Removing line-number prefixes from file reads

GitHub measured 3.1% of total cost saved by removing theirs. FreeCode’s coding-suite A/B measured −10.9% tokens / −23.4% cost with zero quality loss.

The problem, in plain words

FreeCode’s read tool used to prefix every line of every file with its line number:

41: export function truncateOutput(text: string) {
42:   const limit = 500_000;
43:   ...

That’s 3–7 extra characters per line, on every file the model ever reads, re-sent on every later turn. The question is whether the model actually uses those numbers. In FreeCode, on paper, it doesn’t need to:

• The edit tool matches strings, not line numbers — you give it the old text and the new text.
• When the model needs a line number for navigation, it gets one from grep -n or LSP output, not from read.

GitHub found the same in Copilot’s harness: nothing in the editing workflow consumed the prefixes. But “on paper” is an argument, not evidence — maybe the model quietly leans on visible numbering to cite file:line in answers, or to disambiguate repeated strings before an edit. So this shipped as an experiment, not a change.

How I implemented it

One small, deliberately boring change (apps/core/src/tools/read.ts): the per-line prefix is applied only when FREECODE_READ_LINE_NUMBERS is not 0. The env var is read per call inside execute — not at startup — which matters because the A/B runner flips env vars between interleaved trials; a startup-read var would make both sides identical and the experiment would silently measure nothing. The key is allowlisted in the A/B runner’s VARIABLE_ENV_KEYS (apps/core/src/eval/ab.ts) for the same reason.

What stays regardless of the flag: the range footer (“Showing lines 1–200, use offset=201”). Paging through big files works off the offset parameter, not off per-line prefixes, so removing the prefixes doesn’t break re-reads.

The default stayed on until both measurements below were in — per the rule that a default is earned, not asserted. With both recorded, the default flipped on 2026-09-04: prefixes are now off, FREECODE_READ_LINE_NUMBERS=1 restores them.

Performance improvements

Coding suite, 11 cases × 3 trials per side, MiniMax-M3 on both sides, interleaved:

freecode eval ab coding --baseline env:FREECODE_READ_LINE_NUMBERS=1 \
  --candidate env:FREECODE_READ_LINE_NUMBERS=0 --trials 3

Every case passed every trial on both sides — removing the numbers loses nothing on correctness — and the no-prefix side was cheaper on every axis. It even repeated fewer calls, so there’s no hidden recovery detour eating the saving.

The second measurement — the judged suite, because citation quality (“does the model still say file.ts:42 in answers?”) is a judged property the coding suite can’t see — ran with Gemini 3.5 Flash Lite as the independent judge:

Both risk axes are answered: answer quality holds without the prefixes, and the savings repeat on a second suite. The experiment’s verdict is that the line numbers never earned their tokens.