Hat	Stage	Upstream gate	Output
Orchestrator	Execution	/ralph-workspace-plan ran; fix_plans present	ralph/logs/{repo}.log, ralph/{repo}.pid, ralph/dispatch.log

When to Use

Per-repo fix_plans exist and user wants ralph to execute, ideally in parallel across multiple repos.

What This Skill Adds Over ralph Alone

Scope	Source
Within-repo parallelism (rpc.p N)	ai-ralph native
Across-repo parallelism — one loop per registered repo, pidfile + per-repo log	this skill
Continuous mode pass-through — --max-tasks M / positional --parallel N M, plus --max-task-attempts, --respawn-delay, --no-tabs, capability-gated against ralph --help	this skill

Continuous Mode

For long unattended runs over a deep workspace fix_plan, engage ralph’s continuous mode: workers stay saturated up to N concurrent until M total attempts have been spent, or the queue drains.

Resolution: CLI flag > env > project.conf > unset (batch mode).

Flag	Env var	project.conf key	Meaning
--max-tasks M	WB_RALPH_MAX_TASKS	WB_RALPH_MAX_TASKS	Engages continuous mode. Total attempts cap.
--max-task-attempts K	WB_RALPH_MAX_TASK_ATTEMPTS	WB_RALPH_MAX_TASK_ATTEMPTS	Per-task retry cap.
--respawn-delay SEC	WB_RALPH_RESPAWN_DELAY	WB_RALPH_RESPAWN_DELAY	Cooldown between worker respawns.
--no-tabs	WB_RALPH_DISABLE_TABS=true	WB_RALPH_DISABLE_TABS	Force single-pane orchestrator.

The positional form wb.ralph-dispatch --parallel N M is accepted as a shorthand and forwarded identically to ralph’s own --parallel N M shape.

wb.ralph-dispatch --parallel 3 --max-tasks 30
wb.ralph-dispatch --parallel 3 30
wb.ralph-dispatch --parallel 3 30 --no-tabs --respawn-delay 5

Prerequisites

• /ralph-workspace-plan ran — implies upstream artifacts all approved. Dispatch inherits the gate; does not re-check.
• .ralph/fix_plan.md exists and non-empty in every target repo.
• Workbench git-clean. Every target repo git-clean.

Protocol

1. Sanity checks — refuse and print specific failure:

    git -C "$WB_ROOT" status --porcelain
    for r in <selected-repos>; do git -C "$WB_ROOT/repos/$r" status --porcelain; done
    for r in <selected-repos>; do
      [[ -f "$WB_ROOT/repos/$r/.ralph/fix_plan.md" ]] && [[ -s "$WB_ROOT/repos/$r/.ralph/fix_plan.md" ]] \
        || echo "fix_plan missing or empty: $r"
    done
   

2. Select repos — default all with fix_plan; override --repos r1,r2.
3. Pick engine — respects DEVIN_DEFAULT; override --agent.
4. Launch: ./scripts/ralph-dispatch.sh [--repos {csv}] [--agent {engine}].
5. Report launched loops — PIDs, log paths.
6. On status requests, read pidfiles + logs: | Repo | PID | State | Last line |.
7. On completion (all PIDs gone) — read last 30 lines, extract exit code + any PR URL, update EPIC-PIPELINE.md Exec column (~ → ✓ or ✗), remove stale pidfiles.

Do Not

• Launch if workbench or any target repo is dirty.
• Silently ignore non-zero exits. Surface last 30 lines in status summary.
• Leave stale pidfiles.

Source

skills/ralph-dispatch/SKILL.md