shell: only treat digit as fd-prefix redirect at word boundary

[robobun]

What does this PR do?

Fixes the shell lexer treating a digit at the end of a word as an fd-prefix when followed by > / <.

Repro

$`echo abc1>file`       # wrote "abc\n" — lexed as `echo abc` + `1>file`
$`echo test100>out`     # wrote nothing — lexed as `echo test10` + `0>out` (stdin redirect)
$`echo "abc"1>file`     # wrote "abc\n" — lexed as `echo "abc"` + `1>file`
$`./script1<input`      # ran `./script` with `1<input`

POSIX only recognizes an IO number when the digits form their own word. bash writes abc1\n, test100\n, abc1\n, and runs ./script1 respectively.

Root cause

In the '0'...'9' arm of the lexer (src/shell/shell.zig), eat_redirect() was attempted unconditionally whenever a digit appeared in Normal state, then break_word() was called after the redirect was accepted — silently splitting the trailing digit off whatever word preceded it.

Fix

Before entering the fd-redirect path, require that the digit begins a new word:

1. self.word_start == self.j — no text currently being accumulated, and
2. the preceding token (if any) is not a word-continuation token (Text, Var, SingleQuotedText, DoubleQuotedText, CmdSubstEnd, Asterisk, brace tokens) — same classification break_word_impl already uses.

Digits that genuinely begin a word — after whitespace (echo 2>file), at start of input (2>file), or after an operator (echo foo;2>file) — still parse as fd prefixes.

How did you verify your code works?

• USE_SYSTEM_BUN=1 bun test test/js/bun/shell/lex.test.ts -t "op_redirect digit" → fail
• bun bd test test/js/bun/shell/lex.test.ts → 30/30 pass
• bun bd test test/js/bun/shell/parse.test.ts → 17/17 pass
• bun bd test test/js/bun/shell/file-io.test.ts → 25/25 pass
• bun bd test test/js/bun/shell/bunshell.test.ts -t redirect → 28/28 pass
• bun bd test test/regression/issue/12602.test.ts → 3/3 pass

Closes #12602

[coderabbitai]

Walkthrough

A control-flow change in the Zig shell lexer prevents digits that are part of an in-progress word from being parsed as file-descriptor redirections. New regression tests validate parsing of command names with trailing digits versus standalone FD redirects.

Changes

Cohort / File(s)	Summary
Shell Parsing Logic src/shell/shell.zig	Added a check that an accumulating word is not in-progress (word_start != j) before attempting to parse a leading digit as an fd redirection, avoiding misinterpretation of digits inside words.
Regression Tests test/regression/issue/12602.test.ts	Added three tests (two skipped on Windows) that verify commands like ./script1<input.txt keep the trailing digit in the command name, while constructs like cat 0<input.txt are treated as FD redirects.

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The PR successfully addresses issue #12602 by fixing the tokenizer to check whether digits are part of an already-accumulating word before parsing as fd redirects, with comprehensive regression tests validating the fix.
Out of Scope Changes check	✅ Passed	All changes are directly related to fixing the shell redirect parsing issue described in #12602; no unrelated modifications are present.
Title check	✅ Passed	The title accurately describes the main change: preventing digits from being treated as fd-prefix redirects when they're part of an existing word, which directly addresses the core fix.
Description check	✅ Passed	The description includes both required sections: detailed explanation of what the PR does with specific examples and repro steps, plus comprehensive verification with test results.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[robobun]

^{Updated 8:53 PM PT - May 1st, 2026}

❌ @robobun, your commit 58599ca has 3 failures in Build #49939 (All Failures):

• test/js/bun/s3/s3-storage-class.test.ts - code 1 on 🍎 14 aarch64
• test/cli/run/no-orphans.test.ts - timeout on 🍎 14 x64
Details
test/cli/run/no-orphans.test.ts - timeout on 🍎 26 aarch64
Details
test/cli/run/no-orphans.test.ts - timeout on 🍎 14 aarch64
• test/bake/dev-and-prod.test.ts - code 1 on 🪟 2019 x64-baseline

---

🧪   To try this PR locally:

bunx bun-pr 27257

That installs a local version of the PR into your bun-27257 executable, so you can run:

bun-27257 --bun

[claude]

Newest first

✅ 0a2a7 — Looks good!

Reviewed 2 files across src/shell/ and test/regression/issue/: Fixes the shell tokenizer to correctly preserve trailing digits in command names (e.g., ./script1) when followed by redirect operators, preventing the digit from being incorrectly stripped and misinterpreted as a file descriptor redirect number.

Previous reviews

✅ 08afc — Looks good!

Reviewed 2 files across src/shell/ and test/regression/issue/: Fixes the shell tokenizer to correctly preserve trailing digits in command names when followed by redirect operators, preventing incorrect parsing where commands like ./script1 before < or > would have their trailing digit stripped and misinterpreted as a file descriptor redirect.

Previous reviews

✅ d9f72 — Looks good!

Reviewed 2 files across src/shell/ and test/regression/issue/: Fixes the shell tokenizer to correctly preserve trailing digits in command names when followed by redirect operators, preventing incorrect parsing like ./script1<file being misinterpreted as a file descriptor redirect.

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@test/regression/issue/12602.test.ts`:
- Around line 15-20: Capture and assert the exit code of the setup chmod command
and add one negative redirect case: assign the chmod invocation to a variable
(e.g., chmodRes = await $`chmod +x ${dir}/script1`.quiet()) and assert
chmodRes.exitCode === 0; keep the existing successful run (result = await $`cd
${dir} && ./script1<input.txt`.quiet()) and its output assertion, then add a
failing redirect test by running the command against a missing file with
.nothrow() (e.g., bad = await $`cd ${dir} && ./script1<missing.txt`.nothrow())
and assert bad.exitCode !== 0 to cover the error path.

[robobun]

Addressed review feedback in 7ac36fe + 58599ca:

• Added .DoubleAsterisk to the compound-word guard so **2>file keeps the 2 in the glob word
• Moved .DoubleAsterisk into the => true arm of break_word_impl so whitespace after ** emits a .Delimit (matching *) — fixes echo ** 2>file lexing as stderr redirect, and stops echo ** foo from gluing into one compound word
• eat_redirect's '<' arm now consumes the < before checking for <<, so N<file no longer lexes with a spurious append=true and N<<file lexes as a single redirect token
• Dropped the unnecessary Windows skip on the cat 0<input.txt regression test

All shell tests pass on every platform (lex 31/31, parse 17/17, 12602 3/3 on darwin and 1+2skip on Windows).

CI reds are unrelated flakes also hitting other open PRs: no-orphans.test.ts perl-setsid timeout on macOS (from #29930, also failing on #30089), bake/dev-and-prod.test.ts HMR race on one Windows shard (#28211/#29575), plus several darwin shards that expired waiting for runners.