Skip to content

Latest commit

 

History

History
39 lines (30 loc) · 2.2 KB

history.md

File metadata and controls

39 lines (30 loc) · 2.2 KB

How ts-rust got here

This text was in the README before the first preview release (2026-10).

The model experiment

This repository is also an experiment in how capable gpt-5.6-sol is at understanding, validating, and completing a compiler-sized systems port through Codex and its subagents.

There is an important provenance caveat: the archived Codex log labels 236 of the original implementation turns as an earlier model, one later turn as gpt-5.5, and only the brief July 8 resume as gpt-5.6-sol. The existing code is therefore the starting artifact for the gpt-5.6-sol experiment, not evidence that every existing line was produced by that model name.

The first goal runs (June to July 2026)

The implementation logs show three distinct Codex goal runs:

Goal run Outcome Goal-accounted tokens
Full Rust port, June 22–26 Produced the broad prototype. A later audit found 70 failing tests and measured the broad compiler 2.6–3.4x slower than Go for equivalent work. Paused. 207,170,354
minimal-working-v0, June 26 Narrowed the claim to the explicit five-file no-check contract and built its oracle, runtime, and benchmark gates. Completed. 1,439,471
Type-checking parity, June 27–July 8 Added substantial checker work, but never demonstrated full tsgo parity. Paused. 7,438,681
Total 216,048,506 (~216 million)

The estimate uses Codex's goal-level tokensUsed counters for the three runs that were explicitly tied to this repository. Those counters equal uncached input plus output tokens. The raw session log records about 8.34 billion input-plus-output tokens, but about 8.12 billion of those are cached context replay, so that larger number is not a useful estimate of new inference spent on the project. Work between goals and this README update are not included.

The history also explains why breadth is not the same as completion here. The first goal landed hundreds of commits across many compiler subsystems before the parity harness was authoritative. A recovery audit then reduced the scope to one measurable path. The later checker goal expanded the scope again and was suspended before cleanup and full verification were complete.