Ghostwriter

Every reviewer knows the feeling: a PR description says "minor cleanup," and the diff says otherwise. Ghostwriter closes that gap. It runs three isolated Bob 2.0 subagents — Technical Truth-Teller, Scope Auditor, and Reviewer's Ghost — in parallel via Agent mode, each with no visibility into the others' output. Technical Truth-Teller independently describes what the diff actually does and flags anything undisclosed. Scope Auditor reads the linked ticket (document understanding) and checks whether the PR stayed in scope. Reviewer's Ghost predicts the specific, risk-focused questions a senior reviewer would ask. A synthesis step is the only place these three outputs meet. It produces a Communication Debt Score (0–100) and, more novel, Agent Disagreement Detection: it compares each subagent's independently assigned risk level, and when they genuinely conflict, it flags the PR for human review with a note explaining exactly what the disagreement is about. A fourth mode, Review Composer, turns a completed audit into a GitHub-ready review comment — including real one-click suggestion blocks, but strictly limited to small, mechanical fixes, never architectural judgment calls. We validated the full pipeline against 11 real, curated merged PRs from freeCodeCamp/freeCodeCamp (BSD-3-Clause). It caught a real render-scope behavioral bug hidden behind an "update imports" title, a carried-forward TODO bug in a PayPal capture flow, and a genuine risk-assessment disagreement between subagents on that same PayPal PR — proof that independent parallel analysis surfaces things a single pass misses. We also built a live "paste any PR" demo service, which replays these exact Bob-validated prompts against an external model API (for quota reasons, not Bob itself at runtime) so anyone can try it on a real PR during the demo.