Guardrails

Ralph is intentionally conservative.

Fresh context

Each iteration starts from a fresh agent session, then re-anchors against repo state.

Receipt gates

Ralph does not trust a model saying “done.” It checks receipts written by review commands.

Multi-model review

Implementation is checked by a different model through the configured backend. Introduced findings gate the verdict.

No autonomous product decisions

/flow-next:strategy, /flow-next:prospect, and /flow-next:capture are user-triggered surfaces. Ralph does not decide what to build.

Test first

Run one iteration before overnight mode:

scripts/ralph/ralph_once.sh

Guardrail stack

flowchart TB
  Spec["Reviewed spec"] --> Ready["Ready task only"]
  Ready --> Fresh["Fresh session"]
  Fresh --> Evidence["Evidence commands"]
  Evidence --> Receipt["Review receipt"]
  Receipt --> Continue["Continue or stop"]

Human-owned boundaries

Ralph should not decide:

• whether a feature belongs in the product
• whether a risky migration is acceptable
• whether to merge
• whether to ignore a security finding
• whether undocumented behavior is acceptable

Those decisions belong in the spec, review notes, or human merge decision.

Production defaults

For production use:

• Require plan review for risky specs.
• Require implementation review with a real backend.
• Keep iteration caps low until the repo has strong tests.
• Prefer one spec branch per Ralph run.
• Review generated PR bodies before merge.

Ralph is powerful when the workflow is already disciplined. Without crisp specs and review gates, it only makes drift faster.