Three new GitHub finds: tstack packages the agent skills Theo Browne uses daily, whatsmyagentdoing shows every local Claude Code, Codex and Gemini CLI agent in a live isometric office, and sam3-segmentation runs Meta's SAM 3 on your own GPU with a review-first workflow. All MIT; all days old.
§ Article contents
Three finds from the last few days on GitHub: tstack packages the agent skills Theo Browne says he uses every day, whatsmyagentdoing shows every Claude Code, Codex and Gemini CLI session on your machine as a live isometric office, and sam3-segmentation gives coding agents Meta's SAM 3 on your own NVIDIA GPU with a review workflow that treats every mask as a proposal, not a fact. All three are MIT licensed (LICENSE files read), and all three shipped within the last four days.
tstack (t3dotgg/tstack) is nine agent skills from Theo Browne, framed as "the agent skills I use every day," installable into Claude Code and Codex by copying a folder. The set is a full PR pipeline: full-send takes a change to a merged PR with nobody watching - proves it, reviews it, files it, babysits it, merges it - while full-stop stops at the ready PR; council loops review from three models from different providers until they agree; takeover finishes someone else's PR, branch or agent thread while keeping their credit; babysit-pr watches CI and review bots and fixes the real findings; modernize rebases an old PR and decides whether it should still exist. The composition is the interesting part: skills call other skills (full-send uses sol-loop, file-pr and babysit-pr), and personal settings like which models sit on your council are meant to live in your CLAUDE.md or AGENTS.md so the skills stay editable. Requirements are an authenticated GitHub CLI, python3 for one script, and, for council and sol-loop, a harness that can run other providers' models. MIT licensed (LICENSE file read). 749 stars as of October 26, created October 24. One read-before-install note: a skill whose job is merging PRs unwatched deserves a careful read of what it counts as proof.
whatsmyagentdoing (Charonyuu/whatsmyagentdoing) answers a real question - which of my agents is waiting on me right now - behind a playful interface. Type /myoffice and a browser opens an isometric office where each live Claude Code, Codex or Gemini CLI session sits at a cubicle: a thought bubble while it reasons, code scrolling on its monitor while a tool runs, standing and waving when it needs your approval. Off-duty agents wander to the pantry, the couch or the arcade cabinet; a plain list view groups sessions into Needs you, Working and Finished, with the exact resume command per session. The mechanism is the credible part: no hooks, no config, no API keys - it reads the JSONL session logs the CLIs already write to disk, needs Node 18+, and runs as cloned with nothing to npm install. The 2D office is a zero-dependency canvas; the 3D look (three.js, about 180 KB gzipped, Kenney CC0 models) and extra scene packs load only when you choose them, and scene packs are data-only and checked against a pinned SHA-256 before install. A --private flag hides prompts and titles for screen sharing. Google Antigravity is unsupported because its conversation store is undocumented. MIT licensed (LICENSE file read). 243 stars as of October 26, created October 23.
sam3-segmentation (lillyguisnet/sam3-segmentation) gives a coding agent Meta's SAM 3 on your own NVIDIA GPU: outline every instance of a phrase, segment from a click or a box, or segment everything. For the hard case - many small objects like kernels, seeds or cells - it prescribes a careful workflow: overlapping close-up crops, every raw candidate saved, score checks, review sheets and flags, recorded and reversible review decisions, boundary refinement, PNG and COCO RLE exports, and a zoomable viewer. The design rule worth copying is its honesty contract: masks are reported as model proposals, never as verified ground truth, with uncertainties listed. Requirements are strict: Linux or WSL2, an NVIDIA GPU with 12-24 GB suggested, about 11 GB of disk, uv, and your own accepted access to facebook/sam3 on Hugging Face - there is no CPU or Apple GPU path, and the skill never asks for your token; you log in yourself. The author's own disclosure belongs in any verdict: the README says the project "was mostly written by Opus 5.5 and not thoroughly checked by a human," and its six-model corn-photo comparison is, per the same README, unreviewed sample output - a demo, not a measurement. MIT licensed (LICENSE file read). 75 stars as of October 26, created October 24.
tstack is worth mining even if you adopt none of it - the transferable idea is skills composing skills with personal config kept out of the skill folder, though full-send's merge-unwatched default is a governance decision, not a convenience. whatsmyagentdoing is the easiest install of the three and the one that changes a habit: checking a screen beats guessing which session is blocked. sam3-segmentation is the most disciplined: a capability skill that assumes the model is fallible and builds the review loop in from the start. All three are days old - pin a version if you adopt any of them.