omk-ralph

name

omk-ralph

description

Keep working autonomously under the oh-my-kimicli Stop-hook Ralph system until the task is complete, verified, or honestly blocked. Use when the user says keep going, don't stop until done, just finish it, complete this end-to-end, plow through it, or similar no-hand-holding language. Also used by ultrawork as its persistence layer. Do not use for one-shot Q&A, simple lookups, or tasks that should stop after a normal single response.

omk-ralph

Use OMK Ralph as a project-local replay loop.

State file:

./.omk/state/ralph-state.json

The hook initializes this file for /skill:omk-ralph .... While status is "active", Stop injects prompts/ralph/continue.md with the original task. When status becomes "done", Stop injects prompts/ralph/end.md once to summarize the Ralph turn, then allows the final stop. "blocked" allows stop without the end summary because the task is not complete.

State Contract

Keep valid JSON:

{
  "version": 1,
  "workflow": "ralph",
  "status": "active",
  "task": "<original task>",
  "completion_promise": "OMK_RALPH_DONE",
  "iteration": 0,
  "max_iterations": -1,
  "reason": "<current state>",
  "evidence": ["<short proof>"],
  "updated_at": "<ISO-8601 timestamp>"
}

Statuses:

• "active": keep working
• "done": task is complete and verified
• "blocked": progress requires user input, credentials, approval, destructive action, or broader scope

max_iterations: -1 means unlimited. A positive value lets the hook mark the task blocked after the limit is exceeded.

completion_promise is the phrase the hook uses to verify that the agent intentionally marked the state "done". The hook looks for this exact string in the state file before finalizing. OMK_RALPH_DONE is the standard value. A caller (like ultrawork) may replace it with a more specific promise, but the hook requires the stop to include an agreed-upon token so a spurious "done" from a tool call does not end the session prematurely.

Avoid Common Confusion

• The state file is the source of truth; saying "done" in chat is not enough.
• The path is the user's current project directory, not the package directory or global Kimi share dir.
• There is no EnterRalphMode, ExitRalphMode, or <choice>CONTINUE</choice> in OMK Ralph.
• After "done", do not set the state back to "active" just to write the summary.
• Keep evidence short. Record command names/results and key observations, not raw logs or diffs.

Work Rule

Treat the user's original task as the stop condition.

Continue while useful work remains. Do the next concrete action instead of describing future work. Before "done", run the narrowest meaningful proof: targeted tests, lint/typecheck, format check, build, smoke test, or direct inspection when that is the right proof.

If a check fails because of your work, fix it and re-check. If it is unrelated, record that clearly.

Finish

Write "done" only after completion and verification:

{
  "workflow": "ralph",
  "status": "done",
  "task": "<original task>",
  "completion_promise": "OMK_RALPH_DONE",
  "reason": "task completed and verified: OMK_RALPH_DONE",
  "evidence": ["npm run check: passed"]
}

Write "blocked" with a concrete reason when blocked. The hook will allow the blocker response without injecting end.md.