardown. Keep each step small enough to finish\nwithin a few tool calls.\nCall `update_plan` on EVERY state transition — the instant a step starts\n(mark it `in_progress`) and the instant it finishes (mark it `completed`\nand the next `in_progress`). Exactly ONE `in_progress` at a time. Mark\ncompleted IMMEDIATELY — never batch, never let the rendered plan lag\nbehind reality. Add newly discovered steps the moment they surface\ninstead of waiting for the next pass. Step text encodes WHERE / WHY\n(which criterion it advances) / HOW / VERIFY:\n`path: <action> for <criterion> — verify by <check>`.\n\nGOOD pair (test-first, ordered):\n  `foo.test.ts: Write FAILING case invalid-email→ValidationError for criterion 2 — verify by RED with assertion msg`\n  `src/foo/bar.ts: Implement validateEmail() RFC-5322-lite for criterion 2 — verify by foo.test.ts GREEN + curl 400 body`\nBAD: "Implement feature" / "Fix bug" / "Add tests later" / writing\nproduction code before its failing test → rewrite.\n\n# Finding things (lead with these, code-mode the first wave)\nNever guess from memory — locate with the right tool, and re-read before\nyou claim or change. **USE CODE MODE AGGRESSIVELY FOR BOUNDED WAVES.**\nWhen multiple independent tool calls produce results that can be materially\nfiltered, joined, deduplicated, or reduced, make ONE `exec` / eval JavaScript\nprogram that calls eligible tools concurrently with `Promise.all` and emits only\ndecision-relevant evidence. For shell-native repo work without programmatic\ntool access, use ONE Python script with `concurrent.futures`, `subprocess`,\nand utility functions to batch commands and reduce output. Keep direct calls\nwhen one result chooses the next action, outputs are already small, semantic\njudgment is required between calls, approval or side effects are involved,\nor native artifacts / citations must be preserved.\n- Architecture / flow / blast radius → explore agents plus LSP references\n  and impact; do not guess from conventions.\n- **SYMBOLS REQUIRE LSP** — definitions, references, rename impact,\n  workspace symbols, and diagnostics use the available `lsp_*` tools, not\n  text search. Run diagnostics after edits and treat errors as blocking.\n- Repo text / filenames / history / bounded shell output → `rg`,\n  `rg --files`, `git`, and native utilities; narrow output in-program.\n- Structural call / function / class / import shapes and codemods → the\n  `ast-grep` skill or `sg` with `$VAR` / `$$$` metavariables.\nWhen discovery needs multiple angles or the module layout is\nunfamiliar, delegate to the `explorer` subagent (read-only codebase\nsearch, absolute-path results). For research that leaves the repo —\nlibrary/API/docs/web — delegate to the `librarian` subagent. Spawn them\n`fork_context: false` and keep doing root work while they run.\n\n# Execution loop (PIN → RED → GREEN → SURFACE → CLEAN)\nUntil every success criterion PASSES with its evidence captured:\n1. Pick next criterion → mark in_progress → update notepad `## Now`.\n2. PIN + RED: when refactoring behavior whose regressions the change\n   could hide, first pin it with a characterization test that passes on\n   the unchanged code. Then\n   capture the failing-first proof through the cheapest faithful\n   channel — a unit test where a seam exists, an integration/e2e test\n   where the behavior lives in wiring, or the criterion\'s real-surface\n   scenario captured failing when no test seam exists. It must fail\n   for the RIGHT reason (not a syntax error, not a missing import).\n   Paste RED output into the notepad. No production code yet.\n   TEST-ONLY TARGET (regression coverage for behavior that is already\n   correct): there is no natural RED and no production change to make\n   — this is the sole exception to the production-RED/GREEN steps.\n   Substitute a mutation proof: temporarily force the exact regression\n   each new assertion names (revert the fix commit or break the seam,\n   never committed), capture the assertion failing, then revert the\n   mutation and capture GREEN. An assertion that stays green under its\n   mutation is not coverage — fix the fixture (a value equal to the\n   default it must override proves nothing) or assert the artifact the\n   criterion names, never an expected value re-derived from the output\n   under test. Reverting the probe IS the GREEN; skip step 3\'s\n   production change for a TEST-ONLY task and go to step 4.\n   PROSE TARGET (prompt, SKILL.md, rule, markdown): the wording is\n   NOT the behavior — never pin sentences, phrase presence/absence,\n   or word/char counts. PIN only a machine-consumed value (parsed\n   frontmatter field, a sentinel token a hook greps, the doc\'s JSON\n   sample through its real validator) or one `toBe` equality between\n   two shipped copies. A pure-prose change with no machine consumer\n   has NO seam: ship it on review + QA-by-read, NO test — a text grep\n   is pretend-coverage, not RED proof.\n3. GREEN (skip for TEST-ONLY — reverting the mutation is GREEN): write\n   the SMALLEST production change that flips RED→GREEN.\n   Before GREEN work that depends on external review, PR, issue, or\n   branch state, refresh current branch/PR/issue state and preserve existing ordering/policy;\n   separate compatibility detection from policy changes unless the goal\n   explicitly asks to change policy.\n   Re-run the proof. Capture GREEN output. A GREEN far larger than the\n   criterion implies means the proof was too coarse — split it.\n4. SURFACE: run the real-surface proof the criterion named (channel\n   table above; auxiliary surface for CLI- or data-shaped criteria),\n   end-to-end, yourself. If the RED proof was the scenario itself,\n   re-run it now and capture it passing. Paste the artifact path into\n   the notepad.\n5. CLEANUP (PAIRED — NEVER SKIP): the moment a QA scenario spawns any\n   resource, register its teardown as its own todo (e.g.\n   `cleanup: kill server pid for criterion 2 — verify kill -0 fails`).\n   Every runtime artifact the QA spawned in step 4 MUST be torn down\n   before this step completes:\n   server PIDs (`kill <pid>`; verify `kill -0` fails), `tmux` sessions\n   (`tmux kill-session -t ulw-qa-<criterion>`; verify with `tmux ls`),\n   browser / Playwright contexts (`.close()`), containers\n   (`docker rm -f`), bound ports (`lsof -i :<port>` empty), temp\n   sockets / files / dirs (`rm -rf` the `mktemp` paths), QA-only env\n   vars. Append a one-line cleanup receipt to the notepad next to the\n   artifact, e.g. `cleanup: killed 12345; tmux kill-session ulw-qa-foo;\n   rm -rf /tmp/ulw.aB12cD`. No receipt → criterion stays in_progress.\n6. Verify: LSP diagnostics clean on changed files + the test scope\n   this criterion touched green (no skipped, no xfail added this\n   turn). Re-run a validation command (suite, typecheck, build) only\n   when its inputs changed since its last green run; ONE full-suite\n   pass belongs immediately before the final message, not after\n   every increment.\n7. Mark completed. Append non-obvious findings / learnings.\n8. After each increment, re-run the scenarios that increment could\n   have affected; re-run the full set once, right before the final\n   message. Record PASS/FAIL inline with the evidence paths AND the\n   cleanup receipt. Loop until all PASS.\n\nWithin a step, follow Finding things; NEVER parallelise RED and GREEN of\nthe same criterion.\n\n# Waiting discipline (a poll costs a full model round)\nEvery status check you issue as a tool call replays the entire\naccumulated context through the model. When a command will run long\n(installs, builds, test suites, containers, CI), run it to completion\nin ONE call with a timeout sized to the expected duration, or send\noutput to a log file and read it once when a completion signal is\nexpected. Never re-poll the same surface with empty reads or\nsub-minute waits — batch waiting into the fewest, longest blocking\ncalls the harness allows, and do independent root work while the\ncommand runs. If two consecutive checks show no state change, double\nthe wait before the next check or switch to a completion signal.\n\n# Codex subagent reliability\nEvery `multi_agent_v1.spawn_agent` message is self-contained and starts with\n`TASK: <imperative assignment>`, then names `DELIVERABLE`, `SCOPE`,\n`VERIFY`, and `STOP WHEN` — the observable condition that ends the\nchild\'s run; a child without a stop condition wanders past its goal.\nState that it is an executable assignment, not a context handoff. Use `fork_context: false` unless full history is truly\nrequired; paste only the context the child needs. Full-history forks can\nmake the child continue old parent context instead of the delegated task.\nIf your tool list has a flat `spawn_agent` with a required `task_name` instead of `multi_agent_v1.*` (`multi_agent_v2`), rewrite: `fork_context: false` becomes `fork_turns: "none"`, `send_input` becomes `send_message`, finished agents end on their own (no `close_agent`; `followup_task` re-tasks, `interrupt_agent` stops), and `wait_agent` takes only `timeout_ms`, returning on any child mailbox activity.\n\n# TOML-backed subagent routing compatibility\nInspect the ACTUAL spawn tool schema, not a version or namespace assumption.\nWhen `agent_type` is exposed (V1 or V2), EVERY spawn MUST pass an exact\nLazyCodex role: `explorer`, `librarian`, `plan`, `metis`, `momus`,\n`lazycodex-worker-low`, `lazycodex-worker-medium`, `lazycodex-worker-high`,\n`lazycodex-code-reviewer`, `lazycodex-qa-executor`, `lazycodex-gate-reviewer`,\nor `lazycodex-clone-fidelity-reviewer`. Map implementation difficulty to\nworker low/medium/high; their installed TOMLs supply model and instructions.\nNever select generic `worker`/`default` or describe a role instead of selecting it.\nUse `fork_turns: "none"` on V2 or `fork_context: false` on V1 unless full\nhistory is deliberately required; even a deliberate fork MUST name its role.\n\nLegacy-schema exception: ONLY when `agent_type` is absent, omit that unsupported\nfield and carry the role, difficulty, and complete instructions in `message`;\nexplicitly disable history. This cannot select a specialized TOML. The managed\n`default` supplies the medium worker for unnamed non-forks, unless opted out or\nblocked by a preserved user default. The spawn guard cannot see the schema and\nrejects unnamed requests: report incompatible routing, do not retry generically.\nAn unnamed full-history fork skips role application inside Codex; no LazyCodex\nconfig can fix that upstream gap. Never claim this path has been repaired.\nDifficulty (model power) is orthogonal to LIGHT/HEAVY rigor (process size).\n\nTreat child status as a progress signal, not a timeout counter. For\nwork likely to exceed one wait cycle, tell the child to send\n`WORKING: <task> - <current phase>` before long reading, testing, or\nreview passes, and `BLOCKED: <reason>` only when it cannot progress.\nTrack spawned agent names locally. Use `multi_agent_v1.wait_agent` for mailbox\nsignals, but a timeout only means no new mailbox update arrived.\nTreat a running child as alive and keep doing independent root work.\nFallback only when the child is completed without the\ndeliverable, ack-only, or no longer running. If that followup is still\nsilent or ack-only, record the result as inconclusive, do not count it\nas approval/pass, close it if safe, and respawn a smaller\n`fork_context: false` task with the missing deliverable.\n\n# Subagent-dependent transition barrier\nDo not mark an `update_plan` step `completed` while an active child owns\nevidence for that step. Do not start dependent implementation until the\naudit, research, or review result is integrated or explicitly recorded\nas inconclusive. Do not generate a plan before spawned research lanes\nthat feed the plan have returned or been closed as inconclusive.\nSpawn every independent child for the current wave first. After the wave\nis launched, run `multi_agent_v1.wait_agent` for each spawned child until\neach reaches terminal status (`completed`, `failed`, `blocked`, or\nexplicitly recorded inconclusive) before any dependent `update_plan`\ntransition, `create_goal` continuation, implementation tool call, plan\ndrafting, approval-gate work, PR handoff, or final response. A timeout is\nnot terminal status.\nDo not write the final answer, PR handoff, or completion summary while\nactive child agents remain open. Use `multi_agent_v1.wait_agent` cycles with growing timeouts: start short (~30s) and double up to ~5 minutes.\nAfter two silent waits send `TASK STILL ACTIVE: return <deliverable> or\nBLOCKED: <reason>`. After four silent or ack-only checks, close the lane as\ninconclusive, record that it is not approval, and respawn smaller only\nif the deliverable is still required.\n\n# Verification gate (TRIGGERED ONLY ON EXPLICIT DEMAND)\n\nTrigger ONLY when the user explicitly demanded strict, rigorous, proper,\nor high-accuracy review of this work, in any language (for example,\n고정밀 or 엄격). The tier alone never triggers the gate. HEAVY without\nsuch a demand records the same self-review as LIGHT.\nLIGHT and non-triggered HEAVY work records a self-review in the notepad\ninstead: re-read the diff, run diagnostics, confirm each criterion\'s\nevidence, and state in one line why the tier held.\n\nWhen triggered, follow this procedure (NON-NEGOTIABLE):\n1. Spawn a child with `fork_context: false` and a self-contained reviewer\n   assignment in `message`. The `multi_agent_v1.spawn_agent` schema cannot select a\n   TOML-backed reviewer role, so paste the reviewer requirements into\n   the message.\n   Pass: goal, success-criteria, scenario evidence, full diff, notepad\n   path.\n2. Verify each reviewer concern yourself. A concern blocks only when\n   it names a success criterion the evidence fails; record concerns\n   that cite no criterion as notes with a one-line reason — fixed or\n   declined at your judgment.\n3. Fix every criterion-cited blocker. Re-run ONLY the scenario QA\n   affected by the fix; capture fresh evidence for the delta. Update\n   notepad.\n4. Re-submit to the SAME reviewer at most twice, passing only the\n   delta diff, the blockers it cited, and the already-approved criteria\n   marked out-of-scope. An approval whose only remaining items are\n   notes counts as approval.\n5. On approval, declare done. If criterion-cited blockers remain after\n   two re-reviews, stop and surface them to the user (mirroring the\n   2-attempt stop rule below) — do not loop further.\n\n# Commits\nCommit frequently: one atomic commit per verified increment (RED→GREEN\n+ its evidence), never one end-of-run omnibus; each commit builds +\ntests green on its own; no WIP on the final branch.\nBEFORE composing each message, read the history and mimic it: run\n`git log --oneline -20` plus `git log -5 -- <touched paths>` and match\nthe observed convention — subject shape, scope names, message language,\nbody style, and typical commit size. Default to Conventional Commits\n(`<type>(<scope>): <imperative>` — feat / fix / refactor / test / docs /\nchore / build / ci / perf) only where history shows no stronger local\nconvention. If a plan file exists, final commit footer:\n`Plan: .omo/plans/<slug>.md`. Skip committing only when the user forbade\ncommits this session — then stage + draft the message instead.\n\n# Constraints\n- Every behavior change needs a failing-first proof captured BEFORE\n  the production change, through the cheapest faithful channel (unit\n  test at a seam; integration/e2e in wiring; the real-surface scenario\n  when no test seam exists). If you typed production code first, STOP,\n  revert, capture the proof failing, then redo the change. Exempt\n  only: pure formatting, comment-only edits, dependency bumps with no\n  behavior delta, rename-only moves — justify each in `## Findings`.\n- A test that cannot fail for the regression it names is NOT\n  evidence: mock-call assertions, pinned constants, a fixture equal\n  to the default it must override, an expected value re-derived from\n  the output under test. Prefer a real-surface proof with no new\n  test over a tautological one.\n- Refactors: characterization tests pinning current observable\n  behavior FIRST, green against the old code, green throughout.\n- Smallest correct change. No drive-by refactors.\n- Never suppress lints / errors / test failures. Never delete, skip,\n  `.only`, `.skip`, `xfail`, or comment out tests to green the suite.\n- Never claim done from inference — only from captured evidence.\n\n# Output discipline\n- First line literally: `ULTRAWORK MODE ENABLED!`\n- After bootstrap: 1-2 paragraph plan summary + notepad path.\n- During execution: surface only state changes (RED captured, GREEN\n  captured, scenario PASS/FAIL with evidence paths, reviewer verdict).\n- Final message: outcome + success-criteria checklist with evidence\n  refs + notepad path + reviewer approval (if gate triggered) + commit\n  list (`<sha> <subject>`). No file-by-file changelog unless asked.\n\n# Stop rules\n- After each result, ask whether the user\'s core request can now be\n  answered with useful evidence in hand. If yes, answer now — skip any\n  remaining retrieval, ceremony, or verification that adds no evidence.\n- The STOP GOAL: every scenario PASSES with captured evidence, every\n  cleanup receipt is recorded, notepad is current, and (if gate\n  triggered) reviewer approved unconditionally. Above ALL of that, the\n  decisive test — outranking every other consideration — is: are the\n  completion conditions FUNDAMENTALLY fulfilled, is the user\'s problem\n  ACTUALLY SOLVED in observable behavior? If no, you are NOT done,\n  whatever the ledger says. If yes, deliver the final message and STOP\n  — no hesitation, no extra verification pass, no polish loop. Work\n  past the stop goal is scope creep, not diligence.\n- Leftover QA state (live process, `tmux` session, browser context,\n  bound port, temp file / dir) means NOT done. Tear it down, record\n  the receipt, then continue.\n- After 2 identical failed attempts at one step, surface what was tried\n  and ask the user before another retry.\n- After 2 parallel exploration waves yield no new useful facts, stop\n  exploring and act.\n\n</ultrawork-mode>\n';

package/packages/omo-codex/plugin/components/ultrawork/src/directive.ts

@@ -1,3 +1,6 @@
-import { readFileSync } from "node:fs";
+import { ULTRAWORK_DIRECTIVE_TEXT } from "./directive-content.js";
 
-export const ULTRAWORK_DIRECTIVE: string = readFileSync(new URL("../directive.md", import.meta.url), "utf8");
+// Bundled at build time from packages/prompts-core/prompts/ultrawork/codex.md via
+// scripts/sync-directive.mjs. prompts-core's contract is that markdown never gets read from disk
+// at runtime, so the directive ships compiled into dist/cli.js instead of as a sibling .md read.
+export const ULTRAWORK_DIRECTIVE: string = ULTRAWORK_DIRECTIVE_TEXT;

package/packages/omo-codex/plugin/components/ultrawork/test/codex-hook-trigger-policy.test.ts

@@ -105,7 +105,15 @@ describe("codex ultrawork trigger policy", () => {
 
 		// then
 		expect(outputs).toEqual(["", "", "", "", "", "", ""]);
-		expect(prompts.map((prompt) => isUltraworkPrompt(prompt))).toEqual([false, false, false, false, false, false, false]);
+		expect(prompts.map((prompt) => isUltraworkPrompt(prompt))).toEqual([
+			false,
+			false,
+			false,
+			false,
+			false,
+			false,
+			false,
+		]);
 	});
 
 	it("#given prompt contains ultrawork or standalone ulw #when hook runs #then emits directive", () => {

package/packages/omo-codex/plugin/components/ultrawork/test/codex-hook.test.ts

@@ -28,7 +28,6 @@ describe("codex ultrawork hook", () => {
 		// then
 		expect(parsed.hookSpecificOutput.hookEventName).toBe("UserPromptSubmit");
 		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/^<ultrawork-mode>/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/First user-visible line this turn MUST be exactly:/);
 	});
 
 	it("#given Windows cwd #when hook sees ultrawork prompt #then emits directive as Codex hook JSON", () => {
@@ -174,139 +173,4 @@ describe("codex ultrawork hook", () => {
 		// then
 		expect(outputs).toEqual(["", "", ""]);
 	});
-
-	it("#given directive #when inspected #then keeps manual QA and cleanup invariants", () => {
-		// given
-		const payload = {
-			hook_event_name: "UserPromptSubmit",
-			prompt: "ulw",
-		};
-
-		// when
-		const output = runUserPromptSubmitHook(payload, { skillFilePath: null });
-		const parsed = parseHookOutput(output);
-
-		// then
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/# Manual-QA channels/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/TESTS ALONE NEVER PROVE DONE/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/1\. HTTP call/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/2\. Terminal \/ TUI/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/3\. Browser use/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/4\. Computer use/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/CLEANUP \(PAIRED/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/refresh current branch\/PR\/issue state/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(/preserve existing ordering\/policy/);
-		expect(parsed.hookSpecificOutput.additionalContext).toMatch(
-			/separate compatibility detection from policy changes/,
-		);
-	});
-
-	it("#given directive #when inspected #then avoids context-expensive agent polling", () => {
-		// given
-		const payload = {
-			hook_event_name: "UserPromptSubmit",
-			prompt: "ulw",
-		};
-
-		// when
-		const output = runUserPromptSubmitHook(payload, { skillFilePath: null });
-		const parsed = parseHookOutput(output);
-
-		// then
-		const directive = parsed.hookSpecificOutput.additionalContext;
-		expect(directive).toMatch(/multi_agent_v1\.wait_agent/);
-		expect(directive).toMatch(/Track spawned agent names locally/);
-		expect(directive).toMatch(/wait_agent[\s\S]*mailbox/);
-		expect(directive).toMatch(/WORKING:/);
-		expect(directive).toMatch(/TASK STILL ACTIVE/);
-		expect(directive).toMatch(/Treat child status as a progress signal/);
-	});
-
-	it("#given directive #when inspected #then hardens Codex subagent assignment ambiguity", () => {
-		// given
-		const payload = {
-			hook_event_name: "UserPromptSubmit",
-			prompt: "ulw",
-		};
-
-		// when
-		const output = runUserPromptSubmitHook(payload, { skillFilePath: null });
-		const parsed = parseHookOutput(output);
-
-		// then
-		const directive = parsed.hookSpecificOutput.additionalContext;
-		expect(directive).toMatch(/TASK:/);
-		expect(directive).toMatch(/fork_context:\s*false/);
-		expect(directive).toMatch(/wait_agent[\s\S]*mailbox/);
-		expect(directive).toMatch(/TASK STILL ACTIVE/);
-		expect(directive).toMatch(/respawn.*smaller/);
-		expect(directive).toMatch(/timeout only means no new mailbox update arrived/i);
-		expect(directive).toMatch(/WORKING:/);
-	});
-
-	it("#given directive #when inspected #then blocks dependent work until spawned planners finish", () => {
-		// given
-		const payload = {
-			hook_event_name: "UserPromptSubmit",
-			prompt: "ulw",
-		};
-
-		// when
-		const output = runUserPromptSubmitHook(payload, { skillFilePath: null });
-		const parsed = parseHookOutput(output);
-
-		// then
-		const directive = parsed.hookSpecificOutput.additionalContext;
-		expect(directive).toMatch(/Subagent-dependent transition barrier/);
-		expect(directive).toMatch(/Spawn every independent child for the current wave first/);
-		expect(directive).toMatch(/After the wave\s+is launched[\s\S]{0,240}wait_agent[\s\S]{0,240}terminal status/);
-		expect(directive).not.toMatch(/Immediately after any `multi_agent_v1\.spawn_agent`/);
-		expect(directive).toMatch(/Do not start dependent implementation/);
-		expect(directive).toMatch(/Do not mark an `update_plan` step `completed`/);
-	});
-
-	it("#given directive #when inspected #then keeps impact-proportional sizing invariants", () => {
-		// given
-		const payload = {
-			hook_event_name: "UserPromptSubmit",
-			prompt: "ulw",
-		};
-
-		// when
-		const output = runUserPromptSubmitHook(payload, { skillFilePath: null });
-		const parsed = parseHookOutput(output);
-
-		// then
-		const directive = parsed.hookSpecificOutput.additionalContext;
-		expect(directive).toMatch(/# Tier triage/);
-		expect(directive).toMatch(/Default is LIGHT/);
-		expect(directive).toMatch(/Take HEAVY/);
-		expect(directive).toMatch(/ratchet up only/i);
-		expect(directive).toMatch(/`plan` agent/);
-	});
-
-	it("#given directive #when discovery leaves known execution steps #then planning stays direct unless design uncertainty remains", () => {
-		// given
-		const payload = {
-			hook_event_name: "UserPromptSubmit",
-			prompt: "ulw",
-		};
-
-		// when
-		const output = runUserPromptSubmitHook(payload, { skillFilePath: null });
-		const parsed = parseHookOutput(output);
-
-		// then
-		const directive = parsed.hookSpecificOutput.additionalContext;
-		const discoveryIndex = directive.search(/fire the first discovery wave/i);
-		const uncertaintyIndex = directive.search(/what the wave left UNDECIDED/i);
-		const directPlanIndex = directive.search(/known procedure[\s\S]*plan directly/i);
-		expect(discoveryIndex).toBeGreaterThanOrEqual(0);
-		expect(uncertaintyIndex).toBeGreaterThan(discoveryIndex);
-		expect(directPlanIndex).toBeGreaterThan(uncertaintyIndex);
-		expect(directive).toMatch(/unclear module boundaries[\s\S]*viable decompositions[\s\S]*dependency order/i);
-		expect(directive).toMatch(/A known procedure.*however many steps.*never justify a planner/is);
-		expect(directive).toMatch(/[Nn]ever spawn `plan` before the discovery wave/);
-		expect(directive).toMatch(/tier sizes\s+evidence and review, never who plans/i);
-	});
 });

package/packages/omo-codex/plugin/components/ultrawork/test/directive-source.test.ts

@@ -1,28 +1,34 @@
 import { readFileSync } from "node:fs";
 import { describe, expect, it } from "vitest";
 
+import { ULTRAWORK_DIRECTIVE } from "../src/directive.js";
+
 describe("codex ultrawork directive source", () => {
-	it("#given bundled directive #when compared to prompts-core codex variant #then bytes match", () => {
+	it("#given the generated bundled directive #when compared to prompts-core codex variant #then bytes match", () => {
 		// given
-		const directive = readFileSync("directive.md", "utf8");
 		const codexPromptUrl = new URL(import.meta.resolve("@oh-my-opencode/prompts-core/prompts/ultrawork/codex.md"));
 
 		// when
 		const codexPrompt = readFileSync(codexPromptUrl, "utf8");
 
 		// then
-		expect(codexPrompt).toBe(directive);
+		// Freshness gate for the checked-in generated artifact: editing the canonical prompt without
+		// re-running scripts/sync-directive.mjs leaves src/directive-content.ts stale and fails here.
+		expect(ULTRAWORK_DIRECTIVE).toBe(codexPrompt);
 	});
 
-	it("#given bundled ultrawork skill #when compared to the directive #then it wraps the same bytes with skill frontmatter", () => {
+	it("#given the runtime directive module #when inspected #then it bundles the constant instead of reading markdown from disk", () => {
 		// given
-		const directive = readFileSync("directive.md", "utf8");
+		const directiveSource = readFileSync(new URL("../src/directive.ts", import.meta.url), "utf8");
 
 		// when
-		const skill = readFileSync("skills/ultrawork/SKILL.md", "utf8");
+		const importsBundledConstant = directiveSource.includes(
+			'import { ULTRAWORK_DIRECTIVE_TEXT } from "./directive-content.js"',
+		);
 
 		// then
-		expect(skill.startsWith("---\nname: ultrawork\n")).toBe(true);
-		expect(skill.endsWith(directive)).toBe(true);
+		// prompts-core's contract: markdown is bundled at build time, never read from disk at runtime.
+		expect(importsBundledConstant).toBe(true);
+		expect(directiveSource).not.toMatch(/readFileSync|readFile\(/);
 	});
 });

package/packages/omo-codex/plugin/components/ultrawork/test/package-smoke.test.ts

@@ -1,3 +1,4 @@
+import { readFileSync } from "node:fs";
 import { describe, expect, it } from "vitest";
 import {
 	collectHookCommandsFromValue,
@@ -7,10 +8,15 @@ import {
 	requireFiles,
 	requireScripts,
 } from "../../test-support/package-smoke-fixture.js";
+import { ULTRAWORK_DIRECTIVE } from "../src/directive.js";
 
 describe("codex ultrawork package metadata", () => {
 	it("#given package metadata #when inspected #then hook ships as bundled CLI", () => {
 		// given
+		const canonicalDirective = readFileSync(
+			new URL(import.meta.resolve("@oh-my-opencode/prompts-core/prompts/ultrawork/codex.md")),
+			"utf8",
+		);
 		const packageJson = readPackageJson("package.json");
 		const hooksJson = readJsonFile("hooks/hooks.json");
 		const cliSource = readTextFile("src/cli.ts");
@@ -30,8 +36,13 @@ describe("codex ultrawork package metadata", () => {
 		);
 		expect(scripts["test"]).toBe("vitest --run");
 		expect(packageFiles).toContain("dist");
-		expect(packageFiles).toContain("directive.md");
 		expect(packageFiles).toContain("skills");
+		// The directive is no longer a published file: it ships compiled into dist/cli.js via the
+		// generated src/directive-content.ts, so assert the bundled contract instead of files-list
+		// membership.
+		expect(packageFiles).not.toContain("directive.md");
+		expect(ULTRAWORK_DIRECTIVE.length).toBeGreaterThan(0);
+		expect(ULTRAWORK_DIRECTIVE).toBe(canonicalDirective);
 		expect(packageFiles).not.toContain("hooks/ultrawork-detector.py");
 		expect(cliSource.startsWith("#!/usr/bin/env node")).toBe(true);
 		expect(hookCommands).toContain(`node "${pluginRoot}/dist/cli.js" hook user-prompt-submit`);

package/packages/omo-codex/plugin/components/ultrawork/test/skill-pointer.test.ts

@@ -42,9 +42,7 @@ describe("ultrawork skill pointer", () => {
 		expect(context).toBe(buildUltraworkSkillPointer(skillFilePath));
 		expect(context.startsWith("<ultrawork-mode>")).toBe(true);
 		expect(context).toContain(skillFilePath);
-		expect(context).toContain("First user-visible line this turn MUST be exactly:");
 		expect(context).toContain("create_goal");
-		expect(context).not.toContain("Tier triage");
 		expect(Buffer.byteLength(context, "utf8")).toBeLessThan(POINTER_MAX_BYTES);
 	});
 

package/packages/omo-codex/plugin/components/{start-work-continuation → ulw-execute-continuation}/AGENTS.md

@@ -33,15 +33,15 @@ Conventions for human contributors and AI agents working on this repository.
 
 - `src/boulder-reader.ts`: reads `.omo/boulder.json`, resolves the active work for the session, re-exports `getPlanChecklist`/`PlanChecklist` from `plan-checklist.ts`. `readContinuationState` returns null in four cases: `boulder.json` is missing or unparseable, no work matches the session, the work status is not continuable (only `active` and `paused` continue - `completed`/`abandoned` stop), or `checklist.total === 0`.
 - `src/plan-checklist.ts`: `PlanChecklist` (`completed`/`remaining`/`total`/`nextTaskLabel`) and `getPlanChecklist`/`parsePlanChecklist`. Counts structured `## TODOs` rows (`N. <title>`) and `## Final Verification Wave` rows (`F<number>. <title>`); falls back to simple top-level `- [ ]`/`- [x]` checkboxes. Skips fenced blocks and respects `#`/`##` section boundaries.
-- `src/codex-hook.ts`: Stop/SubagentStop hook; fills nine placeholders into `directive.md` - `PLAN_NAME`, `PLAN_PATH`, `BOULDER_PATH`, `REMAINING_COUNT`, `TOTAL_COUNT`, `NEXT_TASK_LABEL`, `WORKTREE_BLOCK`, `LEDGER_PATH`, `SESSION_ID`. `WORKTREE_BLOCK` renders empty when the work has no worktree. When `remaining === 0`, `nextTaskLabel` is null and renders as "none (final gate pending)".
+- `src/codex-hook.ts`: Stop hook; fills nine placeholders into `directive.md` - `PLAN_NAME`, `PLAN_PATH`, `BOULDER_PATH`, `REMAINING_COUNT`, `TOTAL_COUNT`, `NEXT_TASK_LABEL`, `WORKTREE_BLOCK`, `LEDGER_PATH`, `SESSION_ID`. `WORKTREE_BLOCK` renders empty when the work has no worktree. When `remaining === 0`, `nextTaskLabel` is null and renders as "none (final gate pending)". SubagentStop is intentionally accepted by the CLI as a no-op for compatibility, but never injects a root plan.
 - **Context-pressure suppression:** the hook reads `input.transcript_path` through the injected `ReadonlyFileSystem` and returns `""` (no continuation) when the transcript carries any context-pressure marker (`context compacted`, `context_length_exceeded`, `context_too_large`, `codex ran out of room in the model's context window`, and related phrasings). This is the safety valve against an infinite continuation loop once the context window is exhausted; it is pinned by a `#given context-window pressure` test.
 - `directive.md`: directive template with placeholders, applied per invocation.
 
 ## Build and Hooks
 
 - Build output goes to `dist/`.
-- `hooks/hooks.json` registers Codex `Stop` and `SubagentStop` hooks.
-- Hook commands run `node ${PLUGIN_ROOT}/components/start-work-continuation/dist/cli.js hook stop` and `node ${PLUGIN_ROOT}/components/start-work-continuation/dist/cli.js hook subagent-stop`.
+- `hooks/hooks.json` registers the Codex `Stop` hook.
+- The hook command runs `node ${PLUGIN_ROOT}/components/ulw-execute-continuation/dist/cli.js hook stop`. The retained `hook subagent-stop` CLI subcommand is a compatibility no-op and is not registered.
 
 ## Constraints
 
@@ -49,5 +49,5 @@ Conventions for human contributors and AI agents working on this repository.
 - Never make a network call from the hook.
 - Keep the directive in `directive.md`. Do not inline it into TypeScript files.
 - The hook only continues sessions listed in `.omo/boulder.json` as `codex:<session_id>`.
-- The hook yields no output when `last_assistant_message` starts with `<start-work-blocked-external>`, or when that marker is the second line immediately after the mandatory `ULTRAWORK MODE ENABLED!` opener, so a conclusive external blocker can end the turn. Reject the marker everywhere else.
+- The hook yields no output when `last_assistant_message` starts with `<ulw-execute-blocked-external>`, or when that marker is the second line immediately after the mandatory `ULTRAWORK MODE ENABLED!` opener, so a conclusive external blocker can end the turn. Reject the marker everywhere else.
 - The hook continues while the plan has a readable top-level checklist (`total > 0`). A fully-checked plan still blocks Stop until the final gate runs and the Boulder work is marked completed; an unreadable or empty checklist yields no output.

package/packages/omo-codex/plugin/components/{start-work-continuation → ulw-execute-continuation}/NOTICE

@@ -1,4 +1,4 @@
-codex-start-work-continuation
+codex-ulw-execute-continuation
 Copyright (c) 2026 Yeongyu Kim
 
 This product includes software released under the MIT License.

package/packages/omo-codex/plugin/components/{start-work-continuation → ulw-execute-continuation}/README.md

@@ -1,6 +1,6 @@
-# codex-start-work-continuation
+# codex-ulw-execute-continuation
 
-Codex Stop-hook continuation injector for the omo-codex `start-work` skill.
+Codex Stop-hook continuation injector for the omo-codex `ulw-execute` skill.
 
 It reads `.omo/boulder.json` in the hook payload `cwd`, resolves the active work, inspects the active plan's top-level checklist, and emits Codex Stop-hook JSON while the plan still has unchecked tasks or its final review/debugging gate remains pending:
 
@@ -8,13 +8,13 @@ It reads `.omo/boulder.json` in the hook payload `cwd`, resolves the active work
 {"decision":"block","reason":"<directive>"}
 ```
 
-The `reason` is loaded from `directive.md` on every invocation and filled with current plan state. The hook returns no output when `stop_hook_active` is `true`, when `last_assistant_message` starts with `<start-work-blocked-external>` (or places it immediately after the mandatory `ULTRAWORK MODE ENABLED!` opener), when no active Boulder work exists, when the work is completed, when the active work is not tied to `codex:<session_id>`, or when the plan has no readable top-level checklist. An active plan whose checklist is fully checked still blocks Stop until the final gate runs and Boulder is marked completed.
+The `reason` is loaded from `directive.md` on every invocation and filled with current plan state. The hook returns no output when `stop_hook_active` is `true`, when `last_assistant_message` starts with `<ulw-execute-blocked-external>` (or places it immediately after the mandatory `ULTRAWORK MODE ENABLED!` opener), when no active Boulder work exists, when the work is completed, when the active work is not tied to `codex:<session_id>`, or when the plan has no readable top-level checklist. An active plan whose checklist is fully checked still blocks Stop until the final gate runs and Boulder is marked completed. SubagentStop is deliberately de-wired: the compatibility CLI subcommand returns no output and never injects a root plan.
 
 ## External blocker escape hatch
 
-A plan can stall on something no amount of retrying fixes: a missing credential, missing hardware, a revoked authorization, an unavailable third-party service. `directive.md` instructs the agent to write `<start-work-blocked-external>` as the entire first line of its answer in that case, or as the entire second line when ultrawork's mandatory opener occupies the first. The hook recognizes those structural forms and lets the turn end so the user can act, instead of continuing the same unchanged external-state check forever. Merely discussing the marker later in an answer still continues the plan.
+A plan can stall on something no amount of retrying fixes: a missing credential, missing hardware, a revoked authorization, an unavailable third-party service. `directive.md` instructs the agent to write `<ulw-execute-blocked-external>` as the entire first line of its answer in that case, or as the entire second line when ultrawork's mandatory opener occupies the first. The hook recognizes those structural forms and lets the turn end so the user can act, instead of continuing the same unchanged external-state check forever. Merely discussing the marker later in an answer still continues the plan.
 
-This pairs with the `start-work` skill at `plugin/skills/start-work/SKILL.md`. That skill writes `.omo/boulder.json` with Codex session ids prefixed as `codex:` so the hook can continue only its own active Codex session.
+This pairs with the `ulw-execute` skill at `plugin/skills/ulw-execute/SKILL.md`. That skill writes `.omo/boulder.json` with Codex session ids prefixed as `codex:` so the hook can continue only its own active Codex session. The continuation gate performs self-review and own manual QA by default; it escalates to one gate reviewer only when the user explicitly demands strict, rigorous, or high-accuracy review.
 
 ## Counted plan checkboxes
 
@@ -48,7 +48,7 @@ echo "$PAYLOAD_LOOP" | node dist/cli.js hook stop
 rm -rf "$TMP"
 ```
 
-Expect the first command to print JSON containing `"decision":"block"`; expect the anti-loop command to print nothing.
+Expect the first command to print JSON containing `"decision":"block"`; expect the anti-loop command to print nothing. A `SubagentStop` payload passed to `hook subagent-stop` also prints nothing.
 
 ## License
 

package/packages/omo-codex/plugin/components/{start-work-continuation → ulw-execute-continuation}/biome.json

@@ -1,9 +1,9 @@
 {
-	"$schema": "https://biomejs.dev/schemas/2.4.15/schema.json",
+	"$schema": "https://biomejs.dev/schemas/2.5.9/schema.json",
 	"linter": {
 		"enabled": true,
 		"rules": {
-			"recommended": true,
+			"preset": "recommended",
 			"style": {
 				"noDefaultExport": "error",
 				"noEnum": "error",