In a Claude Code skill's frontmatter, the description field is what Claude reads to decide when to load the skill, and disable-model-invocation: true stops Claude from loading it on its own, so it only runs when you invoke it.
How we checked
Claude Code skills docs frontmatter table confirms both fields and their behavior
Why it matters
If you write descriptions as triggers and lock skills with side effects, agents stop firing deploy or merge workflows by surprise.
Try this
Add disable-model-invocation: true to any of your skills that merges, deploys or sends messages.
T3 Code nightly builds include a Working section that holds threads whose background work the agent is waiting on, and threads settle once their PR is merged.
How we checked · Why it matters
How we checked
Oct 25 nightly notes mention Working and thread settling after a merge; Theo's 'every 6 hours' cadence claim not confirmed
Why it matters
This keeps finished or waiting agent threads out of your active inbox when many agents run at once.
Theo's full-send skill has an agent take a change all the way to a merged PR: prove it works, review it, file the PR, watch CI and review bots, then merge. full-stop does the same but stops before merging.
How we checked · Why it matters · Try this
How we checked
tstack README describes full-send and full-stop as stated; skill not executed
Why it matters
It shows a concrete merge gate and proof step for autonomous agents. full-stop is the safer version if you want a human to merge.
Try this
Start with a full-stop style flow, where the agent prepares a PR and a person merges it, before allowing any agent to merge.
tstack includes cross-model review skills. sol-loop has a GPT Sol model review a change in a loop until both agents agree, and council gets reviews from three models from different providers until they agree.
How we checked · Why it matters · Try this
How we checked
README lists sol-loop and council as described; the 3 to 5 round budget is only in the video
Why it matters
Pairing an authoring model with a reviewer from another vendor is a cheap way to catch mistakes that one model misses.
Try this
On one non-critical PR, ask a model from a different vendor to review your main agent's diff and compare what each one finds.
Theo has published tstack, an MIT-licensed GitHub repo of his daily agent skills for Claude Code and Codex. It includes full-send, full-stop, sol-loop, council, takeover, babysit-pr, modernize, file-pr and leaving-pr-comments.
How we checked · Why it matters · Try this
How we checked
Repo fetched: exists, MIT license, README lists the named skills; no sandbox run
Why it matters
You can read how a heavy agent user structures reusable skills for PR work and borrow the parts that fit your own workflow.
Try this
Read the full-stop skill file in the tstack repo and compare its steps with your own PR checklist before adopting anything.
T3 Code is an open-source control surface for coding agents, with desktop, web and mobile apps. It drives Claude Code, Codex, Cursor and other harnesses on your own machine and supports remote access.
How we checked · Why it matters · Try this
How we checked
T3 Code GitHub README confirms the harness list and the mobile/remote apps; not run
Why it matters
One interface for several agent harnesses across machines can make it easier to manage parallel agent threads.
Try this
Read the T3 Code README and its remote-access doc to see whether it fits your setup before installing anything.
Simon Willison released Seam Cutter, a browser tool that cuts an image horizontally or vertically and slides the two halves together, removing a strip from the middle, for example to redact a prompt from a stitched transcript screenshot.
How we checked · Why it matters · Try this
How we checked
Tool page fetched: exists and accepts an image for a seam cut; behavior taken from the blog post, not tested
Why it matters
It is a quick way to remove irrelevant or sensitive prompt text from long agent-transcript screenshots before you share them.
Try this
Open the Seam Cutter page in your browser and remove one section from a long screenshot of an agent transcript.
Theo says T3 Code Nightly added Orchestrator v2, where a main agent thread hands work to subagents running in other harnesses and models, each in its own worktree, which report back.
How we checked · Why it matters
How we checked
Release notes and README checked; no mention of Orchestrator v2 or cross-harness subagents
Why it matters
If shipped, it would allow mixed-vendor subagent review inside one tool, but this is not confirmed yet.