SKILL.md 19.5 KB

Raw Blame History Permalink


name: ralph
description: "[OMX] Self-referential loop until task completion with architect verification"


[RALPH + ULTRAWORK - ITERATION {{ITERATION}}/{{MAX}}]

Your previous attempt did not output the completion promise. Continue working on the task.


Ralph is a persistence loop that keeps working on a task until it is fully complete and architect-verified. It wraps ultrawork's parallel execution with session persistence, automatic retry on failure, and mandatory verification before completion.


Task requires guaranteed completion with verification (not just "do your best")
User says "ralph", "don't stop", "must complete", "finish this", or "keep going until done"
Work may span multiple iterations and needs persistence across retries
Task benefits from parallel execution with architect sign-off at the end


User wants a full autonomous pipeline from idea to code -- use autopilot instead
User wants to explore or plan before committing -- use plan skill instead
User wants a quick one-shot fix -- delegate directly to an executor agent
User wants manual control over completion -- use ultrawork directly


Complex tasks often fail silently: partial implementations get declared "done", tests get skipped, edge cases get forgotten. Ralph prevents this by looping until work is genuinely complete, requiring fresh verification evidence before allowing completion, and using explicit architect native-subagent verification to confirm quality.


Fire independent agent calls simultaneously -- never wait sequentially for independent work
Use run_in_background: true for long operations (installs, builds, test suites)
Always set agent_type when spawning native subagents; use reasoning_effort for per-dispatch intensity when needed
Preserve legacy Ralph tier intent through native reasoning effort: LOW -> low, STANDARD -> medium, THOROUGH -> xhigh

Deliver the full implementation: no scope reduction, no partial completion, no deleting tests to make them pass
Apply the shared workflow guidance pattern: outcome-first framing, concise visible updates for multi-step execution, local overrides for the active workflow branch, validation proportional to risk, explicit stop rules, and automatic continuation for safe reversible steps. Ask only for material, destructive, credentialed, external-production, or preference-dependent branches.
Integrate with Codex goal mode when goal tools are available: inspect the active thread goal with get_goal, preserve it as the top-level stop condition, and only call update_goal({status: "complete"}) after a Ralph completion audit proves the objective is actually achieved.


Pre-context intake (required before planning/execution loop starts):


Assemble or load a context snapshot at .omx/context/{task-slug}-{timestamp}.md (UTC YYYYMMDDTHHMMSSZ).
Minimum snapshot fields:


task statement
desired outcome
known facts/evidence
constraints
unknowns/open questions
likely codebase touchpoints


If an existing relevant snapshot is available, reuse it and record the path in Ralph state.
If request ambiguity is high, gather brownfield facts first. omx explore is deprecated; use normal repository inspection tools/subagents for simple read-only repository lookups and omx sparkshell only for explicit shell-native read-only evidence. Then run $deep-interview --quick <task> to close critical gaps.
Do not begin Ralph execution work (delegation, implementation, or verification loops) until snapshot grounding exists. If forced to proceed quickly, note explicit risk tradeoffs.


Review progress: Check TODO list and any prior iteration state

Continue from where you left off: Pick up incomplete tasks

Delegate in parallel: Route tasks to specialist native agents with explicit agent_type and appropriate reasoning_effort


Simple lookups: reasoning_effort="low" -- "What does this function return?"
Standard work: reasoning_effort="medium" -- "Add error handling to this module"
Complex analysis: reasoning_effort="xhigh" -- "Debug this race condition"
When Ralph is entered as a ralplan follow-up, start from the approved available-agent-types roster and make the delegation plan explicit: implementation lane, evidence/regression lane, and final sign-off lane using only known agent types


Run long operations in background: Builds, installs, test suites use run_in_background: true


Visual task gate (when screenshot/reference images are present):


Run the Visual Ralph verdict step before every next edit.
Require structured JSON output: score, verdict, category_match, differences[], suggestions[], reasoning.
Persist verdict to .omx/state/{scope}/ralph-progress.json including numeric + qualitative feedback.
Default pass threshold: score >= 90.

URL-based visual cloning tasks: When the task description contains a target URL (e.g., "clone https://example.com"), route the work through $visual-ralph. $web-clone is hard-deprecated; Visual Ralph owns the migrated live-URL visual implementation use case and uses its built-in visual verdict step for measured visual scoring.


Verify completion with fresh evidence:


If Codex goal mode is available, call get_goal before final verification to restate the active objective and include it in the evidence checklist.
a. Identify what command proves the task is complete
b. Run verification (test, build, lint)
c. Read the output -- confirm it actually passed
d. Check: zero pending/in_progress TODO items


Architect verification (native role):


<5 files, <100 lines with full tests: task(agent_type="architect", reasoning_effort="medium", prompt="...") minimum
Standard changes: task(agent_type="architect", reasoning_effort="medium", prompt="...")

>20 files or security/architectural changes: task(agent_type="architect", reasoning_effort="xhigh", prompt="...")

Ralph floor: always run an explicit architect native subagent, even for small changes
7.5 Mandatory Deslop Pass:
After Step 7 passes, run oh-my-codex:ai-slop-cleaner on all files changed during the Ralph session.
Scope the cleaner to changed files only; do not widen the pass beyond Ralph-owned edits.
Run the cleaner in standard mode (not --review).
If the prompt contains --no-deslop, skip Step 7.5 entirely and proceed with the most recent successful verification evidence.
7.6 Regression Re-verification:
After the deslop pass, re-run all tests/build/lint and read the output to confirm they still pass.
If post-deslop regression fails, roll back cleaner changes or fix and retry. Then rerun Step 7.5 and Step 7.6 until the regression is green.
Do not proceed to completion until post-deslop regression is green (unless --no-deslop explicitly skipped the deslop pass).


On approval: If Codex goal mode is active, call update_goal({status: "complete"}) before /cancel; report final elapsed time and token-budget usage when the tool returns it. Then run /cancel to cleanly exit and clean up all state files.

On rejection: Fix the issues raised, then re-verify with the same agent_type and reasoning_effort profile


Use ask_codex with agent_role: "architect" for verification cross-checks when changes are security-sensitive, architectural, or involve complex multi-system integration
Skip Codex consultation for simple feature additions, well-tested changes, or time-critical verification
If MCP compatibility tools are unavailable, proceed with CLI/agent verification alone -- never block on external tools
Use omx state write/read --input '<json>' --json for ralph mode state persistence between iterations
Use Codex goal tools when present: get_goal to discover or re-check the active objective, create_goal only when the user/system explicitly requested a new goal and no active goal exists, and update_goal only after the audited objective is fully achieved.
Persist context snapshot path in Ralph mode state so later phases and agents share the same grounding context
Prefer CLI state commands. If an explicit MCP compatibility omx_state call reports that its stdio transport is unavailable/closed, do not retry the same MCP call. Retry once through the supported CLI parity surface with the same payload, preserving workingDirectory and session_id: omx state write --input '<json>' --json, omx state read --input '<json>' --json, or omx state clear --input '<json>' --json. If the CLI path also fails, continue with .omx/context / .omx/plans file-backed artifacts and report the state persistence blocker.


Goal Mode Integration

Codex goal mode is the thread-level completion contract for long-running Ralph work. Ralph state tracks workflow mechanics; goal mode tracks whether the user objective is truly done. When the goal tools are available:


Call get_goal during intake or before the first execution loop when the prompt/hook says an active thread goal exists.
If no goal exists, call create_goal only when the user or system explicitly asked for goal tracking; otherwise continue with Ralph state alone.
Treat goal.objective as binding acceptance scope. Newer user updates can refine the current branch, but do not silently narrow the goal.
Before completion, perform a prompt-to-artifact checklist and completion audit against real evidence:


restate the objective as deliverables/success criteria
map every prompt requirement, named workflow ($ralplan, $ralph), file, command, test, gate, and deliverable to evidence
inspect the actual files, command output, state, and tests behind each checklist item
identify missing, weakly verified, or uncovered requirements and continue if any remain


Call update_goal({status: "complete"}) only when the audit shows no required work remains. Do not use passing tests, Ralph state, or architect approval as proxy proof unless they cover the whole goal.
If goal tools are unavailable, keep working through Ralph state and mention the missing goal-mode evidence in the final report.


State Management

Use the CLI-first state surface for Ralph lifecycle state (omx state write/read/clear --input '<json>' --json). Explicit MCP compatibility tools (state_write, state_read, state_clear) remain acceptable only when already enabled.


On start:
omx state write --input '{"mode":"ralph","active":true,"iteration":1,"max_iterations":10,"current_phase":"executing","started_at":"<now>","state":{"context_snapshot_path":"<snapshot-path>"}}' --json


On each iteration:
omx state write --input '{"mode":"ralph","iteration":<current>,"current_phase":"executing"}' --json


On verification/fix transition:
omx state write --input '{"mode":"ralph","current_phase":"verifying"}' --json or omx state write --input '{"mode":"ralph","current_phase":"fixing"}' --json


On completion (only after the completion audit passes with real evidence):
omx state write --input '{"mode":"ralph","active":false,"current_phase":"complete","completed_at":"<now>","completion_audit":{"passed":true,"prompt_to_artifact_checklist":["<requirement mapped to artifact/evidence>"],"verification_evidence":["<fresh test/build/lint command and result>"]}}' --json


Before the final answer:


Run fresh verification and read the output.
Build prompt_to_artifact_checklist entries that map every user requirement, workflow gate, named file, command, PR/delivery requirement, and stop condition to a concrete artifact or evidence item.
Build verification_evidence entries with concrete commands, exit status, files inspected, PR URLs, or other machine-checkable evidence.
Write the Ralph completion state with a top-level completion_audit field on the Ralph state object. Do not write bare top-level prompt_to_artifact_checklist or verification_evidence fields by themselves; the Stop gate will reject them.
Read the state back with omx state read --input '{"mode":"ralph"}' --json and verify completion_audit.passed === true, a non-empty checklist, and non-empty verification evidence before producing the final answer.
If Codex goal mode is active, call update_goal({status:"complete"}) only after this Ralph audit read-back succeeds.


On cancellation/cleanup:
run $cancel (which should call omx state clear --input '{"mode":"ralph"}' --json)


Scenario Examples

Good: The user says continue after the workflow already has a clear next step. Continue the current branch of work instead of restarting or re-asking the same question.

Good: The user changes only the output shape or downstream delivery step (for example make a PR). Preserve earlier non-conflicting workflow constraints and apply the update locally.

Bad: The user says continue, and the workflow restarts discovery or stops before the missing verification/evidence is gathered.


Correct parallel delegation:

task(agent_type="executor", reasoning_effort="low", prompt="Add type export for UserConfig")
task(agent_type="executor", reasoning_effort="medium", prompt="Implement the caching layer for API responses")
task(agent_type="executor", reasoning_effort="xhigh", prompt="Refactor auth module to support OAuth2 flow")


Why good: Three independent tasks fired simultaneously while explicitly selecting the installed executor native role, so the UI/tracker does not show default subagents; legacy tier intent is preserved through native reasoning effort (LOW -> low, STANDARD -> medium, THOROUGH -> xhigh).


Correct verification before completion:

1. Run: npm test           → Output: "42 passed, 0 failed"
2. Run: npm run build      → Output: "Build succeeded"
3. Run: lsp_diagnostics    → Output: 0 errors
4. task(agent_type="architect", reasoning_effort="medium", prompt="verify completion") → Verdict: "APPROVED"
5. Run /cancel


Why good: Fresh evidence at each step, architect verification, then clean exit.


Claiming completion without verification:
"All the changes look good, the implementation should work correctly. Task complete."
Why bad: Uses "should" and "look good" -- no fresh test/build output, no architect verification.


Sequential execution of independent tasks:

task(agent_type="executor", reasoning_effort="low", prompt="Add type export") → wait →
task(agent_type="executor", reasoning_effort="medium", prompt="Implement caching") → wait →
task(agent_type="executor", reasoning_effort="xhigh", prompt="Refactor auth")


Why bad: These are independent tasks that should run in parallel, not sequentially.


Stop and report when a fundamental blocker requires user input (missing credentials, unclear requirements, external service down)
Stop when the user says "stop", "cancel", or "abort" -- run /cancel

Continue working when the hook system sends "The boulder never stops" -- this means the iteration continues
If architect rejects verification, fix the issues and re-verify (do not stop)
If the same issue recurs across 3+ iterations, report it as a potential fundamental problem


 All requirements from the original task are met (no scope reduction)

 Zero pending or in_progress TODO items

 Fresh test run output shows all tests pass

 Fresh build output shows success

 lsp_diagnostics shows 0 errors on affected files

 Architect verification passed through explicit task(agent_type="architect", reasoning_effort="medium"...) minimum

 Codex goal-mode completion audit passed, and update_goal({status: "complete"}) was called when an active goal exists

 ai-slop-cleaner pass completed on changed files (or --no-deslop specified)

 Post-deslop regression tests pass

 /cancel run for clean state cleanup


PRD Mode (Optional)

When the user provides the --prd flag, initialize a Product Requirements Document before starting the ralph loop.


Detecting PRD Mode

Check if {{PROMPT}} contains --prd or --PRD.

Prompt-side $ralph workflow activation is lighter-weight than omx ralph --prd ....
It seeds Ralph workflow state and guidance, but it does not implicitly launch the
CLI entrypoint or apply the PRD startup gate. Treat omx ralph --prd ... as the
explicit PRD-gated path.


Detecting --no-deslop


Check if {{PROMPT}} contains --no-deslop.
If --no-deslop is present, skip the deslop pass entirely after Step 7 and continue using the latest successful pre-deslop verification evidence.


Visual Reference Flags (Optional)

Ralph execution supports visual reference flags for screenshot tasks:


Repeatable image inputs: -i <image-path> (can be used multiple times)
Image directory input: --images-dir <directory>


Example:
ralph -i refs/hn.png -i refs/hn-item.png --images-dir ./screenshots "match HackerNews layout"


PRD Workflow


Run deep-interview in quick mode before creating PRD artifacts:


Execute: $deep-interview --quick <task>

Complete a compact requirements pass (context, goals, scope, constraints, validation)
Persist interview output to .omx/interviews/{slug}-{timestamp}.md


Create canonical PRD/progress artifacts:


PRD: .omx/plans/prd-{slug}.md

Progress ledger: .omx/state/{scope}/ralph-progress.json (session scope when available, else root scope)


Parse the task (everything after --prd flag)
Break down into user stories:


{
  "project": "[Project Name]",
  "branchName": "ralph/[feature-name]",
  "description": "[Feature description]",
  "userStories": [
    {
      "id": "US-001",
      "title": "[Short title]",
      "description": "As a [user], I want to [action] so that [benefit].",
      "acceptanceCriteria": ["Criterion 1", "Typecheck passes"],
      "priority": 1,
      "passes": false
    }
  ]
}


Initialize canonical progress ledger at .omx/state/{scope}/ralph-progress.json

Guidelines: right-sized stories (one session each), verifiable criteria, independent stories, priority order (foundational work first)
Proceed to normal ralph loop using user stories as the task list


Example

User input: --prd build a todo app with React and TypeScript
Workflow: Detect flag, extract task, create .omx/plans/prd-{slug}.md, create .omx/state/{scope}/ralph-progress.json, begin ralph loop.


Legacy compatibility


During the compatibility window, Ralph --prd startup still validates machine-readable story state from .omx/prd.json.

.omx/plans/prd-{slug}.md remains the canonical storage/documentation artifact, but it is not yet the startup validation source.
If .omx/prd.json exists and canonical PRD is absent, migrate one-way into .omx/plans/prd-{slug}.md.
If .omx/progress.txt exists and canonical progress ledger is absent, import one-way into .omx/state/{scope}/ralph-progress.json.
Keep legacy files unchanged for one release cycle.


Background Execution Rules

Run in background (run_in_background: true):


Package installation (npm install, pip install, cargo build)
Build processes (make, project build commands)
Test suites
Docker operations (docker build, docker pull)


Run blocking (foreground):


Quick status checks (git status, ls, pwd)
File reads and edits
Simple commands


Original task:
{{PROMPT}}