query-graph

query-graph

Turns schema.sql into a queryable graph.db file, then answers a question against the two tables it defines — an availability check on a tool, or a provenance lookup on a specific edge — by running SQL, not by re-reasoning over the Elm Street chat from scratch each time.

Instructions

You are a Claude Code skill implementing the sqlite-backed-graph pattern. Follow these steps in order:

1. Load the database if it isn't already loaded. Check whether graph.db exists in this kit's root. If it doesn't, build it by running schema.sql against a fresh file: sqlite3 graph.db < schema.sql, or, if the sqlite3 CLI isn't available, the Python equivalent: python3 -c "import sqlite3; sqlite3.connect('graph.db').executescript(open('schema.sql').read())". If graph.db already exists, leave it as is rather than reloading it — this step must be safe to skip on a second run.
2. Turn on foreign key enforcement for this connection. SQLite does not enforce REFERENCES constraints unless the connection asks for it, so every query in this skill must run PRAGMA foreign_keys = ON; first, even against an already-loaded graph.db.
3. Resolve the question. Default to this kit's worked question if the user doesn't name one: check whether the stud finder can be borrowed right now, and confirm the ladder's status too — which means running the availability query from step 4 against both tool-ladder and tool-stud-finder.
4. Run the availability query per tool. For a given tool id, find its most recent borrowed edge by id (edge ids climb in the same order the source messages arrived, so the highest id is the latest borrowed edge):

   SELECT id, from_id FROM edges
   WHERE label = 'borrowed' AND to_id = :tool_id
   ORDER BY id DESC LIMIT 1;
   

   If no row comes back, the tool has never left its owns edge and is available. If a row comes back, check for a matching returned edge with a higher id:

   SELECT id FROM edges
   WHERE label = 'returned' AND to_id = :tool_id AND from_id = :borrower
     AND id > :borrowed_edge_id;
   

   A match means available (the tool came back); no match means still out with :borrower. Name the specific edge id(s) that produced the verdict — never report a verdict without the row backing it.
5. Run the provenance query on request, or on the edge that grounded step 4's verdict. For a specific edge id:

   SELECT source_doc, extraction_run_id, schema_version FROM edges
   WHERE id = :edge_id;
   

   This is the receipt-per-edge fields doing their job here: naming which source document, which extraction run, and which schema version produced the edge the verdict rests on.
6. Write the result to output.json (or the path the user requested) in this kit's root, with this shape: queries (an array, one entry per tool or edge checked, each carrying the SQL text run and its raw result), and verdicts (one entry per tool, with tool, available, grounding_edge_id, and — if not available — held_by).
7. Report a summary alongside the file: each tool's verdict, the edge id(s) it rests on, and that verdict's provenance (source document, run id, schema version) via step 5's lookup.