breadcrumb-loop

name	breadcrumb-loop
description	Run a long task across multiple agent contexts using breadcrumb progress files
user-invocable	true

Breadcrumb Loop

Run a long-running task that spans multiple agent contexts. Each agent picks up where the last left off via the /anvil:breadcrumb skill.

Instructions

Given the user's goal in $ARGUMENTS:

1. Create a task slug from the goal (kebab-case, e.g., implement-auth-module)
2. Create the breadcrumb directory at plans/<task-slug>-breadcrumb-log/
3. Write readme.md in that directory with:
   • The full objective (from $ARGUMENTS)
   • Acceptance criteria (infer from the goal, or ask the user if unclear)
   • Any relevant context you know about the codebase
4. Run the loop below, substituting your task slug

Important

Do NOT use run_in_background: true when invoking anvil-repl. The REPL manages long-running execution internally via child agent processes. Always run it in the foreground.

Loop

anvil-repl <<'ANVIL_REPL'
const DIR = "plans/<task-slug>-breadcrumb-log";

for (let i = 1; i <= 100; i++) {
  anvil.log(`Iteration ${i}`);

  // The slash command must lead the prompt — Claude Code only auto-expands
  // skills into <command-name> tags when /command is at the start of the message.
  const result = await anvil.spawn({
    prompt: `/anvil:breadcrumb ${DIR} ${i}`,
    contextShortCircuit: {
      limitPercent: 50,
      message: "Your context is running low. Stop new work and save your progress now.",
    },
  });

  if (/^BREADCRUMB_COMPLETE$/m.test(result)) {
    anvil.log("Task completed!");
    break;
  }
}
ANVIL_REPL

After the Loop

Breadcrumb agents commit their work directly to the branch. When the loop finishes:

1. Run git log --oneline to see what the breadcrumb agents committed
2. Read the latest *-progress.md file in the breadcrumb log directory for a summary