Port of the TypeScript compiler, checker and lsp to Rust, by LLM

TL;DR Highlight

LLMs ported tsc to Rust; Opus 5.5 got furthest in two weeks for about $24k

Who Should Read

Backend or tooling developer weighing cost and quality of letting coding agents do a large port or rewrite

Core Mechanics

OpenAI models spent $400k and stalled near 84%
- GPT-5.6 Sol and GPT 6 Astra used over $400,000 of API-priced tokens.
- Months of /goal loops produced over 1.3M lines of Rust but never got past roughly 84% compatibility.
Opus 5.5 restarted from scratch and went further
- It had a working v0 in 10 hours.
- The author expected it to reuse the Codex code, but it started over.
- It got further than Astra in 1/10th the time.
Two weeks cost about $24,047
- That is API-priced spend, though the author used Claude accounts.
- It equaled 925% to 983% of his $200 plan's weekly limits.
It is a direct port of the Go tsc 7
- It keeps the algorithms and behavior of Microsoft's native Go TypeScript compiler.
- It offers the same CLI (tsc), language server, and API, and all 184,773 ported Go tests pass.

Community Reactions

Skepticism about adoption value
- One commenter asked why use a less maintained, less tested implementation.
- Another said the real question is what this teaches us.
Real-world speedup is small in absolute terms
- A commenter noted real-world project times were only 1-3s better than tsc 7.
- They said percentages can look more impressive than absolutes.
Rust port reimplements Go semantics, hurting performance
- A commenter said it reimplements Go-style strings and much of the Go standard library.
- They argued it jumps through hoops to preserve Go behavior.
Concerns about verification and maintenance
- A commenter worried about holes where original authors thought tests unnecessary.
- Another recalled Bun's rewrite taking 3 months of human labor to fix issues before release.
- A TypeScript team member said they were surprised 3 such ports appeared in a week.

How to Apply

Measure test pass rate first
- Run an agent loop that raises the pass rate against the existing test suite of a library you want to port.
- Record the pass rate numerically to compare models.
Compare models on a small module
- Give the same small port to two models and record time and token cost.
- Also try a new model starting from scratch instead of inheriting stuck code.
Verify the port with output diffs
- Write a script that feeds the same inputs to original and port and compares outputs across several projects.
- Document mismatches as a known-problems list.
Adopt a marker for LLM-written code
- Mark LLM-written sections in the README like the "Slop Line".
- State the verification scope so users can judge the risk.

Try It

copy & run

# Paste into Claude Code (simplified version of the article's method: test pass-rate goal loop)

Goal: port module {{module name}} in {{original library path}} to {{target language}}.
Rules:
- Keep the original algorithms and behavior.
- Port the original tests and run them with {{test command}}.
- Print passed/total tests at each iteration.
- Stop at 100% or if the rate does not improve for 3 iterations in a row, and summarize what blocked you.

Show me the port plan and file structure first, and start after I approve.

Next to try: when done, have it feed 20 identical inputs to the original and the port and list the outputs that differ.

Terminology

/goal loopA way to give an agent a goal and let it keep running until it is achieved. It is like leaving homework to finish overnight.

PGOA build method where the compiler uses real run profiles to optimize hot paths. It is like paving the roads people use most first.

work stealingA scheme where idle workers take leftover tasks from busy workers. It is like a free cashier helping a long line.

Type checkerThe part of a compiler that verifies types and reports errors. It works like a spell checker.

Effect diagnosticsLanguage-service diagnostics specific to the Effect library. tsc-rs builds them in, so no second compiler is needed.