ralphlisted
Install: claude install-skill buildproven/agent-kit
# Ralph — bounded backlog loop
**Usage**: `/bs:ralph [--target-dir <path>] [--inline] [--until "N items"] [--scope all|feature|bug|effort:S|effort:M] [--section all|high|medium|low] [--quality auto|95|98] [--classic] [--parallel] [--reflect-depth standard|deep] [--speculate auto|always|never] [--score-threshold 0.7] [--evidence-dir .claude/ralph-next] [--max-retries N] [--max-ci-retries M] [--max-quality-minutes N] [--no-compact] [--dry-run]`
**Defaults**: 10 items, Linear backlog plus any detected inline list, all
scopes/sections, automatic quality, standard reflection, speculation `auto`,
score threshold `0.7`, and a 45-minute aggregate quality budget.
This entrypoint is intentionally short. The full implementation examples,
failure matrix, parser rules, and compatibility history live in
[`reference.md`](reference.md). Read only the section needed for the current
state; do not copy the reference into another playbook.
## Non-negotiable contract
- You are the orchestrator. Drive exactly `PICK → IMPLEMENT → QUALITY →
REFLECT → DECIDE`; `INIT`, `BLOCK`, `SPLIT`, and `SPECULATE` are explicit
transitions, not informal retries.
- Before unattended work, run `scripts/autonomous-loop-runtime.js admit` with
the long-lived loop's owner PID and `--provider codex|claude` matching the
executing provider. The bundled adapter queries CodexBar; see
`../../docs/provider-usage.md` for per-computer setup. Never expose credentials or raw usage data;
state belongs under `$XDG_STA