v0.4.2 structural edits, verified by default

Ground-truth code edits for AI agents.

Walrus gives agents structural visibility into large codebases and lets them patch exact expressions instead of rewriting whole files.

Install Walrus Read docs
$ curl -fsSL walrus.dev/install | sh
Terminal
$ wally -find FFieldExpander_Default
✓ class FFieldExpander_Default Field.h:128
$ wally -flow Counter::incr --body
return value + step;
$ wally -edit Counter::incr --return "value + step"
✓ verified · 31 tokens · 1 expression
200k files

Indexed in 58s on an 8-core machine.

50–100×

Token reduction on edit-heavy workflows.

What Walrus does

Four primitives. One binary.

-find

Resolve symbols to compact, structured summaries with precise file:line anchors.

-has

Truthy relation checks with exact source locations, not regex guesses.

-flow

Read only the function body you need; skip headers and unrelated boilerplate.

-edit

Surgical expression edits with parse/verify before source is written.

Index

Memory-mapped structural index.

Walrus parses source once, resolves symbols cross-file, and writes a typed index on disk. Every agent process maps the same pages, so parallel agents are cheap.

  • Size: 4.2GB index
  • RAM: ~5GB shared pages
  • Cold build: 58s
  • Incremental: 3ms updates

Surgical edits

31 tokens instead of 20k.

Naive edit

Read full file → rewrite full file → no structural verification.

~20,400 tokens

Walrus edit

Resolve handle → patch expression → verify parse/type context.

31 tokens

CLI

Subprocess-friendly by design.

$ wally -find Counter::incr
class Counter::incr @ Counter.cpp:52

$ wally -flow Counter::incr --body
return value + step;

$ wally -edit Counter::incr --return "value + step"
✓ verified · round-trip ok · audit written

Give your agent structural ground truth.

One install. No daemon. No sidecar service.