Ralph loop

Ralph loop

Run an autonomous loop. The plugin's stop hook re-feeds the loop prompt after every turn, the agent executes exactly one step per iteration, and the loop ends only when the completion promise is genuinely true or a safety rail fires.

Seeding is a separate skill: ralph-loop-setup. This skill assumes a seeded loop, except for a quick inline ad-hoc start.

How the loop works

1. /ralph-loop-setup resolves configuration and runs scripts/seed-ralph-loop.sh, which writes {base}/active.md plus a run directory. {base} is .claude/loop or .cursor/loop, resolved from the agent, with no pointer file.
2. /ralph-loop start verifies the seeded files and executes iteration 1.
3. The plugin hooks take over. After every turn the stop hook re-feeds the body of active.md; the agent reads its own state and runs the next step.
4. The loop ends when the agent emits <promise>TEXT</promise> matching the configured promise, max_iterations is reached, the 200-iteration hard ceiling is hit, or the stall guard sees no state change for 3 consecutive iterations.

Safety rails

Rail	Default	Enforced by
max_iterations	50	stop hook, from frontmatter
hard ceiling	200	stop hook, applies even when unlimited
stall guard	3 unchanged iterations	stop hook, watches the state file
session isolation	owning session only	stop hook, from session_id
fix budgets	per preset	loop state counters

Exhausting a budget never fails the loop: the step advances and the unresolved findings are recorded under ## Notes so a human sees them.

References

• references/loop-protocol.md — the step machine, budgets, guardrails, and why one step per iteration
• references/preset-authoring.md — how to write a preset, with a worked non-engineering example
• references/prompt-authoring.md — completion promises, iteration budgets, escape hatches

Assets

• assets/loop.core.template.md — generic loop body, preset-agnostic
• assets/loop-state.core.template.md — per-run mutable state
• assets/context.core.template.md — per-run static context
• assets/presets/ — engineering-delivery, ad-hoc, custom

Router

Mode is start, status, or cancel. If no mode is given: a seeded {base}/active.md implies start; otherwise say the loop needs seeding and point at /ralph-loop-setup.

start — prompts/start.prompt.md. Verify the seeded loop, confirm the branch where the preset needs one, execute iteration

1. With --prompt "..." and no seeded loop, seed an ad-hoc loop first.

status — prompts/status.prompt.md. Report iteration, current step, budgets used, completed items, and artefacts. Read-only.

cancel — prompts/cancel.prompt.md. Stop the loop and archive the run directory as a record.

Ground rules

• Only output <promise>TEXT</promise> when the statement is completely and genuinely true. Never to escape the loop, never because progress is slow.
• Exactly ONE step per iteration. State lives in files, not in memory.
• Every skill step runs in a fresh sub-agent. Context isolation per step is what keeps a long run sharp.
• Where a preset commits, verify the branch with git branch --show-current first. No Co-authored-by trailers, no emojis in commits.
• If the external ralph-loop-plugin is installed, disable it. This plugin ships its own hooks and running both double-fires the stop hook.