build-subgraph

build-subgraph

Turns a target node id, a depth, and a full graph into a bounded subgraph: the target, its dependency neighborhood out to that depth, and any claim nodes attached to it that disagree with each other — nothing else. The boundedness is the point of this skill, not a side effect, so the output must state how much of the full graph it left out, not just imply it.

Instructions

You are a Claude Code skill implementing the task-scoped-retrieval pattern. Follow these steps in order:

1. Read the full graph first. Open sample-graph.example.json (or the file the user names) and parse its nodes and edges arrays. Record the full node count and full edge count before doing anything else — this is the baseline the boundedness check compares against later.
2. Resolve the target and depth. Take the target node id and depth from the user's request. Default to target normalize_currency_code and depth 1 if the user doesn't specify either — this kit's shipped scenario. Confirm the target id actually exists among the graph's nodes; if it doesn't, stop and report that the target was not found rather than guessing a near match.
3. Build the depth-bounded dependency set. Starting from the target, follow calls edges outward (both directions — subject-to-object as "calls", object-to-subject as "called_by") for exactly depth hops. At depth 1, this is the target's direct callers and direct callees only — not their callers or callees in turn.
4. Separately, pull every claim attached to the target. Ignoring depth entirely for this step, find each node linked to the target through a claims edge — there may be one, several, or none, and they may or may not agree with each other. Then find every contradicts (or similarly flagged conflict) edge running between those claim nodes. Do not pick one claim as the "right" one and drop the rest — carry every attached claim and every conflict edge between them into the result.
5. Assemble the subgraph. The result's node set is: the target, the depth-bounded dependency set from step 3, and the claim nodes from step 4. The result's edge set is exactly the edges from the full graph whose subject and object are both in that node set — no edge gets included on its own if one of its endpoints didn't make the cut.
6. Prove the boundary, don't just assert it — this step must never be skipped. Compare the full graph's node count (step 1) against the subgraph's node count (step 5). List, or at minimum count, which node ids from the full graph were excluded. If the full graph has any node outside the boundary and the subgraph's node count still equals the full graph's, treat that as a signal the traversal is wrong — a "subgraph" that isn't smaller than the graph it was drawn from hasn't scoped anything, and you should re-check steps 3-5 before returning a result.
7. Write the result to output.json (or the path the user requested) in this kit's root, with this shape: target, depth, nodes (the included node objects), edges (the included edge objects), full_graph_node_count, subgraph_node_count, and excluded_node_ids (every node id in the full graph not present in nodes).
8. Report a summary alongside the file: the target, the depth used, how many nodes/edges were included versus the full graph's totals, and which (if any) claim nodes were pulled in for disagreement rather than depth.