Overlook: AI said done. See what actually changed.

[PROBLEM] AI coding agents now take a ticket and come back with a finished pull request and a confident report: "Done. I only changed the article views. No API changes. All tests pass." Is it true? Today you find out by reading the whole diff, or you don't find out at all. Reviewers and tech leads approve more agent PRs than they can read line by line, and the product owner who wrote the request usually can't read a diff. The risk hides in side edits: a shared date helper "cleaned up", an API serializer tweaked on the way, a failing assertion rewritten until it passes. Each looks harmless in a diff; together they reach screens and API contracts nobody asked to touch. [SOLUTION] Overlook reviews the agent's work for you. Paste a GitHub pull request, compare, commit or repository link. Overlook clones the repo, reads the request from the issue the PR closes and the agent's report from the PR description, and has a map up in seconds. It answers four questions before merge: Did the agent stay in scope? Is its report true? Do the original tests still pass? What needs a human decision? [HOW USERS INTERACT] The codebase becomes a 3D isometric city. Folders are blocks, files are buildings; height is lines of code and a bright top band shows how much of the file this task changed. Blue changed inside the requested area, red changed outside it, amber wasn't touched but imports a changed file. The requested area is fenced in blue, proposed by Bob (or guessed from the request when Bob hasn't run) and confirmed or edited by a person on the map, with every verdict recomputed live. Click a building for why it was flagged, what depends on it, the diff, and Approve or Revert (keys A, R, J). Replay the task commit by commit to see where it drifted, click any commit in the history bar, or compare any branch to spot files both sides touched. Decisions become one git revert commit per file, and a receipt with every verdict can be posted on the PR. [TARGET USERS] Reviewers and tech leads drowning in agent PRs, who want the risky changes first instead of the whole diff; product owners who need to know whether what shipped matches what they asked for; and teams running agents in parallel whose branches overlap in ways nobody planned. [WHAT MAKES IT UNIQUE] One rule: evidence decides, Bob explains. Everything git can prove (changed files, API file touches, rewritten or weakened tests, the import graph for JS/TS, Go and Python) is computed by deterministic code; no model decides a fact. IBM Bob does the judgment: drawing the requested area from the brief, splitting the report into typed claims, writing an executable check for each feature claim, and explaining each change in plain language. Tests are run, not trusted: Overlook runs the original pre-agent tests against the agent's code in a throwaway git worktree, so rewriting a test cannot fake a pass. Until they run, "all tests pass" stays unverified. [REAL RESULTS] On the GT-142 demo, the report said "only the article views, no API changes, all tests pass". Overlook found 4 of 6 changed files outside the request, 3 more affected through a shared helper, an API serializer change and a rewritten test: 1 of 4 claims held. Bob audited three real agent tasks in public repos and caught a false claim in both Copilot PRs (github-mcp-server #1645 and playwright-mcp #725) that our own hand-written audits missed. On a real repo, Overlook ran the original 78 tests and all passed. Across 22 merged Copilot, Codex and Devin PRs measured via the GitHub API, 15 of 20 changed a test file and 6 rewrote assertions. [EFFICIENT BY DESIGN] Plain Node.js with zero npm dependencies, a static UI with no build step, and the same verdict code in the CLI, local server, MCP server and browser. 48 tests. Without Bob it still runs as a clearly labelled Draft audit. Try it at overlook-olive.vercel.app, no sign-in or keys, or clone the repo and run npm run site to audit any GitHub link locally, with Bob, saved audits and test runs.