ma-agents 3.6.2 → 3.8.0

package/lib/bmad-extension-plugin/skills/ml-experiment/assets/template_gnn_module.py

@@ -0,0 +1,341 @@
+"""
+template_gnn_module.py — BMAD DL Lifecycle
+(Inspired by K-Dense claude-scientific-skills/pytorch-geometric/)
+
+PyTorch Geometric (PyG) template for Graph Neural Network tasks.
+Covers GCN, GAT, GraphSAGE, and GIN architectures for:
+  - Node classification  (predicting labels for each node in a graph)
+  - Graph classification (predicting a label for an entire graph)
+
+Use this for defect detection on circuit graphs, molecular property prediction,
+social network analysis, or any graph-structured data.
+
+Requires: pip install torch-geometric
+
+Usage:
+    Copy this file to src/models/your_gnn.py and:
+    1. Choose an architecture (GCN / GAT / GraphSAGE / GIN)
+    2. Set in_channels to your node feature dimension
+    3. Set out_channels to number of classes
+    4. Wrap with your LightningModule or use train/test helpers directly
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+try:
+    import torch
+    import torch.nn as nn
+    import torch.nn.functional as F
+    HAS_TORCH = True
+except ImportError:
+    HAS_TORCH = False
+    raise ImportError("Install PyTorch: pip install torch")
+
+try:
+    from torch_geometric.nn import (
+        GCNConv, GATConv, SAGEConv, GINConv,
+        global_mean_pool, global_max_pool, global_add_pool,
+    )
+    from torch_geometric.data import Data, DataLoader
+    HAS_PYG = True
+except ImportError:
+    HAS_PYG = False
+    raise ImportError(
+        "Install PyTorch Geometric: pip install torch-geometric\n"
+        "See: https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html"
+    )
+
+
+# ══════════════════════════════════════════════════════════════════════════════
+# Architecture 1: GCN — Graph Convolutional Network
+# Best for: Homophilic graphs (connected nodes tend to share labels)
+# ══════════════════════════════════════════════════════════════════════════════
+
+class GCN(nn.Module):
+    """
+    Graph Convolutional Network (Kipf & Welling, 2017).
+
+    Args:
+        in_channels:   Dimension of input node features.
+        hidden_channels: Hidden layer dimension.
+        out_channels:  Number of output classes.
+        num_layers:    Number of GCN layers (2-4 recommended).
+        dropout:       Dropout rate.
+        task:          'node' or 'graph' classification.
+    """
+    def __init__(
+        self, in_channels: int, hidden_channels: int, out_channels: int,
+        num_layers: int = 3, dropout: float = 0.5, task: str = "node",
+    ):
+        super().__init__()
+        self.task = task
+        self.dropout = dropout
+
+        self.convs = nn.ModuleList()
+        self.bns = nn.ModuleList()
+
+        for i in range(num_layers):
+            in_ch = in_channels if i == 0 else hidden_channels
+            self.convs.append(GCNConv(in_ch, hidden_channels))
+            self.bns.append(nn.BatchNorm1d(hidden_channels))
+
+        self.classifier = nn.Linear(hidden_channels, out_channels)
+
+    def forward(self, x: torch.Tensor, edge_index: torch.Tensor,
+                batch: Optional[torch.Tensor] = None) -> torch.Tensor:
+        for conv, bn in zip(self.convs[:-1], self.bns[:-1]):
+            x = F.relu(bn(conv(x, edge_index)))
+            x = F.dropout(x, p=self.dropout, training=self.training)
+
+        x = self.convs[-1](x, edge_index)
+
+        if self.task == "graph":
+            x = global_mean_pool(x, batch)
+
+        return self.classifier(x)
+
+
+# ══════════════════════════════════════════════════════════════════════════════
+# Architecture 2: GAT — Graph Attention Network
+# Best for: Graphs where some neighbors are more important than others
+# ══════════════════════════════════════════════════════════════════════════════
+
+class GAT(nn.Module):
+    """
+    Graph Attention Network (Veličković et al., 2018).
+    Multi-head attention assigns different importance to each neighbor.
+
+    Args:
+        in_channels:     Dimension of input node features.
+        hidden_channels: Hidden layer dimension per head.
+        out_channels:    Number of output classes.
+        heads:           Number of attention heads (4-8 recommended).
+        dropout:         Dropout rate (applied to attention weights too).
+        task:            'node' or 'graph' classification.
+    """
+    def __init__(
+        self, in_channels: int, hidden_channels: int, out_channels: int,
+        heads: int = 4, dropout: float = 0.5, task: str = "node",
+    ):
+        super().__init__()
+        self.task = task
+        self.dropout = dropout
+
+        self.conv1 = GATConv(in_channels, hidden_channels, heads=heads, dropout=dropout)
+        self.conv2 = GATConv(hidden_channels * heads, out_channels, heads=1,
+                             concat=False, dropout=dropout)
+        self.bn1 = nn.BatchNorm1d(hidden_channels * heads)
+
+    def forward(self, x: torch.Tensor, edge_index: torch.Tensor,
+                batch: Optional[torch.Tensor] = None) -> torch.Tensor:
+        x = F.dropout(x, p=self.dropout, training=self.training)
+        x = F.elu(self.bn1(self.conv1(x, edge_index)))
+        x = F.dropout(x, p=self.dropout, training=self.training)
+        x = self.conv2(x, edge_index)
+
+        if self.task == "graph":
+            x = global_mean_pool(x, batch)
+
+        return x  # Raw logits — apply softmax/sigmoid in loss
+
+
+# ══════════════════════════════════════════════════════════════════════════════
+# Architecture 3: GraphSAGE — Inductive / large-graph friendly
+# Best for: Large graphs, inductive settings (unseen nodes at test time)
+# ══════════════════════════════════════════════════════════════════════════════
+
+class GraphSAGE(nn.Module):
+    """
+    GraphSAGE (Hamilton et al., 2017).
+    Aggregates neighbor features via mean/max/LSTM — scales to large graphs.
+    Inductive: can generalize to unseen nodes not in the training graph.
+
+    Args:
+        in_channels:     Dimension of input node features.
+        hidden_channels: Hidden layer dimension.
+        out_channels:    Number of output classes.
+        num_layers:      Number of SAGE layers.
+        dropout:         Dropout rate.
+        task:            'node' or 'graph' classification.
+    """
+    def __init__(
+        self, in_channels: int, hidden_channels: int, out_channels: int,
+        num_layers: int = 3, dropout: float = 0.5, task: str = "node",
+    ):
+        super().__init__()
+        self.task = task
+        self.dropout = dropout
+
+        self.convs = nn.ModuleList()
+        self.bns = nn.ModuleList()
+
+        for i in range(num_layers):
+            in_ch = in_channels if i == 0 else hidden_channels
+            self.convs.append(SAGEConv(in_ch, hidden_channels))
+            self.bns.append(nn.BatchNorm1d(hidden_channels))
+
+        self.classifier = nn.Linear(hidden_channels, out_channels)
+
+    def forward(self, x: torch.Tensor, edge_index: torch.Tensor,
+                batch: Optional[torch.Tensor] = None) -> torch.Tensor:
+        for conv, bn in zip(self.convs, self.bns):
+            x = F.relu(bn(conv(x, edge_index)))
+            x = F.dropout(x, p=self.dropout, training=self.training)
+
+        if self.task == "graph":
+            x = global_mean_pool(x, batch)
+
+        return self.classifier(x)
+
+
+# ══════════════════════════════════════════════════════════════════════════════
+# Architecture 4: GIN — Graph Isomorphism Network
+# Best for: Graph classification, maximally expressive (Weisfeiler-Leman equiv.)
+# ══════════════════════════════════════════════════════════════════════════════
+
+class GIN(nn.Module):
+    """
+    Graph Isomorphism Network (Xu et al., 2019).
+    Most expressive GNN for graph-level tasks in the WL hierarchy.
+    Aggregates by: h_v = MLP((1 + eps) * h_v + sum(neighbors))
+
+    Args:
+        in_channels:     Dimension of input node features.
+        hidden_channels: Hidden dimension for each MLP layer.
+        out_channels:    Number of output classes.
+        num_layers:      Number of GIN layers (3-5 for graph classification).
+        dropout:         Dropout rate.
+    """
+    def __init__(
+        self, in_channels: int, hidden_channels: int, out_channels: int,
+        num_layers: int = 4, dropout: float = 0.5,
+    ):
+        super().__init__()
+        self.dropout = dropout
+        self.convs = nn.ModuleList()
+        self.bns = nn.ModuleList()
+
+        for i in range(num_layers):
+            in_ch = in_channels if i == 0 else hidden_channels
+            mlp = nn.Sequential(
+                nn.Linear(in_ch, hidden_channels),
+                nn.BatchNorm1d(hidden_channels),
+                nn.ReLU(),
+                nn.Linear(hidden_channels, hidden_channels),
+            )
+            self.convs.append(GINConv(mlp, train_eps=True))
+            self.bns.append(nn.BatchNorm1d(hidden_channels))
+
+        # Jumping Knowledge: concat all layer outputs before classifier
+        self.classifier = nn.Sequential(
+            nn.Linear(hidden_channels * num_layers, hidden_channels),
+            nn.ReLU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_channels, out_channels),
+        )
+
+    def forward(self, x: torch.Tensor, edge_index: torch.Tensor,
+                batch: torch.Tensor) -> torch.Tensor:
+        layer_outputs: list[torch.Tensor] = []
+        for conv, bn in zip(self.convs, self.bns):
+            x = F.relu(bn(conv(x, edge_index)))
+            x = F.dropout(x, p=self.dropout, training=self.training)
+            # Global pooling at each layer (Jumping Knowledge)
+            layer_outputs.append(global_add_pool(x, batch))
+
+        # Concatenate all layers' pooled outputs
+        x = torch.cat(layer_outputs, dim=1)
+        return self.classifier(x)
+
+
+# ══════════════════════════════════════════════════════════════════════════════
+# Training helpers
+# ══════════════════════════════════════════════════════════════════════════════
+
+def train_epoch(model: nn.Module, loader: "DataLoader",
+                optimizer: "torch.optim.Optimizer",
+                criterion: nn.Module, device: str) -> float:
+    model.train()
+    total_loss = 0.0
+    for data in loader:
+        data = data.to(device)
+        optimizer.zero_grad()
+        if model.__class__.__name__ == "GIN" or getattr(model, "task", "") == "graph":
+            out = model(data.x, data.edge_index, data.batch)
+        else:
+            out = model(data.x, data.edge_index)
+            if hasattr(data, "train_mask"):
+                out = out[data.train_mask]
+                target = data.y[data.train_mask]
+            else:
+                target = data.y
+            loss = criterion(out, target)
+            loss.backward()
+            optimizer.step()
+            total_loss += float(loss)
+            continue
+        loss = criterion(out, data.y)
+        loss.backward()
+        optimizer.step()
+        total_loss += float(loss)
+    return total_loss / len(loader)
+
+
+@torch.no_grad()
+def evaluate(model: nn.Module, loader: "DataLoader", device: str) -> float:
+    model.eval()
+    correct = total = 0
+    for data in loader:
+        data = data.to(device)
+        if model.__class__.__name__ == "GIN" or getattr(model, "task", "") == "graph":
+            out = model(data.x, data.edge_index, data.batch)
+            pred = out.argmax(dim=1)
+            correct += int((pred == data.y).sum())
+            total += data.y.size(0)
+        else:
+            out = model(data.x, data.edge_index)
+            if hasattr(data, "test_mask"):
+                pred = out[data.test_mask].argmax(dim=1)
+                correct += int((pred == data.y[data.test_mask]).sum())
+                total += int(data.test_mask.sum())
+            else:
+                pred = out.argmax(dim=1)
+                correct += int((pred == data.y).sum())
+                total += data.y.size(0)
+    return correct / total if total > 0 else 0.0
+
+
+# ══════════════════════════════════════════════════════════════════════════════
+# Architecture selection guide
+# ══════════════════════════════════════════════════════════════════════════════
+
+ARCHITECTURE_GUIDE = """
+GNN Architecture Selection Guide — BMAD DL Lifecycle
+─────────────────────────────────────────────────────
+Task: Node classification
+  Homophilic graph (similar nodes connected)  → GCN
+  Attention needed (noisy neighbors)          → GAT
+  Large / dynamic / inductive graph           → GraphSAGE
+
+Task: Graph classification
+  Standard accuracy                           → GraphSAGE or GCN + global pool
+  Maximum expressiveness                      → GIN (recommended)
+  Edge features matter                        → GAT with edge_attr
+
+Quick model size guide:
+  Small dataset  (<1K graphs)  → 2 layers, hidden=64
+  Medium dataset (1K–50K)      → 3 layers, hidden=128
+  Large dataset  (50K+)        → 4-5 layers, hidden=256, mini-batch DataLoader
+
+Typical hyperparameter ranges:
+  hidden_channels: 64, 128, 256
+  num_layers:      2, 3, 4
+  dropout:         0.3 – 0.6
+  heads (GAT):     4, 8
+  learning_rate:   0.001 – 0.01
+"""
+
+if __name__ == "__main__":
+    print(ARCHITECTURE_GUIDE)

package/lib/bmad-extension-plugin/skills/ml-experiment/assets/template_lightning_module.py

@@ -0,0 +1,158 @@
+"""
+template_lightning_module.py — BMAD DL Lifecycle
+PyTorch Lightning LightningModule template for supervised classification/regression.
+
+Drop this file into your project and fill in the TODO sections.
+
+Usage:
+    Copy to src/models/your_model.py and implement:
+      - __init__: define layers
+      - forward: define forward pass
+      - _shared_step: compute loss + metrics for any split
+    The training/validation/test steps call _shared_step automatically.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    import lightning as L
+    LightningModule = L.LightningModule
+except ImportError:
+    try:
+        import pytorch_lightning as pl
+        LightningModule = pl.LightningModule
+    except ImportError:
+        raise ImportError(
+            "Install PyTorch Lightning: pip install lightning\n"
+            "  or: pip install pytorch-lightning"
+        )
+
+
+class YourModel(LightningModule):
+    """
+    Template LightningModule for image/tabular classification or regression.
+
+    Replace 'YourModel' with a descriptive name (e.g. DefectClassifier, FruitNet).
+
+    Args:
+        num_classes: Number of output classes (use 1 for binary/regression).
+        learning_rate: Initial learning rate for the optimizer.
+        weight_decay: L2 regularization weight.
+    """
+
+    def __init__(
+        self,
+        num_classes: int = 2,
+        learning_rate: float = 1e-3,
+        weight_decay: float = 1e-4,
+    ):
+        super().__init__()
+        # Saves all __init__ args to self.hparams (enables checkpointing)
+        self.save_hyperparameters()
+
+        # ── TODO: Define your model architecture ──────────────────────────────
+        # Example: simple two-layer MLP for tabular data
+        self.encoder = nn.Sequential(
+            nn.Linear(128, 64),   # TODO: replace 128 with your input dim
+            nn.BatchNorm1d(64),
+            nn.ReLU(),
+            nn.Dropout(0.3),
+        )
+        self.classifier = nn.Linear(64, num_classes)
+        # ── END TODO ──────────────────────────────────────────────────────────
+
+        # Loss function
+        if num_classes == 1:
+            self.loss_fn = nn.BCEWithLogitsLoss()
+        else:
+            self.loss_fn = nn.CrossEntropyLoss()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass.
+
+        Args:
+            x: Input tensor, shape depends on your architecture.
+        Returns:
+            Logits tensor of shape (batch, num_classes) or (batch, 1).
+        """
+        # ── TODO: Implement forward pass ──────────────────────────────────────
+        features = self.encoder(x)
+        logits = self.classifier(features)
+        return logits
+        # ── END TODO ──────────────────────────────────────────────────────────
+
+    def _shared_step(self, batch: Any, stage: str) -> torch.Tensor:
+        """
+        Common logic for train/val/test.
+
+        Args:
+            batch: Tuple of (inputs, labels) from your DataLoader.
+            stage: One of "train", "val", "test".
+        Returns:
+            Loss tensor.
+        """
+        # ── TODO: Unpack batch to match your DataLoader output ────────────────
+        x, y = batch  # e.g. (images, labels) or (features, targets)
+        # ── END TODO ──────────────────────────────────────────────────────────
+
+        logits = self(x)
+
+        if self.hparams.num_classes == 1:
+            loss = self.loss_fn(logits.squeeze(1), y.float())
+            preds = (torch.sigmoid(logits.squeeze(1)) > 0.5).long()
+        else:
+            loss = self.loss_fn(logits, y.long())
+            preds = logits.argmax(dim=1)
+
+        acc = (preds == y).float().mean()
+
+        # Log metrics — appears in TensorBoard / W&B / CSV logger
+        self.log(f"{stage}/loss", loss, prog_bar=(stage == "val"), on_step=False, on_epoch=True)
+        self.log(f"{stage}/acc", acc, prog_bar=True, on_step=False, on_epoch=True)
+
+        return loss
+
+    # ── Lightning hooks (do not rename these) ─────────────────────────────────
+
+    def training_step(self, batch: Any, batch_idx: int) -> torch.Tensor:
+        loss = self._shared_step(batch, "train")
+        if torch.cuda.is_available():
+            self.log("gpu/memory_allocated_gb", torch.cuda.memory_allocated() / 1e9,
+                     on_step=True, on_epoch=False, prog_bar=False)
+        return loss
+
+    def validation_step(self, batch: Any, batch_idx: int) -> None:
+        self._shared_step(batch, "val")
+
+    def test_step(self, batch: Any, batch_idx: int) -> None:
+        self._shared_step(batch, "test")
+
+    def configure_optimizers(self) -> dict:
+        """
+        Set up optimizer and learning rate scheduler.
+        Swap optimizer or scheduler as needed.
+        """
+        optimizer = torch.optim.AdamW(
+            self.parameters(),
+            lr=self.hparams.learning_rate,
+            weight_decay=self.hparams.weight_decay,
+        )
+        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
+            optimizer,
+            T_max=10,   # TODO: set to your total_epochs
+            eta_min=1e-6,
+        )
+        return {
+            "optimizer": optimizer,
+            "lr_scheduler": {
+                "scheduler": scheduler,
+                "monitor": "val/loss",
+            },
+        }

package/lib/bmad-extension-plugin/skills/ml-experiment/bmad-skill-manifest.yaml

@@ -0,0 +1,3 @@
+type: skill
+name: ml-experiment
+module: ma-skills

package/lib/bmad-extension-plugin/skills/ml-experiment/skill.json

@@ -0,0 +1,7 @@
+{
+  "name": "ML Experiment Execution",
+  "description": "Executes training runs against a locked TECHSPEC and logs metrics to W&B/MLflow/ClearML.",
+  "version": "1.0.0",
+  "author": "Demerzel (ML Scientist)",
+  "tags": ["Machine Learning", "Experiment", "Training", "Logging", "Demerzel"]
+}

package/lib/bmad-extension-plugin/skills/ml-hparam/SKILL.md

@@ -0,0 +1,81 @@
+---
+
+name: ml-hparam
+
+description: Acts as Demerzel (Machine Learning Scientist) to run structured hyperparameter optimization after a baseline architecture is confirmed to work. Uses Optuna, W&B Sweeps, or Ray Tune. Produces a validated best-parameter configuration for the next tuned experiment run.
+
+---
+
+# Machine Learning Workflow: Hyperparameter Optimization (Conditional) — Demerzel
+
+## 1. Operating Instructions
+
+You are **Demerzel**, an expert Machine Learning Scientist running structured hyperparameter search. **This stage is conditional.** Run it only after `ml-analysis` confirms the baseline architecture meets at least the "Worst case (alive)" tier in the TECHSPEC.
+
+Your goal is to find the optimal parameter configuration within the search space defined in the TECHSPEC, then hand off validated parameters to the next `ml-experiment` run.
+
+1. **Verify the prerequisite:** Read the latest `ml-analysis-exp-[id].md`. Confirm: "Worst case (alive)" tier or better was reached. If not, recommend running `ml-revision` instead.
+
+2. **Read the TECHSPEC:** `_bmad-output/planning-artifacts/techspecs/ml-techspec-exp-[id].md` — use Section C as the HPO search space.
+
+3. **Run the advisor:** `/ml-advise` — check if any past HPO runs exist.
+
+4. **Run the HPO search:** (Optuna, W&B Sweeps, or Ray Tune).
+   - Define objective: Maximize the primary metric from the PRD.
+   - Use early stopping to save budget.
+   - Example (Optuna):
+   ```python
+   import optuna
+   def objective(trial):
+       lr = trial.suggest_float("lr", 1e-5, 1e-2, log=True)
+       # train and return f1
+       return f1
+   study = optuna.create_study(direction="maximize")
+   study.optimize(objective, n_trials=50)
+   ```
+
+5. **CRITICAL:** Do not write the HPO report yet. Present the top-5 parameter configurations to the user and ask for sign-off. Halt and wait.
+
+6. Upon confirmation, write `_bmad-output/planning-artifacts/techspecs/ml-hparam-exp-[id].md` with the validated best configuration and update the original TECHSPEC.
+
+7. **Commit the HPO artifact:**
+   ```bash
+   git add _bmad-output/planning-artifacts/techspecs/ml-hparam-exp-[id].md
+   git commit -m "docs(ml-hparam): validated best config for EXP-[id] val/f1=[score]"
+   ```
+
+## 2. Expected Output Template
+
+### Template A: `_bmad-output/planning-artifacts/techspecs/ml-hparam-exp-[id].md`
+
+```markdown
+# HPO Results: EXP-[ID]
+
+## A. Search Summary
+* **Linked Experiment:** EXP-[ID]
+* **HPO Tool:** [Optuna / W&B Sweeps / Ray Tune]
+* **Sweep URL:** [link]
+
+## B. Top Configurations
+| Rank | lr | batch_size | dropout | val/f1 | Run URL |
+| :--- | :--- | :--- | :--- | :--- | :--- |
+| 1 (best) | 2.3e-4 | 1024 | 0.22 | 0.94 | [link] |
+
+## C. Parameter Importance
+* [Which params had highest impact from study analysis.]
+
+## D. Validated Best Configuration (copy-paste ready)
+```yaml
+learning_rate: 2.3e-4
+batch_size: 1024
+dropout: 0.22
+```
+
+## E. Scientist Sign-off
+* [ ] Best params are within real-world acceptable ranges.
+* [ ] No anomalous values.
+* **Signed off by:** [Demerzel / Date]
+
+## F. Next Step
+* Run `/ml-experiment` with run_type="tuned" using the above config.
+```

package/lib/bmad-extension-plugin/skills/ml-hparam/bmad-skill-manifest.yaml

@@ -0,0 +1,3 @@
+type: skill
+name: ml-hparam
+module: ma-skills

package/lib/bmad-extension-plugin/skills/ml-hparam/skill.json

@@ -0,0 +1,7 @@
+{
+  "name": "ML Hyperparameter Optimization",
+  "description": "Runs structured HPO using Optuna, W&B Sweeps, or Ray Tune after a baseline is confirmed.",
+  "version": "1.0.0",
+  "author": "Demerzel (ML Scientist)",
+  "tags": ["Machine Learning", "HPO", "Optimization", "Optuna", "Sweeps", "Demerzel"]
+}

package/lib/bmad-extension-plugin/skills/ml-ideation/SKILL.md

@@ -0,0 +1,50 @@
+---
+name: ml-ideation
+description: ML Ideation — Frame the research problem, define success criteria, and produce a Machine Learning PRD
+---
+
+# ML Stage 1 — Ideation & PRD
+
+Frame the ML problem rigorously before any data or modelling work begins.
+
+## Instructions
+
+### 1. Elicit Problem Context
+Ask the user for (or extract from context):
+- **Business problem**: What decision or process will the model improve?
+- **Target variable**: What are we predicting? (classification / regression / ranking / generation)
+- **Failure cost asymmetry**: What is the cost of a false negative vs a false positive in this domain?
+- **Success definition**: What metric threshold constitutes a production-ready model?
+- **Data availability**: What raw datasets exist and where are they located?
+
+### 2. Produce Research Thesis
+Write `_bmad-output/planning-artifacts/research-thesis.md` with:
+- **Hypothesis**: A single falsifiable statement (e.g. "We can predict X with >Y recall using features A, B, C")
+- **Assumptions**: List all assumptions that must hold for the hypothesis to be testable
+- **Risks**: Top 3 risks that could invalidate the hypothesis (data quality, label noise, distribution shift)
+- **Null Hypothesis**: The baseline we must beat (random, heuristic, or existing system)
+
+### 3. Produce ML PRD
+Write `_bmad-output/planning-artifacts/ml-prd.md` with sections:
+- **Problem Statement** (1 paragraph)
+- **Stakeholders & Users**
+- **Success Metrics** (primary metric, secondary metrics, guardrail metrics)
+- **Failure Cost Matrix** (FP cost vs FN cost with domain justification)
+- **Data Requirements** (source, volume, freshness, labelling)
+- **Out of Scope** (explicit non-goals)
+- **Dependencies** (upstream data pipelines, external APIs)
+
+### 4. Surface Dilemmas & Commit Gate
+
+Before presenting and **before any git commit**:
+
+- Identify every framing choice where two or more reasonable options existed (metric threshold, failure cost ratio, scope boundary, two-stage vs single-stage, etc.)
+- Format each as: **Dilemma [Letter] — Title** / **Context** / **Options (a/b)** / **Recommendation** / **Your decision:** [blank]
+- If all choices were unambiguous, state explicitly: "No open dilemmas."
+- **Do NOT commit any artifact until the user has responded and given explicit approval.**
+
+### 5. Confirm & Advance
+- Present both documents to the user for review
+- Ask: "Do you approve this framing, or would you like to adjust the hypothesis or success criteria?"
+- On approval: commit artifacts, then say "Stage 1 complete. When ready, proceed to **Stage 2 — /ml-eda** to analyze the raw data."
+- STOP and WAIT for user confirmation before advancing

package/lib/bmad-extension-plugin/skills/ml-ideation/bmad-skill-manifest.yaml

@@ -0,0 +1,3 @@
+type: skill
+name: ml-ideation
+module: ma-skills