| name | ultrawork |
| description | Autonomous high-throughput execution for complex tasks. Trigger when the user says ulw, ultrawork, keep going, finish it, complete this end-to-end, or asks for no-hand-holding work across multiple files, tests, investigation, or review. Uses omk-ralph hook state for persistence and omk-review as the final quality gate. Do not use for simple one-shot answers or tiny edits. |
Ultrawork
Rule: the user says what; Ultrawork decides how and keeps going. Constant asking defeats
the purpose — the user chose Ultrawork to delegate decisions, not to be interrupted for
every one of them.
Ultrawork is a coordinator, not a fixed list of OMK-only abilities. Use the current session's
available Kimi skills, native tools, project skills, user skills, and plugin skills as the source of
truth.
Use OMK Ralph persistence for every Ultrawork task:
./.omk/state/ralph-state.json
Keep completion_promise as OMK_RALPH_DONE unless the user provides a clearer promise. Use
max_iterations: -1 for no-hand-holding completion work.
Start
- State
ULTRAWORK MODE ENABLED when responding directly — sets the user's expectation
that autonomous, no-hand-holding mode is active.
- Ensure Ralph state exists and has
"workflow": "ralph" and "status": "active".
- Restate the target in one sentence.
- Decide whether a Capability Selection Pass is useful.
- Decide whether Plan Mode is required.
- Create a concise todo list for the main shards.
- Start work immediately after any required plan is approved.
The OMK Stop hook replays the original task while Ralph state is active. After "done", the hook
injects ralph/end.md once for the final completion summary. "blocked" stops without that
summary because the task is not complete.
If Ralph state is missing, create it manually before doing other work:
{
"version": 1,
"workflow": "ralph",
"source_skill": "ultrawork",
"status": "active",
"task": "<original user task>",
"completion_promise": "OMK_RALPH_DONE",
"iteration": 0,
"max_iterations": -1,
"phase": "starting",
"skill_selection_status": "not_evaluated",
"selected_skills": [],
"plan_required": "auto",
"plan_status": "pending",
"review_required": true,
"reason": "ultrawork started",
"evidence": ["initialized by ultrawork"],
"updated_at": "<ISO-8601 timestamp>"
}
Capability Selection Pass
Use this pass only when a discovered skill may materially change execution. Do not perform
it as a ceremony for small, obvious, or purely local tasks — burning tokens and adding
latency for no gain.
When useful, inspect the current session's Available Skills list. Consider Kimi built-in skills,
project skills, user skills, and plugin skills equally; OMK skills are not preferred by default.
For each likely useful skill:
- Decide whether it would materially change execution.
- Read its
SKILL.md from the listed path only if likely useful.
- Follow selected skill instructions during the task.
If you select skills, update Ralph state after this pass:
{
"skill_selection_status": "done",
"selected_skills": [
{ "name": "<skill>", "reason": "<why it changes execution>" }
]
}
If no skill check is worth the overhead, leave skill_selection_status as "not_evaluated" and
continue. Do not write an empty selection just to satisfy a template.
Reconsider this pass when the task scope changes.
Plan Gate
Use Kimi's native EnterPlanMode before execution when the task is non-trivial. The cost
of a few minutes planning is negligible compared to redoing hours of work because the
architecture was wrong from step one:
- new feature or meaningful behavior change
- multiple files or modules
- architecture, API, schema, permission, hook, or prompt-system change
- multiple valid approaches with different tradeoffs
- unclear scope or important user-visible choices
- broad cleanup/refactor where the plan affects safety
Skip Plan Mode only for small explicit fixes, pure read-only investigation, or small tasks
where a quick exploration is enough to choose an obvious path — for these, the planning
overhead exceeds the benefit. For small tasks, explore briefly and execute directly.
When Plan Mode is used, the plan must name selected skills, implementation shards, verification,
review gate, and Ralph completion condition. After ExitPlanMode approval, execute the approved
plan without re-planning unless new facts change the scope.
Autonomy
Continue without asking when the repo, tests, docs, or user wording imply a reasonable
default. Asking on every minor decision defeats the point of Ultrawork — the whole reason
the user invoked this mode is to delegate and walk away.
Ask only when the next action needs:
- credentials or external access
- destructive changes
- approval for broad or irreversible edits
- a user-visible choice with materially different behavior, cost, compatibility, or security
If unsure and the choice is local, reversible, and low impact, choose the smallest change that matches existing patterns and record the assumption.
Execution
Repeat until complete:
- Pick the highest-leverage unfinished todo — what unblocks the most downstream work?
- Gather missing context with project tools or focused subagents.
- Delegate only independent shards with non-overlapping write scope — overlapping writes
cause merge conflicts that waste time resolving.
- Edit narrowly in local style.
- Run the narrowest meaningful check — faster feedback means faster iteration.
- Fix failures caused by this work.
- Update todos and Ralph evidence — stale state causes duplicated or skipped work.
Do not stop at next steps when the next step is safe and available — each stop costs a
full user round-trip.
Keep Ralph evidence compact: command plus result, not raw output. The context window is
shared between evidence and useful working state; raw output pushes out what matters.
Subagents
Subagents do not run Ultrawork, Ralph, plan approval, review gates, or final completion.
Subagents are stateless — they cannot maintain the state machine or persist decisions.
The root agent owns orchestration.
Use subagents only for bounded work:
- focused read-only exploration
- independent implementation shards
- plan or review critique
- verification that can run without blocking the root agent
Tell subagents to return concise findings, changed files, checks run, and blockers. After they
return, the root agent updates todos, Ralph evidence, selected skills, plan status, review status,
and final done or blocked state.
Review Gate
Before marking Ralph "done", run omk-review for meaningful code, config, prompt,
workflow, or documentation changes. The review catches regressions and quality issues
before the user sees them — skipping it means shipping blind.
Review target priority:
- user-specified scope
- files changed during Ultrawork
- staged diff
- working-tree or branch diff
If omk-review finds real Critical/High/Medium issues caused by this work:
- fix them
- re-run relevant verification
- update review evidence
- keep Ralph
"active" until clean or blocked
Stop
Mark ./.omk/state/ralph-state.json:
"done" only when the target is complete, verified, review-blocking issues are resolved, and
reason or evidence includes OMK_RALPH_DONE
"blocked" only when progress requires user input, credentials, approval, destructive action, or broader scope
"active" while work remains
Final response:
Outcome:
- ...
Changed:
- ...
Verified:
- ...
Review:
- ...
Remaining risk:
- ...
Keep it short. Completed final responses are driven by ralph/end.md; do not narrate the whole
internal process.