A coding-agent session rendered as a film instead of a log. Every read, edit and test run becomes a visible beat, and a claim of success with no passing test behind it is flagged on screen.
A seasonal stage: purple houses with lit windows, jack-o'-lanterns, bats overhead and a ghost-white gremlin when a test fails. About once a minute something leaps at the screen, and one button turns the scares off.
While the agent thinks, the lead now strolls round its building and looks it over. The set picker scrolls properly inside VS Code, and the extension is on Open VSX for Cursor, Windsurf and VSCodium.
A map of every file the agent touched in a session: grouped by folder, sized by lines changed, coloured new, edited, deleted or read, and checked against git so shell edits count too.
It shows up in four places: a one-line summary under each Claude Code reply, the Changes view on the stage, an Agent changes list in VS Code, and npx agent-theatre changes --pr, which posts the map to the pull request after asking.
One night, seven channels. npx agent-theatre, the VS Code extension, a Claude Code plugin with a /theatre command, a Homebrew tap, the MCP Registry, and agent-theatre init for Cursor, Windsurf, Codex and Gemini CLI.
Also tonight: main is protected, CI runs on three operating systems, secret scanning and push protection are on, and the repo history was checked for tokens before any of it went public.
Open VSX is uploaded but not yet active, so the site points Cursor users at the .vsix until that namespace is claimed.
The stage picks a world for your repo: an island, an orbital station, a night city, a monastery in the mist. Each one has its own layout, light, silhouette, instrument and ambience, and every sound is synthesised at runtime. There are no audio or image files in the repository.
Every kind of work now has a tool. The character builds with a hammer, fixes with a wrench, lays bricks with a trowel, runs the tests at the beacons and chases gremlins with a net. Subagents arrive at the dock.
When the agent goes quiet, the house lights dim and the character sits down. When it needs you, the stage says so and points you back to the terminal. Export any run as an mp4 and post it.
The third product starts the way the other two did: with the honest part first. A generic agent-event schema that carries paths and counts but never file contents or prompts, with a test that plants secrets in every field and asserts zero leaks.
Then the claim check. If the agent says the job is done and no test has passed since its last edit, a gremlin hides behind it. A test run that exits zero with no counts is not evidence either.
npx debuganything why <pid> against any process, container or pod. It attaches, picks the right tools for the symptom, and returns ranked findings where every finding cites an evidence id. Then it detaches.
The Python package ships the same contract for LangChain, LlamaIndex and CrewAI agents. No breakpoints, no restarts, no instrumentation. It is the layer before the step debugger, for when you do not yet know where to put the breakpoint.
Verification quality release and verify-anything · LastSearch
Debug Anything, first ship · Debug Anything v0.2.0
LastSearch goes live · LastSearch
The first night · LastSearch
You pick what’s next.
Suggest one on GitHub or in Discord. A vote is a thumbs-up on the idea’s GitHub issue, so every count is public; new votes show here within a minute or so. The top idea on Monday morning is the one that gets built.
