← ClaudeAtlas

qualitylisted

Autonomous, revision-bound quality loop: deterministic gates, independent review, CI, and optional merge.
buildproven/agent-kit · ★ 0 · Code & Development · score 61
Install: claude install-skill buildproven/agent-kit
# Quality Skill Run autonomously from the exact `--manifest <path>` supplied by the wrapper; reject every other argument. The manifest is the sole source of campaign state. Never infer state from environment inheritance, session IDs, globbing, mtimes, or a “latest” pointer. Never source or eval manifest content. Campaign identity is deterministic for repository, PR, base, HEAD, scope, level, and merge intent. An identical request resumes its evidence, deadline, attempts, and terminal state. Changing provider order or retrying an unchanged HEAD cannot mint budget. Workers are ephemeral (`--no-session-persistence` / `--ephemeral`) and must obtain the same operator admission as Ralph before a new campaign. Each fenced Bash block starts a fresh shell. Begin every executable block with this resolver; it finds the installed runtime and fails closed when absent: ```bash QUALITY_SCRIPTS_DIR="$(for d in "${CLAUDE_PLUGIN_ROOT:-}" "${CLAUDE_KIT_ROOT:-}" "$HOME/.claude" .; do [ -n "$d" ] || continue; resolver="$d/scripts/quality-runtime-dir.sh"; [ -f "$resolver" ] || continue; bash "$resolver"; exit $?; done; echo "quality runtime not found" >&2; exit 1)" || exit $? ``` A host interruption remains visible in terminal history. Resuming the exact manifest continues unfinished exact-head phases only after the repository lease is reacquired and no execution owner remains active. A valid descendant archives the prior-head interruption, retains cumulative provider evidence and budgets, an