Expert Tools
A nowcasting RL task built in verifiers 0.4.0 has to close four routes to its own answer
Written against Prime Intellect's TaskData, Task, Taskset and Toolset classes and built on real ALFRED vintages, a payrolls taskset let a command in the agent's box on the subprocess runtime read January 2024's hidden answer of 353,000 jobs.
September 2026shadcn/ui gives custom UI a source code boundary
The strongest starting point is not the library with the prettiest demo; it is the layer your team can change without fighting its API.
Base UI keeps custom React components changeable
Use Base UI for interaction, own the visual source, and test the assembled component where automation stops.
August 2026Moving a date from Spark to Snowflake takes two casts, and SQLGlot writes them
A no-dependency library for more than 30 SQL dialects, honest that its lenient parser will accept queries a real engine would reject.
Grapple.nvim tags a file by name, and the name points to a different file on every branch
The git_branch scope gives each branch its own tag file on disk, so checking out another branch loads a different set of names under keymaps that never change.
beartype type-checks a whole list by reading one random item of it
A list of bytes where the annotation promised strings slips past mypy at author time and trips beartype's decorator at the call, which names the exact item that broke the hint.
Jujutsu undoes a botched rebase back to the exact original commit IDs
Every jj command commits the working copy and writes itself into one continuous operation log, and jj undo works by stepping back through that log to the exact prior state.
Outlines masks every invalid token as a model decodes, so the output fits your schema by construction
The guarantee holds on a local model, where the library controls the logits, and thins to the provider's own feature the moment you point it at a hosted API.
Atuin makes your shell history answer questions bash can't
Every command lands in a local SQLite file with its directory, exit code, and duration, and the encryption that syncs it covers less than the word suggests.
grug-far opens a project-wide search as an editable buffer and writes your edits back across files
Delete a match from the results buffer and it drops out of the write; edit a match and Sync rewrites that line in the file it came from.
VisiData turns each new question about a table into a keystroke on the current column
The sheet is already loaded and every transform lands on the column under the cursor, so asking a new question costs a keystroke instead of a rewrite.
Serena gives a coding agent a language server, so it reads and edits code by symbol instead of grepping for it
The agent addresses code by symbol name instead of grepping whole files into context, though the free backend that does the work is narrower than the 40-language banner suggests.
July 2026files-to-prompt folds a slice of a repo into one prompt, file by file
Behind the single command sits a 334-line script that has not been touched in seventeen months, still doing exactly the job it shipped to do.
oil.nvim edits a remote server's files the same way it edits a local directory: as text
Renaming, moving, and deleting a file are edits to a line, and the plugin deletes for good unless trash is switched on first.
Pydantic Monty pauses a running snippet mid-call and resumes it in a different process
Version 0.0.19 starts in microseconds and blocks the filesystem and network by default, but it still can't run a match statement or import itertools.
ast-grep matched and rewrote 9 nullable-call sites across 5 files in the TypeScript compiler's source
The pattern binds to the parse tree, not the text, so it skips every look-alike sitting inside a comment, a string, or a call with the wrong number of arguments.
py-spy profiles a running Python process without changing a line of it
It reads a running interpreter's memory from outside the process, which is why attaching needs no code change and why Linux makes you clear ptrace first.