deepx-ai/dx-compiler/dx-agent-brainstorm

/dx-agent-brainstorm — Brainstorm & Plan Before Compilation

**RIGID skill. HARD-GATE: No code, no commands, no compilation without

an approved plan.**

Trigger Words

"plan", "brainstorm", "new model", "compile this", "convert and compile"

Type: RIGID

HARD-GATE skill. You MUST NOT generate config files, run dxcom, run torch.onnx.export, or execute any compilation step until the plan is approved by the user.

Phase 1: Context Check

Gather facts before asking questions. Run silently:

ls *.onnx *.pt *.pth 2>/dev/null                    # Input model format
dxcom --version >/dev/null 2>&1 && echo "DX-COM: ok" # Tool available?
ls dx_com/calibration_dataset/ 2>/dev/null | head -5  # Calibration data?

Model Acquisition Rule (MANDATORY)

If NO local .pt, .pth, or .onnx file is found for the requested model:

1. Identify the official download source (Ultralytics releases, torchvision,

timm, ONNX Model Zoo, Hugging Face)

1. Download the model file into the session working directory
2. Proceed with the full compilation pipeline — do NOT stop at providing

download instructions or generating only config.json

The user expects a compiled `.dxnn` file, not a recipe.

Common download methods:

• Ultralytics YOLO: from ultralytics import YOLO; YOLO("yolo11n.pt")
• TorchVision: torchvision.models.resnet50(weights="DEFAULT")
• Timm: timm.create_model("efficientnet_b0", pretrained=True)
• ONNX Model Zoo / Hugging Face: wget or huggingface_hub.hf_hub_download()

Gate: Know the input format and tool availability. If model file is missing, download it before proceeding.

Phase 2: Key Decisions (Ask the User)

Q1 — Conversion needed? If PyTorch (.pt/.pth): "Your model is PyTorch. I'll convert to ONNX with opset 13, batch=1, static shapes. Confirm input shape (e.g., 1,3,640,640)?"

Q2 — Calibration data? DX-COM always quantizes to INT8. There is NO precision selection parameter (no FP16/FP32 option). NEVER ask the user to choose output precision. "Do you have calibration images representative of inference data? How many? (recommend 50-200). Calibration method: EMA (default), MinMax, or IQR?"

Q3 — PPU? (detection models only) "Enable PPU post-processing? Type 0 = anchor-based (YOLOv3-v7), Type 1 = anchor-free (YOLOv8+). Default: OFF."

Q4 — ONNX simplification? "Run onnx-simplifier? May break node names, risk precision loss. Default: NO."

Gate: User answered Q1-Q2 minimum. Q3-Q4 asked if applicable.

Phase 3: Build Execution Plan

## Compilation Plan
**Model**: <name>  **Format**: <ONNX|PyTorch>  **Target**: dx_m1
 
### Steps:
1. [ ] Create session directory: dx-agent-dev/<session_id>/
2. [ ] <if PyTorch> Convert to ONNX → /dx-agent-compiler-convert
3. [ ] Validate ONNX (opset, batch=1, static shapes, single output)
4. [ ] Generate config.json (match input name from ONNX graph)
5. [ ] Symlink calibration dataset
6. [ ] Compile ONNX → DXNN → /dx-agent-compiler-compile
7. [ ] Validate output → /dx-agent-compiler-validate
8. [ ] Generate session report
 
**Routing**: /dx-agent-compiler-convert → /dx-agent-compiler-compile → /dx-agent-compiler-validate

Gate: User approves the plan ("looks good", "yes", "go ahead").

Phase 4: Pre-Flight Check

ALL must pass before execution:

test -f "<model_path>" && echo "PASS: model exists" || echo "FAIL"
dxcom --version >/dev/null 2>&1 && echo "PASS: DX-COM ok" || echo "FAIL"
WORK_DIR="dx-agent-dev/$(date +%Y%m%d-%H%M%S)_<model>_<task>"  # local timezone (NOT UTC)
mkdir -p "${WORK_DIR}" && echo "PASS: output dir ok" || echo "FAIL"
ls dx_com/calibration_dataset/*.jpg 2>/dev/null | wc -l

Gate: All pre-flight checks pass. If any FAIL, stop and resolve.

Phase 5: Route to Skills

1. PyTorch input → /dx-agent-compiler-convert
2. ONNX ready → /dx-agent-compiler-compile
3. Always → /dx-agent-compiler-validate
4. pre_optimize requested, or PPU type 2 / ppu → pre_optimize migration → /dx-agent-compiler-pre-optimize (before compile)

Anti-Patterns (NEVER Do)

• Starting compilation without asking Q1-Q2
• Running onnx-simplifier without explicit user approval
• Enabling PPU without user opt-in
• Skipping pre-flight checks
• Generating config.json before inspecting ONNX input names