[FEATURE] Feature-gate heavy dependencies to reduce 80MB binary size

[paultendo]

What variant of Codex are you using?

CLI (v0.106.0)

What feature would you like to see?

Summary

The compiled codex binary is 80MB on macOS arm64. Analysis of the source (codex-rs/) and the binary itself shows this is primarily due to heavy dependencies compiled unconditionally, with no feature gating for optional functionality. The runtime architecture is solid (async throughout, adaptive streaming, render caching); this is purely a dependency/packaging concern.

For comparison, the source is ~420K lines of Rust across 67 workspace crates, with a Cargo.lock containing 1,043 crates.

Findings

1. Heavy dependencies compiled unconditionally

Several large libraries are always included even though they serve optional functionality:

Dependency	Estimated binary contribution	Used for	Could be optional
image (jpeg, png, gif, webp)	~5-8MB	Clipboard paste, inline image viewing (~100 lines of app code)	Yes, via feature flag
syntect + two-face	~4-6MB	Syntax highlighting in TUI (highlight.rs, 1.3KB file)	Yes, could lazy-load themes
OpenTelemetry (5 crates)	~3-5MB	Observability, not used in basic CLI mode (#12913 notes it's not even emitting metrics)	Yes, via feature flag
rama (8 crates)	~2-4MB	HTTP/SOCKS5 network proxy for sandboxing; using ~10% of the framework	Yes, could use lighter alternative
sqlx with migrate feature	~1-2MB	State/history storage; migration tooling not needed at runtime	Partially, drop migrate feature

2. tokio compiled with features = ["full"]

The chatgpt crate depends on tokio with ["full"], which pulls in every tokio subsystem (io, net, fs, signal, process, sync, time, rt-multi-thread, macros). Other crates in the workspace use more targeted feature sets. This inflates the final binary.

3. Both regex and regex-lite included

regex = "1.12.3"
regex-lite = "0.1.8"

Both are workspace dependencies. regex-lite exists specifically as a smaller alternative to regex; having both defeats the purpose.

4. ts-rs (TypeScript type generation) compiled into the CLI binary

ts-rs = "11"

This is a build-time code generation library. If it's only used for generating TypeScript types from Rust structs, it should be a build dependency or gated behind a feature, not compiled into the release binary.

5. All four image formats compiled

image = { version = "^0.25.9", default-features = false, features = ["jpeg", "png", "gif", "webp"] }

For a CLI tool, JPEG and PNG would cover the vast majority of use cases. GIF and WebP decoders add binary weight for edge cases.

6. Core crate is a monolith

The core crate contains: all tool specifications, OpenAI integrations, skill framework, state management, authentication, configuration, models, and threading. Every binary in the workspace that depends on core pulls in all of this. Splitting it into core-api, core-tools, core-models etc. would allow dead code elimination to work more effectively.

What's already done well

Credit where it's due. The build configuration is excellent:

[profile.release]
lto = "fat"
split-debuginfo = "off"
strip = "symbols"
codegen-units = 1

And the runtime architecture is strong:

• Adaptive streaming with queue-depth-based gear switching (smooth/catch-up modes)
• Frame-rate-limited TUI rendering with transcript overlay caching
• Proper tokio::task::spawn_blocking() for sync operations
• Event-driven throughout, no blocking in main loop
• Platform-specific binary distribution (only the current platform's binary is installed)

Suggested improvements

High impact (estimated 15-25MB savings):

• Gate image behind a clipboard-images feature flag. Most CLI users never paste images.
• Gate OpenTelemetry behind an otel feature flag. It's not even emitting metrics currently (codex exec emits no OTel metrics; codex mcp-server emits no OTel telemetry at all #12913).
• Gate syntect behind a syntax-highlighting feature flag, or lazy-load theme databases.
• Evaluate whether rama can be replaced with a lighter proxy implementation for the subset of features used.

Medium impact (estimated 3-8MB savings):

• Replace ["full"] tokio features in the chatgpt crate with the specific features actually used.
• Remove migrate from sqlx features in the release binary.
• Consolidate regex and regex-lite to one.
• Move ts-rs to a build dependency or gate behind a codegen feature.
• Drop GIF and WebP from image features if usage data supports it.

Lower effort, longer-term:

• Split the core crate into smaller modules to improve dead code elimination.

Conservatively, feature-gating the top four dependencies could bring the binary from 80MB down to ~55-60MB with no loss of default functionality (users who need otel or image paste could opt in).

Additional information

Analysis performed on the open source codex-rs/ directory. Binary inspected: @openai/codex-darwin-arm64 v0.106.0 (Mach-O 64-bit arm64, 74MB __TEXT segment, 440 exported symbols, properly stripped).

Environment: macOS Darwin 25.2.0 arm64, Node v20.19.4 (npm install)