ralph

name	ralph
description	MCP-owned Ralph loop around background evolve_step jobs
mcp_tool	ouroboros_ralph
mcp_args	{"lineage_id":"$lineage_id"}

/ouroboros:ralph

MCP-owned Ralph loop around background evolve_step jobs. "The boulder never stops."

Usage

ooo ralph --lineage-id <lineage_id>
/ouroboros:ralph --lineage-id <lineage_id>

# For a plain natural-language request, run `ooo interview` + `ooo seed` first,
# then call the MCP tool with a fresh lineage_id and the validated Seed YAML.

Trigger keywords: "ralph", "don't stop", "must complete", "until it works", "keep going"

How It Works

Ralph is owned by the ouroboros_ralph MCP tool. In non-plugin runtimes, the tool starts one background Ralph job, runs repeated evolve_step generations inside that job, and stops only when QA passes, convergence is reached, a terminal evolution action occurs, cancellation is requested, or max_generations is reached. In OpenCode plugin mode, the MCP tool returns a delegated_to_plugin envelope with job_id=None; the bridge plugin dispatches a child Task session that owns the loop instead of creating a local JobManager job.

The client skill should not reimplement the loop. Deterministic frontmatter dispatch is limited to the router's named --lineage-id option so raw trailing text is never treated as lineage identity. Raw natural-language ooo ralph "<request>" input must flow through the validated Seed path before any mutating Ralph loop starts. Until a lineage id and optional Seed YAML are prepared, ouroboros_ralph returns structured input guidance instead of starting a job. Once the inputs are prepared, start the MCP-owned Ralph surface once, then follow either the returned job tools path or the OpenCode Task widget path.

Instructions

When the user invokes this skill:

Load MCP Tools (Required first)

The Ouroboros MCP tools are often registered as deferred tools that must be explicitly loaded before use. Do this before preparing input or calling Ralph:

1. Use the active runtime's tool-discovery capability to find and load the Ralph/job MCP tools:

   tool discovery query: "+ouroboros ralph job"
   

2. The loaded tools may be exposed under plugin-prefixed names such as mcp__plugin_ouroboros_ouroboros__ouroboros_ralph. Use the actual tool names returned by runtime tool discovery; the bare names below are the canonical MCP tool names for documentation.
3. Confirm that ouroboros_ralph and the job tools (ouroboros_job_wait, ouroboros_job_status, ouroboros_job_result, and ouroboros_cancel_job) are callable. If the tools are unavailable, stop and tell the user that Ralph requires the Ouroboros MCP runtime.

Ralph Flow

1. Prepare lineage input:

   • If the user provides an existing lineage_id and explicitly wants to continue it, reuse that lineage_id and omit seed_content unless they explicitly provide an updated Seed.
   • If the user provides Seed YAML for a new Ralph run, use it as seed_content and generate a fresh lineage_id for this run. Keep lineage_id separate from Seed, interview, and session IDs so separate Ralph runs over the same Seed do not collide.
   • If the user provides only a plain natural-language request, do not treat it as a direct ooo ralph "<request>" command, do not freehand Seed YAML, and do not pass raw text as seed_content. Route through the authoritative Seed path first: ooo interview to capture requirements, then ooo seed / ouroboros_generate_seed to produce validated Seed YAML with the normal ambiguity gate. After Seed generation, call the MCP tool with a fresh lineage_id and that validated Seed YAML as seed_content; do not use the raw request text. If an interview/seed session already exists in context, reuse that validated Seed output instead of regenerating it.

2. Start Ralph by calling ouroboros_ralph with:

   • lineage_id: existing lineage id for an explicit continuation, otherwise a freshly generated stable id for this Ralph run, such as ralph-<short-slug>-<uuid>; do not use a Seed/interview id by itself
   • seed_content: valid Seed YAML for generation 1 when starting a new lineage
   • execute: default true
   • parallel: default true
   • skip_qa: default false
   • project_dir: explicit target project directory when known
   • max_generations: default 10 unless the user requests a tighter bound

3. Handle the start response:

   • If response.meta.job_id is present, report it concisely and retain the job cursor from response.meta.cursor:

Active Conductor decision policy

For attention_required, use at most one short-lived read-only verifier. If the host has no verifier primitive, surface the evidence and do not mutate. Otherwise VERIFY → DECIDE from recommended_host_actions → LOG selected with ouroboros_record_conductor_decision → ACT only a menu-listed registered tool → LOG exactly one completed, failed, or declined outcome. Ralph may apply a conductor directive only to the first and sole bounded successor generation (max_generations=1), and only when it is deterministic and non-relaxing. Never silently retry or weaken the approved shared contract.

These are English canonical host instructions. Render them naturally in the user's conversation language.

Tool Mapping

Skill action	MCP tool
Start Ralph loop	ouroboros_ralph
Wait for progress	ouroboros_job_wait
Fetch final result	ouroboros_job_result
Cancel loop	ouroboros_cancel_job
Inspect current status	ouroboros_job_status

The Boulder Never Stops

This is the key phrase. Ralph does not give up:

• Each failure is data for the next attempt.
• Verification drives the loop.
• Only success, convergence, terminal failure, cancellation, or max-generation limits stop it.

RFC #1392 State Breadcrumb Footer

Your final response MUST end with exactly one breadcrumb footer line:

◆ <current state> → next: <recommended action>

Derive <current state> from live session state via ouroboros_session_status when that MCP projection is available; otherwise derive it from this skill's actual outcome. Never use a linear Step N of M footer because Ouroboros is an evolutionary loop. When the next action is genuinely a choice, list 2-3 honest options in the next: clause. The breadcrumb line must be the last line of the response.