| name | rb:full |
| description | Use when running the end-to-end lifecycle: plan, work, verify, review, and compound chained in one command. Runs autonomously on the happy path; halts and asks the user on `/rb:verify --full` gate failure (HALTED_VERIFY_FAILED), review verdict BLOCKED (HALTED_REVIEW_BLOCKED), REQUIRES CHANGES (HALTED_REVIEW_REQUIRES_CHANGES), or missing/unparsable consolidated review (HALTED_REVIEW_UNKNOWN). Reaches COMPLETED autonomously once the compound solution doc is written โ no final user-acknowledgment prompt. |
| when_to_use | Triggers: "do everything", "full lifecycle", "hands-off", "plan and implement", "end to end". |
| argument-hint | <feature description OR plan path> |
| effort | xhigh |
Full Workflow
/rb:full runs the full plan-work-verify-review-compound cycle autonomously.
Skill body owns the complete cycle from main session.
STEP 0: Read the cycle + state-machine references
Read ${CLAUDE_SKILL_DIR}/references/cycle-patterns.md and
${CLAUDE_SKILL_DIR}/references/state-machine.md. Apply the state
transitions, blocker handling, and recovery protocols as the
canonical procedure. Each child skill (/rb:plan, /rb:work,
/rb:verify, /rb:review, /rb:compound) runs its own STEP 0
playbook read on entry.
Cycle
/rb:brainstorm (optional) โ if requirements vague/rb:plan โ reuses pre-bound namespace; produces plan.md via main-session research fanout/rb:work โ executes plan tasks; clears active-plan marker on completion/rb:verify โ runs full verification stack (no plan-path arg; resolves from current branch)/rb:review โ main-session review fanout (no plan-path arg; resolves from diff)/rb:compound โ captures learnings; accepts plan path arg
Use /rb:full for well-scoped work where the user wants the full cycle.
Skip for vague requirements or trivial fixes.
State Machine (summary)
INITIALIZING โ [DISCOVERING] โ PLANNING โ WORKING โ VERIFYING โ REVIEWING โ COMPOUNDING โ COMPLETED
DISCOVERING is optional (/rb:brainstorm). INITIALIZING transitions
directly to PLANNING when brainstorm is skipped. Branch rules:
${CLAUDE_SKILL_DIR}/references/state-machine.md ยง "Phase Transitions".
/rb:full skill body writes **State**: to progress.md at every
transition. The plan-stop-reminder.sh hook checks for this field to
skip the manual plan-presentation reminder during autonomous runs.
PLAN_DIR Tracking
- Derive
PLAN_DIR once at INITIALIZING; track as local variable through
the cycle.
- Use PLAN_DIR for skill-body's own progress.md State writes only.
- Do NOT pass PLAN_DIR as CLI arg to
/rb:verify or /rb:review.
- Do pass
${PLAN_DIR}/plan.md to /rb:compound.
- After
/rb:work clears the active-plan marker, continue using local
PLAN_DIR. Do NOT re-read the marker mid-cycle.
Slug Pre-Bind Protocol (summary)
/rb:full creates the plan namespace and writes initial progress.md
BEFORE invoking /rb:plan. /rb:plan reads .claude/ACTIVE_PLAN
directly with strict guards (NOT via active-plan-marker.sh get,
which has disk-glob fallbacks) and reuses the pre-bound slug.
Completion Criteria
A workflow is COMPLETED when:
Laws
- Never skip verification โ Always run full verification suite
- Never hide blockers โ Log them in progress.md and scratchpad.md
- Re-read plan.md after compaction โ Checkboxes remain source of truth
- Prefer small steps โ Checkpoint frequently
- Maintain state externally โ Files, not memory
- Delegate appropriately โ Don't do specialist work in main session
- Ask when uncertain โ Better to clarify than assume
Detail References
For implementation detail (state machine table, transition rules, slug
pre-bind protocol with example bash, marker concurrency notes,
Verification Gate ordering, Initial Progress Schema, Integration Points),
see:
${CLAUDE_SKILL_DIR}/references/state-machine.md${CLAUDE_SKILL_DIR}/references/cycle-patterns.md${CLAUDE_SKILL_DIR}/references/execution-steps.md${CLAUDE_SKILL_DIR}/references/safety-recovery.md${CLAUDE_SKILL_DIR}/references/example-run.md
