context-graph-mining

name

context-graph-mining

description

INVOKE THIS SKILL when mining the procurement-agent Arize project for patterns in agent decisions vs human overrides, building a context graph, and proposing updates to procurement-agent rules. Use after a batch of reviews has been captured (process.run + override.run spans with annotations) to surface (a) where the agent's recommendations systematically diverge from reviewer overrides, (b) tribal knowledge embedded in reviewer reasoning that should become explicit rules, and (c) concrete code-change proposals against the procurement-agent codebase.

Context Graph Mining

This skill mines a procurement-agent Arize project for decision patterns, builds a context graph view of how the organization actually decides, and proposes concrete updates to the agent's rules and prompts.

When to use

• After a batch of human overrides has been applied (e.g. Vera Fye reviewing 100+ requests via procurement-reviewer/).
• When you want to surface rule-reality gaps — where the agent's encoded rules disagree with how the org actually decides.
• When you want concrete diff proposals against data/policies.json, data/vendors.json, or the evaluator system prompt.

What you have access to

• The procurement-agent Arize project — export traces, spans, and annotations via the arize-trace skill or the arize.exporter.ArizeExportClient Python SDK.
• The procurement-agent/ codebase — src/agent/evaluator.py, src/agent/tools.py, data/policies.json, data/vendors.json.
• The procurement-reviewer/src/reviewer.py system prompt — Vera Fye's institutional knowledge as currently encoded.

Workflow

Phase 1 — Discovery