Rat Auto Design

Prerequisites

None — this is the full pipeline entry point. Setup is handled automatically.

Execution

# Parse mode
TEAM_MODE = "--no-team" not in ARGUMENTS

if TEAM_MODE:
    # ═══ TEAM MODE: Skill sequences phases directly ═══

    # State management
    Write(".rat/state/rat-auto-design-state.json",
      { "schema_version": "3.0", "status": "running",
        "current_phase": 1, "execution_mode": "team",
        "phases": { "1": {"status":"pending"}, "2": {"status":"pending"},
                    "3": {"status":"pending"}, "4": {"status":"pending"},
                    "5": {"status":"pending"}, "6": {"status":"pending"} } })

    # Phase 1: Research (team)
    Skill(skill="rtl-agent-team:rtl-p1-research-team", args="$ARGUMENTS")
    # Gate check + state update

    # Phase 2: Architecture + RefC (team)
    Skill(skill="rtl-agent-team:rtl-p2-arch-team", args="Phase 1 complete")
    # Gate check + state update

    # Phase 3: uArch + BFM (team)
    Skill(skill="rtl-agent-team:rtl-p3-uarch-team", args="Phase 2 complete")
    # Gate check + state update

    # Phase 4: RTL Implementation (team)
    Skill(skill="rtl-agent-team:rtl-p4-implement-team", args="Phase 3 complete")
    # Gate check + state update

    # Phase 5: Verification (team)
    Skill(skill="rtl-agent-team:rtl-p5-verify-team", args="Phase 4 complete")
    # Gate check + state update

    # Phase 6: Design Review (no team — sequential only)
    Bash("mkdir -p reviews/phase-6-review")
    Task(subagent_type="rtl-agent-team:p6-review-orchestrator",
         prompt="Execute Phase 6 design review. Context: Phase 5 PASS.")

    # Mark completed (stop-gate.sh reads top-level "status" to allow exit)
    # Write completed status BEFORE cleanup so stop hook sees it
    # Portable in-place JSON edit — avoids GNU/BSD sed -i differences (macOS BSD sed requires `-i ''`).
    Bash("python3 -c 'import json, pathlib; p = pathlib.Path(\".rat/state/rat-auto-design-state.json\"); d = json.loads(p.read_text()); d[\"status\"] = \"completed\"; p.write_text(json.dumps(d, indent=2))'")
    # Cleanup state
    Bash("rm -f .rat/state/rat-auto-design-state.json")

else:
    # ═══ SEQUENTIAL MODE: Delegate to autopilot-orchestrator ═══
    Task(subagent_type="rtl-agent-team:autopilot-orchestrator",
         prompt="Execute full RTL autopilot pipeline. --no-team. User input: $ARGUMENTS")

Team mode uses native teams within each phase for parallel execution (Orchestrator as Teammate pattern). Each phase team skill handles its own TeamCreate/TeamDelete lifecycle with a coordinator teammate + 3-5 workers. Sequential mode delegates everything to the autopilot-orchestrator.

Output

• Design artifacts: docs/phase-1-research/ through docs/phase-5-verify/
• Phase 6 design note and review verdicts: reviews/phase-6-review/ (Phase 6 is the design review + documentation phase; its outputs live under reviews/, not docs/)
• rtl/*/*.sv — lint-clean, unit-tested RTL modules
• sim/ — testbenches, regression results, and coverage data
• reviews/ — per-phase review reports with final compliance verdict