A goal-like loop that decomposes work into systematic, evidence-bound ultrawork steps. Use when the user wants a goal loop or durable, checkpointed execution.
Install
The skill file stays in the publisher's repository.
Paste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add code-yeongyu/oh-my-openagent --skill ulw-loop -a cursor -y
Download packages/omo-codex/plugin/components/ulw-loop/skills/ulw-loop/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does ulw-loop live?
The publisher keeps it at https://github.com/code-yeongyu/oh-my-openagent.
Which agents can use ulw-loop?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.
A goal-like loop that decomposes work into systematic, evidence-bound ultrawork steps. Use when the user wants a goal loop or durable, checkpointed execution.
name: ulw-loop
description: "A goal-like loop that decomposes work into systematic, evidence-bound ultrawork steps. Use when the user wants a goal loop or durable, checkpointed execution."
metadata:
short-description: Goal-like ultrawork loop for systematic decomposition
ulw-loop
Use this skill when the user asks for ulw-loop, ulw, durable goal execution, evidence-led work, manual QA, or checkpointed long-running delivery.
This skill is intentionally compact. The full workflow lives in references/full-workflow.md. Read only the sections needed for the current phase, then execute them exactly.
Required First Steps
Open references/full-workflow.md.
Read through Bootstrap (including its tier triage), Execution Loop, the Manual-QA channels table, and the Stop Rules before running any ULW command or recording evidence.
Open references/define-goal.md and register the run's goal by it. Goal creation is NEVER skipped: shape the objective and every success criterion by that reference before any implementation.
If the task has code edits, tests, QA, or commit work, follow the full workflow's delegation and evidence rules. Tests alone never prove done.
Non-Negotiables
Use the ulw-loop CLI state under .omo/ulw-loop; do not hand-edit goal state.
Register goals up front, shaped by references/define-goal.md (omo-agent-toolkit ulw-loop create-goals, then create_goal from the printed handoff), and mirror every atomic step into the live update_plan checklist: one ultra-granular step per action, exactly one in_progress, transitions marked the instant they happen. Use the ultrawork directive’s Task resources and todo rules; read those sections of the bundled ultrawork skill if absent from context.
After any compaction or context loss, re-read brief + goals + ledger FIRST plus omo-agent-toolkit ulw-loop status --json, then resume; never re-plan from scratch.
Every ulw-loop command needs the session scope: pass --session-id <id> (the printed handoff and resume directive carry it; CODEX_THREAD_ID in the environment also resolves it). The CLI refuses unscoped state (ULW_LOOP_SESSION_SCOPE_REQUIRED) instead of touching the shared .omo/ulw-loop root.
If omo-agent-toolkit ulw-loop create-goals says this session's aggregate is already complete, start unrelated new work with a fresh --session-id <new-id> (passed on every later call) instead of steering or forcing the completed state. Use --force only to intentionally overwrite completed evidence.
Every success criterion needs observable evidence from a real surface: a channel (terminal/TUI via the xterm.js web terminal, HTTP, browser, computer-use) or, for CLI- or data-shaped criteria, an auxiliary surface (CLI stdout, DB diff, parsed config dump).
Evidence is bound to the tree it was captured at (git rev-parse --short "HEAD^{tree}"); it goes stale only when tracked content changes — a rebase or amend that keeps the tree identical keeps it valid. When the tree differs, re-run at the current HEAD and re-record, never relabel or regenerate. Record only after cleanup receipts exist.
Delegate code edits, test writes, fixes, and QA execution to right-sized Codex subagents when the workflow requires it.
Every spawn_agent message starts with TASK:, then names DELIVERABLE, SCOPE, and VERIFY; put role and specialty instructions inside message; use fork_turns: "none" (v1: fork_context: false) unless full history is truly required.
Plan and reviewer agents may run for a long time; spawn them in the background and keep doing independent root work. Between wait_agent calls, back off — double the timeout up to ~5 minutes — instead of spinning short cycles.
For work likely to exceed one wait cycle, require the child to send WORKING: <task> - <current phase> before long reading, testing, or review passes, and BLOCKED: <reason> only when it cannot progress.
Track spawned agent names locally. Use wait_agent for mailbox signals, not proof of completion. A timeout only means no new mailbox update arrived. Treat a running child as alive.
While children run, surface the active subagent count, agent names, and latest WORKING: phase.
Fallback only when the child is completed without the deliverable, ack-only after followup_task, explicitly BLOCKED:, or no longer running. Then record inconclusive and respawn a smaller fork_turns: "none" task with the missing deliverable.
Use git-master for git-tracked edits: inspect recent and touched-path commit history, then commit each verified work unit atomically in the repository's observed language, scope, and message style with only that unit's files staged. Never carry verified units into a later omnibus commit.
Team mode: decide it, do not default to it
Solo execution with parallel background task workers is the default. A team (team_create) adds per-member briefing, shared-state, and relay overhead, so it must be paid for by the work's shape. Decide ONCE, when the plan's work units are known, and record the verdict plus its reason in the notepad.
Stand up a team when BOTH hold:
The units' scopes overlap in a way you cannot cleanly cut. They touch the same module, contract, or migration, so one unit's discovery changes what another should do. Fire-and-forget workers cannot exchange that mid-flight; teammates can, because the lead relays it.
Running them at the same time actually finishes sooner. The units are each substantial and none is merely waiting on another's output. Two units where the second only consumes the first's result are a sequence, not a team.
When the units are genuinely independent — separate files, no shared contract — spawn parallel background task workers instead and avoid the team coordination overhead entirely. When the work is one cohesive unit, do it yourself. Overlap alone is not enough: near-identical units that would collide on the same lines are faster done in sequence by one worker.
Under team mode, isolate and land per unit:
One git worktree per member, never a shared checkout — concurrent members editing one working tree corrupt each other's diffs and evidence. Give each member its own branch off the base and its own worktree path. Apply the same resource rules when creating member worktrees.
Merge per work unit, as each unit is verified. A member's unit lands when its own evidence is captured and its gates are green; it does not wait for the slowest sibling. Integrate each merged unit back into the base the others branch from, so overlapping members rebase onto real merged work rather than guessing at it.
Conflicts are the lead's job. When two members' units touch the same lines, the lead decides the order they land and tells the later member what changed; members never resolve a sibling's conflict blind.
Codex Tool Mapping
Codex exposes ONE subagent surface per session — check your tool list. GPT-5.6 (sol/terra) get the flat MultiAgentV2 tools (primary); GPT-5.5 and gpt-5.6-luna get the namespaced multi_agent_v1.* set (fallback row). The workflow's orchestration examples map to:
Intent
MultiAgentV2 (gpt-5.6 sol/terra)
Spawn a worker
spawn_agent({"task_name":"<lower_snake_id>","message":"TASK: act as <role>. ...","fork_turns":"none"}) — task_name+message required; fork_turns:"none" = no parent history; do NOT set agent_type/model/reasoning_effort
When translating load_skills=[...], include the requested skill names in the spawned agent's message.
Driver goal lifecycle
The Codex thread goal is a DRIVER the loop instructs, never a gate. checkpoint takes an OPTIONAL --codex-goal-json snapshot, records it verbatim in the ledger, and never rejects on its status or objective; the advice arrives in nextActions. A driver completed early yields advice to create_goal again with the plan's objective verbatim; paused, usage_limited, and budget_limited yield resume advice; a differing objective is a warning, not a refusal. Malformed snapshot input is the only failure, reported as ULW_LOOP_CODEX_GOAL_JSON_INVALID.
How do I install ulw-loop?
Download packages/omo-codex/plugin/components/ulw-loop/skills/ulw-loop/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does ulw-loop live?
The publisher keeps it at https://github.com/code-yeongyu/oh-my-openagent.
Which agents can use ulw-loop?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.