Context Gap Analyzer

Code is necessary but insufficient context. The delta between "what the code says" and "what a competent team member knows" is where agents fail most expensively. Identify that delta, prioritize by agent impact, surface focused questions a human can answer.

What agents cannot derive from code: Why (rationale), What conventions aren't enforced by tooling, How things interact beyond repo boundaries, When to break vs. follow a pattern, Who owns what.

PHASE 1 — Context Harness Discovery

Discover all context infrastructure already in place. Actively explore — don't rely on a checklist.

Layer 1: File-level — Scan repo root + 2-3 levels deep for "instructions for someone working here." For each: read it, inventory topics, note scope (repo/dir/module), format, audience.

Layer 2: Toolchain — MCP servers, skills/plugins, hooks/middleware (git hooks, CI, pre-commit, linting), IDE/editor configs.

Layer 3: Runtime — Connected services (Drive, Notion, Confluence, Slack), env/secrets, package metadata.

Layer 4: Delivery classification