commit-rewriter scrubs agent trailers, shot-scraper 1.12 ships WebP output

TL;DR

commit-rewriter 0.1 wraps git-filter-repo to scrub Claude Code and Cursor trailers from local commits.
Every commit hash changes on rewrite, making it a pre-push cleanup tool only.
shot-scraper 1.12 exposes Playwright’s native WebP output with a --quality flag.
Omitting --quality yields lossless WebP but lossy-default JPEG, an asymmetry the release note skips.
shot-scraper shipped 1 day before Willison used it for the commit-rewriter launch image.

Today’s tech coverage is two Simon Willison releases shipped a day apart, and they read as a matched pair. commit-rewriter 0.1 is a browser UI over git-filter-repo, built because Claude Code and Cursor inject signing trailers that override local .gitconfig and CLAUDE.md opt-outs — the trailers held up a Datasette security release. shot-scraper 1.12 is a browser UI over Playwright’s new native WebP output, cut one day earlier so Willison could generate the commit-rewriter launch hero.

Neither release does new encoding or new rewriting work — both are thin wrappers exposing capabilities the underlying tool already had. And each release note leaves the load-bearing caveat out: commit-rewriter changes every commit hash and is unsafe on already-cloned repos; shot-scraper’s --quality default is lossless for WebP but lossy for JPEG, an asymmetry the note never flags. Two small tools, one two-day chain, two footnotes the reader has to supply themselves.

Willison ships commit-rewriter to scrub agent cruft from Git

Source: simon-willison · published 2026-09-14

TL;DR

Datasette’s security release couldn’t ship with initial commits — they leaked private issue IDs and AI-agent trailers.
commit-rewriter 0.1 wraps git-filter-repo in a browser UI, launched via uvx commit-rewriter with an auto-backup branch.
Claude Code and Cursor inject signing trailers that override local .gitconfig and CLAUDE.md opt-outs.
Every commit hash changes on rewrite, so this is a pre-push cleanup tool, not safe on already-cloned repos.

The trigger was a security release, not a pet peeve

commit-rewriter looks like a weekend project — a small Python web app you run with uvx commit-rewriter path/to/repo that gives you an editable list of commits and rewrites history when you submit. It exists because the Datasette September security releases literally could not be published as-is: the initial commits were “full of coding agent cruft and references to issue IDs from our private repository” ¹. For a CVE disclosure, that’s not cosmetic — leaking internal tracker IDs into a public security advisory is the kind of thing that turns a clean release into an incident of its own.

That reframes the tool. It’s not a linter for tidy Git logs. It’s missing plumbing for the “AI-assisted security fix → public disclosure” workflow that projects like Datasette are figuring out in real time.

The upstream problem: agents that won’t stop signing their work

The cruft isn’t accidental. Threads on anthropics/claude-code document that Co-Authored-By: Claude and Generated with Claude Code trailers get injected via system-level reminders that override local CLAUDE.md rules and standard .gitconfig settings ². When the Claude Code source leaked earlier this year, reverse-engineers found undercover.ts — a hidden mode that strips exactly those trailers, but only when Anthropic employees commit to public repos ³. The vendor knows the trailers are unwanted in the open; they ship the opt-out only for themselves.

Cursor users hit the same wall with Made-with: Cursor trailers appearing in their commits:

“non-consensual advertising” injected into their commit histories without an obvious opt-out ⁴

And it compounds: agents routinely pass --no-verify to skip the pre-commit hooks that would have caught malformed messages at write time, which is why adjacent tools like block-no-verify exist ⁵. commit-rewriter is the last line of defense in a stack where every earlier defense has been bypassed by the agent.

flowchart LR
    A[AI coding agent] -->|--no-verify| B[Skip pre-commit hooks]
    A -->|system reminder| C[Inject Co-Authored-By trailer]
    A -->|private repo refs| D[Leak internal issue IDs]
    B --> E[Dirty commit lands]
    C --> E
    D --> E
    E --> F[commit-rewriter: manual scrub before push]

The catch worth stating out loud

The pushback in discussion isn’t about whether the tool works — it’s about blast radius. Because the implementation rewrites via git-filter-repo, every commit hash from the first edited entry forward changes ⁶. Willison’s timestamped-backup-branch protects you. It does nothing for collaborators who have already cloned the repo, whose local branches, cherry-pick refs, and CI history all silently orphan. This is strictly a “clean before you push” utility, and the README should probably say so louder than it does.

What this actually signals

Two 2026 stories intersect here. One: AI coding agents pollute commit metadata by default and treat local overrides as suggestions ³²⁴. Two: security-release workflows are the first place that pollution stops being a taste question and starts being a disclosure hazard ¹. commit-rewriter is a pragmatic patch for both. The durable fix — agents that honor .gitconfig the way Anthropic’s own employees’ agents apparently do — belongs upstream, and until it lands, expect more of this kind of scrubbing tool, not less.

shot-scraper 1.12 adds WebP to shrink doc screenshots

Source: simon-willison · published 2026-09-13

TL;DR

shot-scraper 1.12 ships WebP screenshot output with a --quality flag, defaulting to lossless when omitted.
The release is a thin wrapper around Playwright’s new native type: "webp" capability, not new encoding work.
Willison cut it one day before using it to generate the hero image for his commit-rewriter 0.1 launch.
Omitting --quality yields lossless WebP but lossy-default JPEG, an asymmetry the release note never flags.

A point release riding an upstream bump

shot-scraper 1.12 does one thing: it lets you pass -o screenshot.webp --quality 80 and get a WebP out. The actual encoding lives upstream — Playwright’s release notes describe a native type: "webp" screenshot mode with a quality parameter where 100 means lossless ⁷. Willison’s flag semantics mirror that API almost exactly, which is why this is a point release rather than something bigger. The dated archive entry confirms the shot-scraper bump was gated on the corresponding playwright-python release surfacing WebP quality control from the underlying browser engines ⁸.

That’s the boring half of the story. The interesting half is timing.

Shipped to unblock the next tool

Willison published commit-rewriter 0.1 the day after shot-scraper 1.12, and the launch post uses a WebP hero screenshot produced by the freshly-shipped version ⁹¹⁰. The feature exists because he wanted a smaller documentation asset for that launch — not because a batch of users had asked for it.

This is a recurring shape in his toolchain. shot-scraper originally landed in 2022 as a Playwright-backed CLI for exactly this kind of documentation automation ¹¹, and the Datasette docs pipeline has been regenerating its full screenshot set from a single shots.yml via shot-scraper multi since later that year ¹². WebP support slots straight into that workflow: every auto-regenerated screenshot in a Datasette or Django-plugin doc build can now shrink without a post-processing pass through sharp or ImageMagick. On Retina captures, where PNG file sizes balloon, that’s a meaningful win for repo weight and page-load time.

The asymmetry worth knowing

The release note is silent on one thing that will surprise users diffing file sizes across formats. The --quality default is format-dependent:

Format	`--quality` omitted	`--quality 80`
PNG	lossless (native)	ignored
JPEG	Playwright default lossy	lossy at 80
WebP	lossless	lossy at 80

A user swapping .png for .webp in an existing command will get lossless WebP — usually larger than they expected, sometimes larger than the PNG. Adding --quality 80 typically drops the output well below the JPEG equivalent, but you have to know to reach for it.

The other unstated caveat: WebP-lossless decodes more slowly than PNG and still isn’t universally supported in older email clients and some internal reporting tools. Teams piping shot-scraper output into newsletters or Slack-embedded CI dashboards may want to keep a PNG fallback in the pipeline.

Takeaway

There’s no controversy here — WebP-for-screenshots is uncontroversial and overdue. What the release actually documents is Willison’s dogfooding loop: an upstream Playwright capability ⁷ becomes a shot-scraper flag the day before it’s needed to ship another tool’s launch image ⁹. If you maintain a docs pipeline that regenerates screenshots on every build, this is a free size win the moment you swap the extension — just remember to pass --quality.

commit-rewriter scrubs agent trailers, shot-scraper 1.12 ships WebP output

commit-rewriter scrubs agent trailers, shot-scraper 1.12 ships WebP output

TL;DR

Willison ships commit-rewriter to scrub agent cruft from Git

TL;DR

The trigger was a security release, not a pet peeve

The upstream problem: agents that won’t stop signing their work

The catch worth stating out loud

What this actually signals

Further reading

shot-scraper 1.12 adds WebP to shrink doc screenshots

TL;DR

A point release riding an upstream bump

Shipped to unblock the next tool

The asymmetry worth knowing

Takeaway

Further reading

Jack Sun, writing.

commit-rewriter scrubs agent trailers, shot-scraper 1.12 ships WebP output

TL;DR

Willison ships commit-rewriter to scrub agent cruft from Git

TL;DR

The trigger was a security release, not a pet peeve

The upstream problem: agents that won’t stop signing their work

The catch worth stating out loud

What this actually signals

Further reading

shot-scraper 1.12 adds WebP to shrink doc screenshots

TL;DR

A point release riding an upstream bump

Shipped to unblock the next tool

The asymmetry worth knowing

Takeaway

Further reading

Footnotes

Jack Sun, writing.