Charonyuu/whatsmyagentdoing

The problem

Once you run several coding agents at once — a Claude Code tab here, a Codex run there, a subagent spawned mid-task — you lose track of which one is still thinking, which is waiting on an approval, and which finished ten minutes ago. The only feedback is whichever terminal window is focused. whatsmyagentdoing parses the session logs the agents already write to disk and turns them into a single live view of every session on the machine.

What's different

The usual fix is to wire up hooks or a statusline script, or to leave each terminal to speak for itself. This project skips configuration entirely: it reads the JSONL transcripts under ~/.claude/projects/, ~/.codex/sessions/ and the Gemini CLI equivalent, tails them read-only, and pushes status over Server-Sent Events. Two things stand out. First, the presentation: an isometric office where working agents sit at desks and finished ones wander to the pantry or the ping-pong table, plus a plain list view when you actually want to scan states. Second, it is packaged as an agent skill (/myoffice), so installation can be delegated to the agent itself instead of following a setup doc.

Who it's for

Developers who routinely run three or more agent sessions in parallel and want ambient awareness without alt-tabbing. The office view and the --private flag make it usable on a screen share. It is less useful if you run one agent at a time — the terminal already tells you everything — or if you use an agent whose transcripts live in an undocumented binary store; the README explicitly says Google Antigravity is not supported.

Cost of adoption

Node.js 18+ is the only requirement; there is no npm install and no runtime dependencies. The clone lands in ~/.claude/skills/myoffice or ~/.codex/skills/myoffice, which is a per-agent-skill convention rather than a global binary. The risk is squarely on log-format stability: the README admits the formats are not public APIs, so a vendor update can silently degrade statuses. The 3D scene pulls ~180 KB of bundled three.js and ~1.3 MB of models, but only when you select High. MIT licensed; the 3D assets are Kenney CC0.

Verdict

This is a one-day-old project with roughly 180 stars and a new-board appearance at +118 stars a day — momentum is real but entirely unproven, and there is no trend history to place it against. The idea is genuinely useful and the implementation is honest about its limits, which is more than most agent dashboards manage. But it is built directly on undocumented transcript formats from three products moving fast, so expect breakage. Worth trying if you already run a fleet of agents; not worth depending on until it has survived a few upstream releases.