Goal: Progressive automation pipeline

Objective

A working pipeline where mechanical operations (scoring, validation, enrichment) run via scripts, inference-delegable operations (type classification, tag generation, description writing) run via local model, and both are accessible through MCP tools — so that any agent with MCP access can improve the repository without deep context.

Key results

1. score-file.py scores any file and is available via MCP
2. infer-triage-frontmatter.py enriches triage files via local model and is available via MCP
3. At least one non-triage improvement operation (infer-description, infer-tags) works on published files via local model
4. The MCP server exposes all mechanical and delegable operations (currently 9 tools; target: score + individual inference tools)
5. An agent can run improve-weakest-file in a directory and produce a measurable quality improvement without human guidance

Serving plans

• 0015 (MCP tooling)
• 0016 (improve-files skill family)

Constraints

• MCP server needs restart to pick up tools 8 and 9
• Ollama must be running for inference-based operations
• Local model quality is variable — verification scripts needed