neuron-explorer
Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
A modern dependency updater with strong defaults. He never sleeps. Neither do your CVEs.
Developer toolsRust · PklGitHub @sweepiesOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
Count Patchula vants to update your deps. A modern dependency updater with strong defaults.
Taken from the repository README
Loading the full README.
Read the full README ↗The Count reads your repositories' manifests, looks up newer versions, and opens one pull request per update, with release notes. One process handles the whole batch: repositories run concurrently, and every package is looked up once however many repositories use it.
mise use -g --minimum-release-age 0 packslip:sweepies/count-patchula
Other ways to install him are on the installation page.
Invite the Count into a repository with patchula.pkl:
amends "PatchulaConfig.pkl"
prs { dashboard = true }
Then, with a GitHub App installation token in PATCHULA_TOKEN:
patchula run --dry-run acme/my-app # look first
patchula run acme/my-app # then let him bite
Lockfile tools, and the versions of the tools in mise.toml, come from
mise, so it needs to be on the PATH. See getting started.
The CLI is the quickest way to try him, and the GitHub
Action runs him on a
schedule. The recommended way to run him is the GitHub App
server: one long-running
process keeps every cache warm, batches work across repositories, and acts on
PR checkboxes the moment you tick them. It idles cheaply between batches, so
usage-billed hosts such as Railway suit it well. Each release publishes it as
a container image,
ghcr.io/sweepies/count-patchula, for linux/amd64 and linux/arm64. Once it is up, patchula run --remote runs on it from your machine, signed in with GitHub, with no
token or toolchain of your own.
patchula tui) to browse updates, and in a checkout, pick
some and write them into your files. patchula apply writes them without
asking, for scripts and pre-commit hooks.Against Renovate 44.173.1 on six large public repositories (mastodon, sentry, grafana, monica, jellyfin, kafka), each tool alone on the same c7i.xlarge:
| Count Patchula | Renovate | ||
|---|---|---|---|
| All six, cold | 150.5 s | 920.6 s | 6.1× faster |
| All six, warm | 14.8 s | 274.3 s | 18.5× faster |
| Peak memory, grafana cold | 494 MB | 3.6 GB | 7.3× less |
The benchmarks page has the method and every run.
The compatibility matrix
lists every manager, datasource and feature, and migrating a
config maps
renovate.json options onto Pkl.
Install mise, then mise trust and mise install.
mise run build # build the workspace
mise run test # offline tests, including the real lockfile tool tests
mise run lint # every hk check, Clippy (warnings denied) included
mise run hooks:install # the pre-commit and commit-msg hooks
mise run ci # every check CI runs, in parallel
mise run docs:dev # the documentation site for this checkout
Builds go through mr boxington for a shared
Cargo cache. Tests never touch the network: they replay HTTP fixtures the Count
recorded himself. Tools only the tests need, like the lockfile engines, are
task-scoped, so a plain mise install does not fetch them.
Count Patchula is an independent implementation under the MIT licence. Renovate (AGPL-3.0) may be run as a black box to compare results, but its source code, preset data and test fixtures are never read, copied or adapted into this repository.
Count Patchula would not exist without @jdx. His work inspired it, and it is built on his tools at every step: mise installs every toolchain and lockfile engine, aube writes the npm and pnpm locks, mr boxington caches the builds, usage generates the CLI reference and completions, tak runs the benchmarks, communique writes the release notes, and packslip signs the releases. Thank you, Jeff.
MIT