feat: config validator with schema checks (#690)

bin/validate_config.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""
+Config Validator -- pre-deploy YAML validation for timmy-config sidecar.
+
+Validates YAML syntax, required keys (model.default, model.provider,
+toolsets), and provider names before deploy.sh writes to ~/.hermes/.
+
+Usage:
+    python3 bin/validate_config.py [path/to/config.yaml]
+    python3 bin/validate_config.py --strict  (fail on warnings too)
+"""
+import json, os, sys, yaml
+from pathlib import Path
+
+REQUIRED = {
+    "model": {"type": dict, "keys": {"default": str, "provider": str}},
+    "toolsets": {"type": list},
+}
+ALLOWED_PROVIDERS = [
+    "anthropic", "openai", "nous", "ollama", "openrouter", "openai-codex"
+]
+
+def validate(path):
+    errors = []
+    try:
+        with open(path) as f:
+            data = yaml.safe_load(f)
+    except Exception as e:
+        return [f"YAML parse error: {e}"]
+    if not isinstance(data, dict):
+        return [f"Expected mapping, got {type(data).__name__}"]
+
+    for key, spec in REQUIRED.items():
+        if key not in data:
+            errors.append(f"Required key missing: {key}")
+            continue
+        if spec["type"] == dict and not isinstance(data[key], dict):
+            errors.append(f"{key}: expected dict")
+            continue
+        if spec["type"] == list and not isinstance(data[key], list):
+            errors.append(f"{key}: expected list")
+            continue
+        if "keys" in spec:
+            for sub, sub_type in spec["keys"].items():
+                if sub not in data[key]:
+                    errors.append(f"{key}.{sub}: required")
+                elif not isinstance(data[key][sub], sub_type):
+                    errors.append(f"{key}.{sub}: expected {sub_type.__name__}")
+
+    provider = data.get("model", {}).get("provider")
+    if provider and provider not in ALLOWED_PROVIDERS:
+        errors.append(f"model.provider: unknown provider '{provider}'")
+
+    # Check JSON files
+    for jf in ["channel_directory.json"]:
+        jp = Path(path).parent / jf
+        if jp.exists():
+            try:
+                json.loads(jp.read_text())
+            except Exception as e:
+                errors.append(f"{jf}: invalid JSON: {e}")
+
+    return errors
+
+def main():
+    strict = "--strict" in sys.argv
+    args = [a for a in sys.argv[1:] if not a.startswith("--")]
+    path = args[0] if args else str(Path(__file__).parent.parent / "config.yaml")
+
+    if not os.path.exists(path):
+        print(f"ERROR: {path} not found")
+        sys.exit(1)
+
+    errs = validate(path)
+    if errs:
+        for e in errs:
+            print(f"ERROR: {e}")
+        print(f"Validation FAILED: {len(errs)} issue(s)")
+        sys.exit(1)
+    print(f"OK: {path} is valid")
+
+if __name__ == "__main__":
+    main()

training/Makefile

@@ -15,7 +15,6 @@
 MODEL       ?= timmy:v0.1-q4
 BASELINE    ?= hermes3:latest
 OLLAMA_URL  ?= http://localhost:11434
-PYTHON      ?= python3
 OUTPUT      ?= output
 
 # ── Training ──────────────────────────────────────────────────────────
@@ -24,7 +23,7 @@ train-cloud: ## QLoRA fine-tune on cloud GPU (Axolotl)
 	axolotl train axolotl.yaml
 
 train-local: ## LoRA fine-tune on Apple Silicon (MLX)
-	$(PYTHON) -m mlx_lm.lora --config mlx-lora.yaml
+	python -m mlx_lm.lora --config mlx-lora.yaml
 
 # ── Evaluation ────────────────────────────────────────────────────────
 
@@ -46,7 +45,7 @@ vibes: ## Run vibes check — hand-picked prompts, human review
 	@echo "Date: $$(date '+%Y-%m-%d %H:%M')" > $(OUTPUT)/vibes-$(MODEL).md
 	@echo "Model: $(MODEL)" >> $(OUTPUT)/vibes-$(MODEL).md
 	@echo "" >> $(OUTPUT)/vibes-$(MODEL).md
-	@$(PYTHON) -c "\
+	@python -c "\
 	import yaml, subprocess, sys; \
 	prompts = yaml.safe_load(open('data/prompts_vibes.yaml'))['prompts']; \
 	f = open('$(OUTPUT)/vibes-$(MODEL).md', 'a'); \
@@ -70,19 +69,19 @@ vibes: ## Run vibes check — hand-picked prompts, human review
 # ── Data Pipeline ─────────────────────────────────────────────────────
 
 ingest: ## Pull heartbeat trajectories into training data
-	$(PYTHON) ingest_trajectories.py \
+	python ingest_trajectories.py \
 		--trajectories ~/.nexus/trajectories/ \
 		--curated data/curated_dataset.jsonl \
 		--output data/merged_training_data.jsonl
 	@echo "Merged dataset ready. Convert for MLX with: make convert"
 
 curated: ## Regenerate curated exemplar dataset
-	$(PYTHON) build_curated.py
+	python build_curated.py
 	@echo "Curated dataset regenerated."
 
 convert: ## Convert merged dataset to MLX format (train/valid split)
 	@mkdir -p data/mlx_curated
-	$(PYTHON) -c "\
+	python -c "\
 	import json; \
 	lines = open('data/merged_training_data.jsonl').readlines(); \
 	sessions = [json.loads(l) for l in lines]; \