--- name: ultragoal-goal description: "Ultragoal for Codex. Use when the user asks for ultragoal in Codex, a hook-backed or runner-backed Codex goal loop, a rubric-backed Codex /goal, durable goal execution, evidence-led completion, panel verification, or to turn a brain dump into a Codex-native goal. Do not use for trivial one-turn tasks unless explicitly invoked." --- # Ultragoal Goal for Codex Use this skill to convert an open-ended request into a file-backed Ultragoal spec, audit the rubric, then attach Codex native Goal mode to that spec. The Codex plugin bundles a Codex Stop hook gate, SessionStart context/bootstrap hooks, and a SubagentStop evidence gate for the `ultragoal-executor` role. Stop-hook blocking is verified working on codex-cli 0.142.x **in trusted project roots** — the gate keeps Codex working until the rubric and verifier/panel verdict hold. In untrusted directories Codex downgrades blocking to advisory, and older builds may too; there, `npx ultragoal run --codex --headless ""` is the supported enforcement fallback: it starts Codex with the same file-backed contract, bypasses hook trust for that vetted automation run, inspects the active goal after each turn, and resumes with the remaining rubric/verifier work until the file is marked done or the runner cap is reached. Known limitation (unlike the Claude gate): Codex goals are not scoped per session — the gate enforces any active goal in the repo whose `session`/`codex_thread` field is empty, for every Codex session there. Run one active Codex goal per repo at a time; a second concurrent goal needs its own git worktree. ## Inputs - User brief or brain dump. - Repository guidance (`AGENTS.md`) and any existing `.ultragoal/memory/` files. - Current Codex goal state, if `get_goal` is available. - Codex hook state if the user reports hook warnings; ask them to open `/hooks` only when trust is blocking the bundled gate. - Mode hint from `npx ultragoal run --codex`: `MODE: interactive-interview` or `MODE: headless-autonomous`. ## First Reads After deciding this skill applies, read: 1. `references/rubric-guide.md` 2. `references/qa-capability-map.md` If present in the repo, also read `.ultragoal/memory/MEMORY.md` and any active `.ultragoal/goals/**/goal.md` before drafting. ## Workflow 1. **Triage** - If the task is trivial and the user did not explicitly ask for a goal, do it directly instead of creating ceremony. - If another Codex goal is active and its objective differs, stop and ask whether to pause/clear it or only draft a spec. Do not call `create_goal` over an unrelated active goal. 2. **Ground before asking** - Always consult memory and repo context before asking: read project instructions, `.ultragoal/config.md` if present, `.ultragoal/memory/MEMORY.md` if present, and enough relevant source/docs to avoid asking discoverable questions. - Ask only questions whose answer changes what gets built, the rubric, the verification rung, or a safety/scope boundary. Use a recommended default when the repo and brief support one. - Size the interview by ambiguity, stakes, and run length: clear reversible goals can get one short batch or no questions; vague, long, risky, release-grade, or user-specified "deep/max" goals need a deeper interview before the draft is armed. 3. **Mode-specific interview** - **Interactive interview mode** (`MODE: interactive-interview`, default for `npx ultragoal run --codex ""`): act like Claude `/ultragoal:goal`. Ask decision-shaped questions with the recommended default first; include a one-line reason/tradeoff for each option. Prefer one concise batch, but use additional rounds for deep/high-stakes goals. After drafting, provide a recap before arming: objective, key decisions, assumptions/defaults you made, non-goals, verification plan, and budget/rigor choice. Then ask one standalone **Arm goal** question and wait for explicit user confirmation before setting `status: active`, `codex_goal: active`, or calling `create_goal`. - **Headless autonomous mode** (`MODE: headless-autonomous`, used by `npx ultragoal run --codex --headless ""`): do not ask the user questions. Make conservative assumptions from the brief, repo, and memory; record explicit defaults in the goal file Context/Verification log so the user can audit what was assumed; then arm and execute. Headless must never wait for an arm prompt. 4. **Write the draft** - Prefer `.ultragoal/goals/active//goal.md` for new hook-backed goals so Claude and Codex use the same canonical active-goal shape. - If the repo already uses `.ultragoal/codex-goals//goal.md`, continuing that legacy path is acceptable; the Codex gate enforces both locations. - Use `status: draft` until the user explicitly asks to start/arm/run it. - Keep the draft readable by other harnesses. Required shape: ```markdown --- slug: type: codex-goal status: draft created: codex_goal: pending verify: on|panel budget: --- # Objective # Context # Rubric - [ ] - check: `` - [ ] - check: BROWSER or SIMULATOR or MANUAL DEVICE STEP - - [ ] VERIFIER: independent sign-off recorded in the Verification log # Stop conditions - # Constraints - - check: `` # Verification log ``` 5. **Audit the draft** Run: ```bash node /scripts/rubric-audit.mjs .ultragoal/codex-goals//goal.md ``` Resolve `` to this skill directory. A `BLOCKER` is a draft defect; revise and rerun before arming. A `WARN` must be fixed or recorded as an intentional tradeoff in Context. 6. **Recap before arming** - Summarize the objective, key decisions, non-goals, proof strategy, and remaining warnings. - In interactive interview mode, ask a standalone **Arm goal** question and wait for explicit user confirmation before arming. - In headless autonomous mode, record the defaults/assumptions that replaced user answers and arm without asking. 7. **Arm Codex native Goal mode** - If `create_goal` is available, call it with a compact objective under 4,000 characters. Do not set a token budget unless the user explicitly requested one. - After arming, update the draft frontmatter to `status: active` and `codex_goal: active`. - If the user requested "max rigor" or "panel verification", set `verify: panel`; completion then requires all three fresh-context lenses: `checks`, `refute`, and `constraints`. - Objective template: ```text Work from .ultragoal/codex-goals//goal.md. Complete every unchecked rubric item with recorded evidence, use the named real-surface QA channels, keep the file updated, and do not mark the goal complete until the final verifier item passes. ``` - If goal tools are unavailable, tell the user to run: ```text /goal Work from .ultragoal/codex-goals//goal.md until every rubric item has evidence and the verifier item passes. ``` 8. **Execute** - Keep `update_plan` synchronized with the next few rubric items. - In substantial goals, keep the root Codex thread as the orchestrator: planning, updating the goal file, integrating results, and deciding what proof is still missing. Delegate implementation subtasks to `ultragoal-executor` when Codex subagent tools are available. Give each executor prompt `TASK`, `DELIVERABLE`, `SCOPE`, and `VERIFY` sections. - Every `ultragoal-executor` completion must create one non-empty receipt under `.ultragoal/evidence/` and end its final response with `ULTRAGOAL_EVIDENCE_RECORDED: `. The bundled SubagentStop hook blocks missing, empty, symlinked, or out-of-tree receipts for a small retry budget. - Use browser, simulator, Computer Use, curl, tmux, or CLI evidence according to the QA map. Tests are supporting evidence, not proof of a UI or external-service claim. - After each passed item, update the checkbox and add one evidence line under it or in `# Verification log`. - Use `ultragoal-verifier` for independent verification when Codex subagent tools are available. If `verify: panel`, dispatch or run three independent verifier passes at the end. Each must append exactly `ULTRAGOAL-VERIFIED: PASS rubric= lens=` to the goal file, bound to the current rubric hash. - If Codex prints a hook-trust warning, tell the user to run `/hooks` and trust the Ultragoal bundled hooks, or rerun through `npx ultragoal run --codex --headless ""`. - Do not call `update_goal(status="complete")` until every rubric item, constraints check, and final verifier item is complete. - Use `update_goal(status="blocked")` only when the same blocking condition has repeated for the required Codex goal turns and no meaningful progress is possible. 9. **Resume** - On continuation or after compaction, read the goal file and `get_goal` before doing anything else. - Continue from unchecked rubric items; do not re-plan from scratch unless the file is incoherent or the user changes scope. ## Codex-Specific Notes - Codex native Goal mode remains the visible task tracker; the Ultragoal file is the stronger contract, and the Codex Stop hook or runner enforces that contract. - Installed plugin hooks require Codex hook trust. Interactive users review them with `/hooks`; headless `npx ultragoal run --codex --headless` passes `--dangerously-bypass-hook-trust` because the command is explicitly launching this vetted plugin. - The SessionStart bootstrap links bundled custom agents into `~/.codex/agents/`. If the first session reports that agents were linked, restart Codex before relying on `ultragoal-executor` or `ultragoal-verifier`. - Codex Stop hook behavior can vary by CLI version. If a hook does not block the turn, treat the hook as an advisory reminder and rely on the headless runner plus the verifier/panel rubric before completion. - Prefer Codex Browser Use for local web UI QA and Computer Use for simulator/desktop GUI flows when command-line proof is not faithful. - Keep the goal objective short and put long details in the draft file, because native goal text has a small limit.