I ran an autoresearch loop to optimize the executable size of zoxide. ~ 40.2% reduction

[haydonryan]

Thank you for an amazing tool - I use it every day on every computer and VM use.

I know AI is a polarizing topic for developers so I wanted to start a discussion before sending in PRs etc.

I'm playing around a lot with AI, and I decided to do an karparthy style auto-research experiment and decided to use one of my favorite rust apps as the test subject.

Initial prompt was done on chatgpt (instant) and I fed that into deepseek v4 flash (nvidia/DeepSeek-V4-Flash-NVFP4) running locally on my server.

A few things I noted in this experiment:

1. it tried to rip out colorization but that was because it didn't pick up that no changes in appearance was allowed. I undid those changes.
2. It uses build-std and the nightly toolchain (since build-std isn't in stable)
3. It did use the pattern of "implement it myself vs pull in crates.

Anyways let me know what you think? Want a PR? Want this to just go away? :)

Everything below here was written by AI.

Binary size optimization results for zoxide

#	Optimization	Before	After	Delta	Cumulative
0	Baseline (original Cargo.toml, stable toolchain)	1,100K	—	—	1,100K
1	panic = "abort" in release profile	1,100K	1,008K	-92K (8.4%)	1,008K
2	Replace ouroboros with plain struct; make Dir owned	1,008K	1,007K	-1K (0.1%)	1,007K
3	Disable clap default features (keep color, help, usage, error-context)	1,007K	997K	-10K (1.0%)	997K
4	anyhow with default-features = false	997K	983K	-14K (1.4%)	983K
5	Replace time crate with manual YYYY-MM-DD HH:MM:SS parsing	983K	949K	-34K (3.5%)	949K
Stable total		1,100K	949K	-151K (13.7%)
6	-Z build-std=std,panic_abort + panic = "immediate-abort" (nightly)	949K	658K	-291K (30.6%)	658K
Nightly total		1,100K	658K	-442K (40.2%)

What each change does

1. panic = "abort" — Eliminates landing pads and unwind tables for every function, since this binary never unwinds. Saves ~92K.

2. Remove ouroboros — The Database struct used ouroboros for a self-referential pattern (Vec<u8> with borrowed Vec<Dir>). Replaced with a plain struct using owned String paths. Removes the proc-macro crate entirely.

3. Prune clap features — Disabled suggestions (did-you-mean), env (env-var parsing), debug, unicode. Kept color, help, usage, error-context so appearance is identical.

4. anyoh default-features=false — Disables backtrace capture in error chain objects. No functional change to error messages.

5. Replace time crate — Only consumer was atuin.rs, parsing one timestamp format (YYYY-MM-DD HH:MM:SS). Replaced with ~30 lines of manual parsing. Removes time, time-core, time-macros, deranged, num-conv, powerfmt, nom, minimal-lexical from the dependency tree.

6. build-std + immediate-abort — Rebuilds the standard library with all panic formatting, backtrace capture, DWARF debug-info readers (gimli, addr2line, object), and unwind tables stripped. Requires nightly + -Z build-std.

Why the time crate removal saved 233K with build-std vs only 34K on stable

Without build-std, the standard library is a precompiled blob — removing time only drops its own code (~34K). The standard library's date/time formatting and parsing code paths remain linked in regardless, because libstd ships as a single monolithic archive.

With build-std, libstd is compiled from source with full LTO. Removing time means the compiler can now prove that std's own date/time formatting/parsing machinery is unreachable and eliminate all of it — not just the time crate itself, but every libstd function that was only reachable through time. The multiplier comes from LTO being able to see through the entire call graph, including the standard library.

Appearance/behavior preserved

• ANSI styled help output (bold/underline) — unchanged
• All 16 unit tests pass on stable
• All subcommands, error messages, database format unchanged
• Performance benchmarks show no regression (all commands complete in <5ms)
  96 tests pass on stable (all 16 unit tests, plus the nix-dev-gated shell compatibility tests that require external shell binaries).

[ajeetdsouza]

Hey, thanks. A lot of these changes trade stability for a very small binary size reduction:

1. panic = "abort" loses debuggable panic messages. While we may not have any panics in zoxide itself, we might have them in our dependencies.
2. Not much of a reduction.
3. The clap features are useful and don't seem to cost much. debug and env are disabled by default. Removing unicode would break zoxide's ability to add paths containing UTF-8 characters.
4. I don't think backtraces are enabled by default on anyhow.
5. This change would only be significant if we also do 6.
6. Same problem with panic=abort. Also, nightly is a problem for stability - I would accept a PR to enable build-std once it's available on stable.

[haydonryan]

Thanks for the feedback :) Appreciate it.

I had a feeling that might have been the case with the nightly toolchain.
5 just replaces the time crate with a parsing function (in that sense it is extra code to mange) but it doesn't need 6.

Have a great one.