drydock-cli 3.0.119__tar.gz → 3.0.121__tar.gz

{drydock_cli-3.0.119/drydock_cli.egg-info → drydock_cli-3.0.121}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.119
+Version: 3.0.121
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

drydock_cli-3.0.121/drydock/builtin_skills/ml-data.md

@@ -0,0 +1,15 @@
+---
+name: ml-data
+description: Organize, prep, split, normalize, or convert a dataset (incl. HDF5/JSON)
+---
+Data-engineering task: $ARGS
+
+1. INSPECT first — format (H5/CSV/JSON/images), shapes, dtypes, class balance
+   (`np.bincount`), missing values.
+2. Splits → stratified train/val/test; fit any scaler/encoder on TRAIN ONLY and apply
+   to val/test (NEVER fit on val/test — that leaks). Use a fixed random_state.
+3. HDF5 ↔ JSON → preserve shapes, dtypes, and attrs; `h5py` for arrays (compression for
+   big ones), `arr.tolist()` when a JSON needs the values; keep large arrays in H5 and
+   only metadata/paths in JSON.
+4. Save outputs EXACTLY as specified (names, keys, dtypes) and validate by reloading and
+   checking shapes/counts.

drydock_cli-3.0.121/drydock/builtin_skills/ml-debug.md

@@ -0,0 +1,18 @@
+---
+name: ml-debug
+description: Diagnose and fix a broken PyTorch training run (NaN loss, not learning, shapes)
+---
+Diagnose and FIX the broken training: $ARGS
+
+Be systematic — reproduce first, read the exact symptom, then:
+- Loss NaN/Inf → LR too high (lower it), log/sqrt of ≤0, divide-by-zero, unnormalized
+  inputs, or NaNs in data. Add `torch.autograd.set_detect_anomaly(True)` and
+  `nn.utils.clip_grad_norm_`.
+- Loss flat / not learning → missing `opt.zero_grad()` or `opt.step()`; wrong loss
+  (softmax BEFORE CrossEntropyLoss); LR too low; model not in `train()`; frozen params;
+  labels/logits misaligned. Print the loss and a param's grad-norm each step.
+- Shape mismatch → print `.shape` at each layer; CrossEntropyLoss wants logits (N,C) and
+  Long targets (N,).
+- Train↑ val↓ → overfitting; regularize / augment / early-stop.
+Overfit a tiny batch first to isolate the bug. Fix, rerun, and confirm it converges and
+meets the target before finishing.

drydock_cli-3.0.121/drydock/builtin_skills/ml-finetune.md

@@ -0,0 +1,17 @@
+---
+name: ml-finetune
+description: Fine-tune a model — full or LoRA (peft) — to a target metric
+---
+Fine-tuning task: $ARGS
+
+1. Load the base/pretrained model and inspect it (`model.named_modules()`), and load
+   the fine-tune data (with a held-out eval split).
+2. Choose the method:
+   - LoRA (train ~0.1% of params): `from peft import LoraConfig, get_peft_model`;
+     r=8, lora_alpha=16, lora_dropout=0.05, target_modules = the attention linear
+     layers you found; `get_peft_model(...)`; call `print_trainable_parameters()` to
+     CONFIRM only adapters train. Save with `model.save_pretrained('adapter')`.
+   - FULL: train all params with a SMALL LR (2e-5–5e-5), AdamW, a few epochs; freeze
+     anything that should stay fixed. `model.save_pretrained(dir)`.
+3. Train, evaluating on the held-out set; small LR + few epochs — don't destroy the
+   pretrained features. Report the eval metric and save exactly what's requested.

drydock_cli-3.0.121/drydock/builtin_skills/ml-metrics.md

@@ -0,0 +1,16 @@
+---
+name: ml-metrics
+description: Compute a full classification metrics report (accuracy/precision/recall/F1/MCC/ROC-AUC/confusion)
+---
+Compute an evaluation-metrics report for: $ARGS
+
+1. Load the ground-truth labels and the predictions (and predicted probabilities /
+   scores if ROC-AUC is needed). Confirm shapes and the positive label / class order.
+2. Using scikit-learn compute: accuracy, precision, recall, F1, MCC, ROC-AUC, and the
+   confusion matrix. For binary, report the counts TP/FP/TN/FN
+   (`tn,fp,fn,tp = confusion_matrix(y_true,y_pred).ravel()`). For multiclass use
+   `average='macro'` (and per-class arrays) and `roc_auc_score(..., multi_class='ovr')`.
+3. Save the metrics as JSON with clear keys, and plot the ROC curve
+   (`roc_curve` → plot tpr vs fpr) to an image if asked.
+4. Print a readable table. Double-check definitions: precision=TP/(TP+FP),
+   recall=TP/(TP+FN), F1=2PR/(P+R), MCC uses all four cells.

drydock_cli-3.0.121/drydock/builtin_skills/ml-rl.md

@@ -0,0 +1,16 @@
+---
+name: ml-rl
+description: Implement and train a reinforcement-learning agent to solve an environment
+---
+Reinforcement-learning task: $ARGS
+
+1. Inspect the environment API — `reset()`, `step(action)` return values, the state
+   and action spaces. Look at the actual code before coding the agent.
+2. Choose the method:
+   - Small discrete states → tabular Q-learning: `Q=np.zeros((nS,nA))`; epsilon-greedy
+     action; Bellman update `Q[s,a]+=alpha*(r+gamma*Q[s2].max()-Q[s,a])`; decay epsilon.
+   - Large/continuous → a neural policy: DQN (replay buffer + target network) or
+     REINFORCE (maximize sum of log_prob(a)*return).
+3. Train until the average episode reward rises and the greedy policy solves the task.
+   alpha≈0.5, gamma≈0.95, eps 0.1–0.2 decaying. Save the policy/Q-table exactly as
+   asked and verify the greedy policy reaches the goal.

drydock_cli-3.0.121/drydock/builtin_skills/ml-train.md

@@ -0,0 +1,19 @@
+---
+name: ml-train
+description: Build and train a PyTorch model on a dataset to a target metric
+---
+You are doing a machine-learning TRAINING task: $ARGS
+
+Work in this order:
+1. INSPECT the data first — load it, print shapes, dtypes, number of classes, and
+   which splits exist (train/val/test). Never assume; look.
+2. BUILD an appropriate model in PyTorch — a CNN for images, an MLP/Transformer for
+   vectors/sequences. Keep it small unless the task demands otherwise.
+3. Write a CORRECT training loop: AdamW, CrossEntropyLoss on RAW logits (no softmax),
+   `opt.zero_grad()` every step, mini-batches via DataLoader, `model.train()` for
+   training and `model.eval()` + `torch.no_grad()` for evaluation, and move BOTH the
+   model and each batch to the same device. Seed torch for reproducibility.
+4. Sanity-check: overfit a single tiny batch to near-zero loss first.
+5. Track the train and validation metric each epoch; stop when the target is met.
+6. Produce EXACTLY the artifacts the task asks for (file names, formats, order), then
+   re-read the requirements and verify each one before you finish.

{drydock_cli-3.0.119 → drydock_cli-3.0.121}/drydock/recipes.py

@@ -274,6 +274,8 @@ def retrieve_recipes(task_text: str, k: int = 2) -> list[tuple[str, str]]:
     overlap. Empty if nothing scores — never inject irrelevant noise."""
     if not task_text:
         return []
+    if not isinstance(task_text, str):  # defensive: never crash the agent loop
+        task_text = str(task_text)
     words = set(_WORD.findall(task_text.lower()))
     scored = []
     for title, keywords, body in RECIPES:

{drydock_cli-3.0.119 → drydock_cli-3.0.121}/drydock/tools/__init__.py

@@ -750,6 +750,10 @@ def _capture_screen(out: str) -> str | None:
         elif sys.platform == "darwin":
             subprocess.run(["screencapture", "-x", out], capture_output=True, timeout=30)
         else:  # Linux/BSD — try grabbers in order of ubiquity (X11 + wayland)
+            if not (os.environ.get("DISPLAY") or os.environ.get("WAYLAND_DISPLAY")):
+                # Fail fast — a grabber on a headless box blocks until timeout.
+                return ("no display ($DISPLAY/$WAYLAND_DISPLAY unset) — cannot capture "
+                        "the screen on a headless server.")
             grabbers = (["scrot", "-o", out], ["import", "-window", "root", out],
                         ["gnome-screenshot", "-f", out], ["grim", out],
                         ["spectacle", "-b", "-n", "-o", out], ["maim", out])

{drydock_cli-3.0.119 → drydock_cli-3.0.121/drydock_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.119
+Version: 3.0.121
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.119 → drydock_cli-3.0.121}/drydock_cli.egg-info/SOURCES.txt

@@ -30,6 +30,12 @@ drydock/tool_registry.py
 drydock/tuning.py
 drydock/web.py
 drydock/builtin_skills/__init__.py
+drydock/builtin_skills/ml-data.md
+drydock/builtin_skills/ml-debug.md
+drydock/builtin_skills/ml-finetune.md
+drydock/builtin_skills/ml-metrics.md
+drydock/builtin_skills/ml-rl.md
+drydock/builtin_skills/ml-train.md
 drydock/builtin_skills/rmf-categorize.md
 drydock/builtin_skills/rmf-control.md
 drydock/builtin_skills/rmf-poam.md

{drydock_cli-3.0.119 → drydock_cli-3.0.121}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 # PyPI distribution name is drydock-cli (the established install name, continued
 # from the retired v2 fork); the import package + CLI command stay `drydock`.
 name = "drydock-cli"
-version = "3.0.119"
+version = "3.0.121"
 description = "Drydock — a local, provider-agnostic terminal coding agent for local LLMs"
 readme = "README.md"
 requires-python = ">=3.11"

{drydock_cli-3.0.119 → drydock_cli-3.0.121}/tests/test_recipes.py

@@ -53,3 +53,18 @@ def test_not_injected_when_disabled():
     cfg = {"model": "gemma4", "max_tokens": 8192, "context_limit": 65536, "recipes": False}
     sysp = _run_capture(cfg, "recover the deleted password from launchcode.txt forensic")
     assert "TECHNIQUE RECIPES" not in sysp
+
+
+def test_ml_skills_bundled():
+    from drydock.skills import load_skills
+    sk = load_skills(".")
+    for name in ("ml-train", "ml-metrics", "ml-finetune", "ml-debug", "ml-rl", "ml-data"):
+        assert name in sk, f"bundled ML skill /{name} missing"
+    assert "MCC" in sk["ml-metrics"].body or "mcc" in sk["ml-metrics"].body.lower()
+
+
+def test_retrieve_handles_non_string():
+    from drydock.recipes import retrieve_recipes, recipe_context
+    for v in (b"train a cnn", ["train", "cnn"], 5, None):
+        assert isinstance(retrieve_recipes(v), list)   # never raises
+    assert isinstance(recipe_context(123), str)

{drydock_cli-3.0.119 → drydock_cli-3.0.121}/tests/test_screenshot.py

@@ -57,3 +57,15 @@ def test_macos_uses_screencapture(monkeypatch):
     monkeypatch.setattr("subprocess.run", lambda a, **k: calls.setdefault("argv", a))
     tools._capture_screen("/tmp/x.png")
     assert calls["argv"][0] == "screencapture"
+
+
+def test_headless_linux_fails_fast(monkeypatch):
+    import time
+    monkeypatch.setattr(tools, "_IS_WINDOWS", False)
+    monkeypatch.setattr("sys.platform", "linux")
+    monkeypatch.delenv("DISPLAY", raising=False)
+    monkeypatch.delenv("WAYLAND_DISPLAY", raising=False)
+    t = time.time()
+    err = tools._capture_screen("/tmp/x.png")
+    assert time.time() - t < 1.0            # no 30s grabber hang
+    assert err and "display" in err.lower()

{drydock_cli-3.0.119 → drydock_cli-3.0.121}/tests/test_skills.py

@@ -55,7 +55,11 @@ def test_empty_body_skipped(tmp_path):
 def test_no_user_or_project_skills_yields_only_builtins(tmp_path):
     # With no user/project skills, only the bundled built-ins (RMF) are present.
     loaded = skills.load_skills(str(tmp_path))
-    assert set(loaded) == {"rmf-control", "rmf-categorize", "rmf-review", "rmf-poam", "stig-assess", "stig-remediate"}
+    assert set(loaded) == {
+        "rmf-control", "rmf-categorize", "rmf-review", "rmf-poam",
+        "stig-assess", "stig-remediate",
+        "ml-train", "ml-metrics", "ml-finetune", "ml-debug", "ml-rl", "ml-data",
+    }
 
 
 def test_create_skill_user_scope(tmp_path, monkeypatch):