GitHub - press-pass/gifhub: Autonomous QA bug-hunt harness: an agent walks the full app flow, finds bugs, fixes them, and ships each as a PR with inline before/after GIFs. Target app included as a submodule.

An autonomous QA bug-hunt harness for any web-app repo.

Point a coding agent at your app; it walks the entire user flow like a real person, hunts for bugs, and for every bug it finds it:

1. records a GIF of the bug happening,
2. fixes it,
3. proves the fix with a UI end-to-end test,
4. records a GIF of that e2e test passing, and
5. opens a PR for the bug with the before + after GIFs embedded inline, so a non-technical reviewer can watch the bug, then watch it disappear — without reading a line of code.

Demo

A bug-hunt PR a reviewer can read top to bottom — the bug it found, the before/after GIFs playing inline, and the root cause — without opening the diff:

Install as a Claude Code plugin

gifhub ships as a Claude Code plugin, so you can drop these skills into any repo:

/plugin marketplace add press-pass/gifhub
/plugin install gifhub

That installs three skills — bug-hunt (find + fix + GIF PRs), prod-bug-report (report-only), and get-or-create-per-worktree-dev-env (the isolation helper bug-hunt leans on) — plus the /bug-hunt slash command. The skills auto-trigger when you ask an agent to hunt for bugs. Run /plugin update gifhub to pull changes. To try it without installing, from a clone:

claude --plugin-dir /path/to/gifhub

Requirement: per-worktree environments

gifhub parallelizes by running subagents in separate git worktrees, each with its own isolated, running copy of the app. So the one thing your repo needs is a way to stand up an isolated environment per worktree — a command that boots the app on a unique port with isolated data / state / auth, so parallel agents never collide.

You don't have to build this by hand. The get-or-create-per-worktree-dev-env skill gets it if your repo already has it, or creates it if it doesn't — checking for a Conductor setup first (per-workspace provisioning out of the box), and otherwise writing a gifhub.toml in Conductor's schema plus a scripts/worktree-env.sh runner. It then proves the isolation (two worktrees side by side: distinct ports, isolated data, isolated auth) before declaring success. bug-hunt invokes it automatically before fanning out, and a hook blocks the fan-out until isolation is verified. If a repo genuinely can't isolate per worktree, gifhub runs single-threaded.

Point it at your repo

Tell the agent (or bake into the prompt):

1. Env setup + App URL — usually nothing to do: get-or-create-per-worktree-dev-env gets-or-creates the isolated per-worktree env and returns the boot command + URL. Point it at an existing setup command if you already have one.
2. E2E — the command that runs the UI e2e suite (Playwright, Cypress, …), and the fact that it can record video (needed to produce the GIFs).

The debug step

The bug-hunt skill is the full brief. With the plugin installed it runs as the /bug-hunt slash command or auto-triggers when you ask an agent to hunt for bugs. PROMPT.md carries the same brief as a copy-paste prompt for agents without the plugin.

Report-only mode

Want bug reports without fixes — e.g. run against production — instead of fix-and-PR? The prod-bug-report skill walks the live app's flows like a user and writes each bug up with reproduction steps (steps: / expected behavior: / actual behavior:). No code changes; it just produces reports.

How it works

The mechanics that make the GIFs reproducible and the e2e reliable are in METHODOLOGY.md, bundled with the bug-hunt skill:

• one e2e spec produces both GIFs — before (run against the buggy build, fails) and after (run against the fix, passes),
• GIFs embed inline in the PR body even for private repos, via committed ?raw=true image URLs,
• favor deterministic UI surfaces for stable e2e (forms, inputs, toggles) over async/AI-driven UI,
• one bug per worktree → one PR.