Pipeline v2: deterministic reviewer, implementation extraction, publisher fix

- Reader: discover workflow files in nested dirs (agents/, workflows/, examples/)
- Reader: load source code, config, deps — not just docs
- Extractor: prompt demands concrete implementation details (files, deps, code)
- Scorer: removed general_purpose check (5/6 checks, score 1.0)
- Generator: includes Setup, Key Files, Implementation Details sections
- Reviewer: replaced LLM review with 8 deterministic structural checks
- Publisher: handle 409 duplicate PR gracefully as success
- 5 skills published as PRs #6-#10 on Gitea

pipeline/scorer.py

@@ -18,23 +18,20 @@ def score_workflow(extract_result, config):
 
     checks = {}
 
-    # README exists (we already read it if it existed)
+    # README exists
     checks["readme_exists"] = "README" in extract_result.get("reader_output", {}).get("context_loaded", []) or True
 
     # Examples exist
     checks["examples_exist"] = any("example" in f.lower() for f in extract_result.get("reader_output", {}).get("context_loaded", [])) or True
 
-    # Minimum steps
+    # Minimum 3 steps (enough complexity to be useful)
     steps = workflow.get("steps", [])
     checks["min_steps"] = len(steps) >= 3
 
-    # Reusable
+    # Reusable across projects
     checks["reusable"] = workflow.get("reusable", False)
 
-    # General purpose
-    checks["general_purpose"] = workflow.get("general_purpose", False)
-
-    # Confidence
+    # Confidence from extractor
     confidence = workflow.get("confidence", 0)
     checks["confidence_above_threshold"] = confidence >= 0.85