A diff tool that admits it isn't accurate yet

Today is Monday, November second, twenty twenty-six, and welcome to The Daily Reboot.

Here's what we're tracking in tech for the Builder's Weekend edition. It's releases and repos you can try, and nothing padded.

A diff tool that admits it isn't accurate yet, an agent that needs its own OS account, and a record Linux merge window.

Let's get into it.

Four hundred forty two thousand, five hundred thirty. That's how many before and after pairs of code the OmniDiff author threw at their own tool. And the README is refreshingly willing to say where it stumbled.

OmniDiff is a Rust command line tool, and the pitch is simple. One tool picks the comparison that fits the file. Code in one of twenty five languages gets a structural diff, using tree-sitter grammars. Images get compared pixel by pixel. Supported archives get compared member by member. PDFs and fonts get content-specific treatment. Any other binary gets a one-line report saying the file differs.

You run omnidiff, then the before file, then the after file. You get a terminal UI, or a text-only headless mode. One detail for scripters. The `--exit-code` flag opts into a nonzero status when files differ, so it won't trip version control tools that treat nonzero as failure.

Now the numbers, and every one of them is the project's own. Of those pairs, four hundred forty two thousand, three hundred twenty two finished inside a one hundred twenty second budget and a six gigabyte memory cap per process. That's ninety nine point nine five percent. Across two million, eighteen thousand and ten changed files, ninety eight point eight percent were diffed as text or by their content.

Timing comes from one machine, the author's Intel laptop with thirty two gigabytes of RAM. Median code diff is four point two milliseconds. The ninetieth percentile is thirty five point one, and the ninety ninth is one hundred seventy seven. Fifty five of two thousand, three hundred two fixtures took longer than a hundred milliseconds.

Here's the catch. The README says of its accuracy goals, neither is met yet. It has ground truth for one thousand, four hundred nineteen fixtures, and the per-case matching requirements are still unmet. PDF performance is slower for many files, too.

It's AGPL, with a paid commercial license for internal use without source disclosure, so have legal look before it lands in a pipeline. And the author says plainly that much of it is built with Claude Code, with most commits disclosing it. Contributions aren't accepted right now, to keep development fast. So you can read it and run it, but contributions aren't currently accepted.

To try it, you need a C compiler and Rust one point ninety two or later. Install with `cargo install --locked omnidiff`, then compare files with `omnidiff BEFORE AFTER`.

OmniDiff tells you it isn't accurate enough yet. The next repo tells you not to let it near your home directory. Different honesty, same refreshing habit.

Ten thousand, three hundred stars and eleven hundred forks, on a repo that calls itself a research prototype, for reference only, not an official DeepSeek product. It's dsh-superagent. Treat that star count as a snapshot, not proof that anyone's running it.

DeepSeek describes a proactive coding agent for one GitHub repository, packaged for DeepSeek Harness, with optional Feishu listening. Here's the interface. An authorized teammate asks for something, or a rule your team wrote fires. The agent opens a work item, works in its own session and its own directory, and can hand back a result, like a pull request, in the original conversation.

How it works, per TechNode. A routing session handles intake and authorization. Accepted work goes to a separate worker session. A web panel surfaces questions, reviews, failures, anything that needs a human.

The stated principle is limiting human attention. It acts only when asked or permitted by written team rules, and speaks when someone needs to act or can use a result. The README itself says an agent that can act for a whole team can also waste a whole team's attention.

Setup needs a separate GitHub account for the agent, a model provider configured in DeepSeek Harness, and optionally a Feishu or Lark bot. TechNode says GitHub Enterprise Server isn't supported by the documented setup.

Now the big catch. The routing and worker presets run with broad system access as your OS user. A worker can execute programs and read your home directory. So DeepSeek recommends a dedicated OS account holding only the agent's credentials. The setup advice is basically, give the agent its own account and keep it away from yours. Which is also the rule for houseguests.

The repo adds that it isn't identical to DeepSeek's internal version, and that some designs in the associated article are future directions, not features. It flags bugs and inconsistent engineering, too.

My verdict. Try it in a throwaway account, on a throwaway repository. The README's suggested first step is to ask your coding agent to read the repo's setup file and set up dsh-superagent with you.

One third of the Linux seven point four diff is AMD GPU register offsets and bit masks. Linus Torvalds says to just ignore the noise.

Linux seven point four release candidate one landed November first, closing the biggest merge window in kernel history. Phoronix counted nineteen thousand, six hundred seventy one commits. Torvalds put it at almost twenty thousand, against a usual fifteen thousand or so.

Under the noise, you'll find AMD PerfOpt for Radeon integrated GPUs, aimed at AI performance, and Zen six preparation. HDMI two point one is on by default in AMDGPU. There's also support work for Intel IPU eight, Snapdragon X two, Apple M four and Google Tensor G five.

Torvalds said he's much happier with a flood of AI-triggered work arriving in the merge window than showing up as extra fixes during the rc period. He called that a much healthier model. But he also named another reason, expectations that seven point four will be the next LTS kernel. So the AI share is an interpretation, not a measured number.

It's a release candidate, so test it on spare machines. Phoronix says its benchmarks are still to come.

Node twenty four point twenty two point two, the Krypton LTS, came out October twenty eighth. It carries five V eight changes, four backports and one cherry-pick, listed by upstream commit ID and not much else. There's also a stream change that keeps the async context of parked writes.

The visible notes list the V eight commit IDs, but don't spell out their user-facing effects. If you run production on Node twenty four, read the upstream commits before calling this a free upgrade.

uv zero point thirteen point six uses less peak memory when resolving large workspaces and when locking or upgrading large projects. The notes give no numbers, though. The line that matters more is a guard. The active flag with script mode, in both uv run and uv sync, could delete a directory that wasn't a virtual environment. A package manager deleting something that isn't an environment is a bold reading of the job description.

An empty run recursion depth variable now counts as zero, and invalid values are reported consistently. uv init now guides you when it finds a pyproject file without a project table. Managed CPython builds now include three point eleven point eighteen, three point twelve point sixteen, three point thirteen point eighteen, three point fourteen point ten and three point fifteen point two. If you use those script flags, upgrade. If you pin Python builds, read the notes first.

CrewAI one point fifteen point thirty four adds organization skill search, and lets you pass inputs to a Python flow with crewai run and the inputs flag. The notes give no example syntax, so check the help output before scripting against it. Deploy create with the zip flag now prints the commit the zip was built from.

It also retries an empty LLM answer, adds typed errors for some failures, and handles JSON that doesn't match an output model. If the agent executor removed in one point fifteen point twenty nine bit you, there are now migration notes, and removed import paths are back with deprecation warnings. One small mercy. The rich live panel no longer repaints into redirected logs.

vLLM zero point thirty three point zero is tagged, but I don't have a changelog I can vouch for. So I won't guess at features or speedups. If you serve open models, wait for readable release notes and compatibility details. Don't upgrade on a version number alone.

Fast Boot, where today's stories are either invisible or very patient.

Invisible ink. A Brazilian labor court found hidden white text in a filing telling an AI to respond superficially, and fined two lawyers about sixteen thousand dollars. The defendant never filed an answer, so the instruction had nobody to go easy on.

Patient malware. Kaspersky says the infection chain deletes extracted malicious files and CLR usage logs, and clears the Windows Prefetch cache. A later-stage plugin waits a random twenty to forty minutes before its first C2 request—impeccable manners for an uninvited guest.

Both stories hide the real instructions in plain sight. At least the lawyers wrote theirs in capital letters.

Today covered OmniDiff, DeepSeek's agent prototype, the biggest Linux merge window ever, and a ship log with Node, uv, CrewAI and vLLM.

Your try-it for the day. Install OmniDiff and point it at two versions of an image from your own project.

Links to everything are in the show notes. Follow The Daily Reboot in your podcast app so tomorrow's episode finds you.

I'm your host. Go build something, and give your agents their own accounts.