feat(grep): treat - operand as standard input

[trieloff]

Summary

grep did not accept - as a FILE operand. Found while reviewing #327 (grep -f), which added -f - for reading patterns from stdin and made the neighbouring gap obvious. There is no issue for it.

$ printf 'apple\n' | grep apple -
grep: -: No such file or directory     # just-bash, exit 2
apple                                  # GNU grep 3.12, exit 0

- is a POSIX-documented operand meaning standard input, and the busybox grep spec suite already exercises it — six cases were parked in spec-tests/grep/skips.ts with the reason "- stdin arg not supported".

This PR implements the GNU semantics. Every expectation below was verified against GNU grep 3.12 (/opt/homebrew/opt/grep/libexec/gnubin/grep), never macOS BSD grep.

Details

The label. GNU prints (standard input) wherever a real file name would appear — the multi-file file:line prefix, -l/-L listings, -c counts, and the file-line form of context lines. (Context output with -A/-B/-C is not exercised in the new tests: just-bash omits GNU's -- separator between context groups, a pre-existing divergence that predates this PR — grep -A1 pat f1 f2 on main already omits it, - or no -.)

$ printf 'apple stdin\n' | ggrep -n apple f1.txt - f2.txt
f1.txt:1:apple
(standard input):2:apple stdin
f2.txt:2:apple pie

A lone - gets no prefix (one operand), and -h suppresses it exactly as it does for files. -l/-L still print (standard input) under -h, since those are the output rather than a prefix.

Stream semantics for repeated -. stdin is a stream, so only the first - sees content; every later one reads EOF:

$ printf 'apple\n' | ggrep -c apple - - -
(standard input):1
(standard input):0
(standard input):0

Implemented by marking each queued - entry with stdinAtEof at operand-collection time, so the parallel batch executor stays deterministic.

- with -r. GNU only forces the file-name prefix under -r when recursion can actually descend into a directory, and stdin is never a directory: grep -r pat - prints bare lines, while grep -r pat - sub prefixes both sources (two operands). showFilename now requires a real path target before -r forces the prefix. This deliberately leaves the pre-existing grep -r pat single-file.txt divergence (just-bash prefixes, GNU does not) untouched — out of scope here.

- bypasses name-based filtering. Glob expansion, recursion and --include/--exclude all filter on a file name; stdin has none, so GNU applies none of them. printf 'apple\n' | ggrep --exclude='*' apple - still prints apple.

Empty vs absent stdin. Empty stdin is a valid empty stream: no output, -c prints 0, exit 1 — same as GNU with < /dev/null. When just-bash has no stdin at all, - reads empty, consistent with how cat - already behaves in this repo.

Exit codes. Unchanged rules, now fed by stdin too: 0 on match, 1 on none, 2 when any operand errored. A missing operand alongside - still searches stdin and exits 2; -q short-circuits to 0 on a match even then, per SUSv3 (busybox spec case at busybox-grep.tests:55).

Two adjacent pre-existing divergences are deliberately not addressed (both reproduce on main without -): grep pat - dir exits 0 where GNU exits 2 (directory operands don't set the error flag), and -q suppresses the No such file or directory diagnostic that GNU still prints to stderr — the new -q test asserts just-bash's current (empty) stderr, and the corresponding fixture routes stderr to /dev/null, so neither pins the GNU wording.

-H is not implemented in just-bash at all (it is rejected as an unknown option, on main too). Out of scope; -h interaction is covered.

Composition with #327 (grep -f). Both -f - and a - operand consume stdin, and in GNU the first consumer wins — patterns are read before input, so -f - drains the stream and the - operand then reads EOF:

$ printf 'a\n' | ggrep -f - -
$ echo $?
1

The two PRs are independent (this one is based on upstream/main, not on #327). #327 renames files → operands in the arg loop and introduces stdinUsedForPatterns; this PR adds a let stdinConsumed = false next to the operand-expansion loop. Whichever lands second needs one mechanical edit: initialise stdinConsumed from stdinUsedForPatterns instead of false, which reproduces GNU's first-consumer-wins behaviour for free. No other overlap — #327 touches the pattern-collection half of the function, this PR touches the operand-expansion and per-file-read half.

Composition with #314 (repeated -e). #314 also renames files → operands and adds combinePatterns; it does not touch operand expansion, showFilename or the per-file read. Both PRs delete from spec-tests/grep/skips.ts, but from non-adjacent hunks (#314 removes the "Multiple -e patterns" block, this PR the "- for stdin argument position" block). Rebase is mechanical.

Tests

GNU grep version used for every expectation and every fixture: GNU grep 3.12 (Homebrew, /opt/homebrew/opt/grep/libexec/gnubin/grep).

• packages/just-bash/src/commands/grep/grep.stdin-operand.test.ts — 28 unit tests, 286 lines, full-string stdout/stderr assertions plus exit codes. Covers the exact repro, the (standard input) label across -n/-l/-L/-c/-o/-v, -h, repeated - (2× and 3×), -m, -r prefixing, --include/--exclude exemption, -- terminator, -e, - as PATTERN, empty stdin, absent stdin, and the missing-operand + -q exit codes.

• packages/just-bash/src/comparison-tests/grep-stdin-operand.comparison.test.ts + fixtures/grep-stdin-operand.comparison.fixtures.json — 27 comparison cases / 27 fixtures, all "locked": true so BSD-grep boxes cannot silently re-record them. The re-record command is documented in the test file header:

  PATH=/opt/homebrew/opt/grep/libexec/gnubin:$PATH \
    RECORD_FIXTURES=force pnpm test:run \
    src/comparison-tests/grep-stdin-operand.comparison.test.ts
  

• packages/just-bash/src/spec-tests/grep/skips.ts — 4 busybox spec tests unskipped and now passing: grep - (specify stdin), grep - nofile (specify stdin and nonexisting file), grep -q - nofile (... match), grep -L exitcode 0 #2. Two entries stay skipped with corrected reasons: grep - infile (the spec runner rewrites input to /tmp/input, so the file-name prefix differs — a runner artifact, not a grep bug) and grep -s nofile - (-s is not implemented).

Verification numbers on this branch:

• pnpm typecheck — clean; pnpm lint (biome + workflow security + lint:banned) — clean; pnpm knip — clean (2 pre-existing configuration hints only).
• pnpm test:comparison — 36 files, 594 passed, 0 failed.
• grep + rg + search-engine + grep spec suites — 33 files, 1158 passed, 90 skipped, 0 failed (grep spec alone: 332/332).
• pnpm test:run — 14342 passed, 97 skipped, 6 failed. The 6 failures are pre-existing python3/WASM sandbox tests (just-bash.bundle, code-exec-exploit-regression, defense-in-depth-independence, python-sqlite-information-disclosure ×2, worker-protocol-runtime-desync); confirmed identical (6 failed / 26 passed across those 5 files) on a stashed, unmodified tree.

Not touched: -L's exit-code inversion (GNU keys -L's exit status off whether a line was selected, not off whether anything was printed) is being fixed separately. Every -L case here is one where both the current and the corrected rule agree.

🤖 Generated with Claude Code

https://claude.ai/code/session_019DayCPuYZEmv4VszJXTHT3

[vercel]

@trieloff is attempting to deploy a commit to the Vercel Labs Team on Vercel.

A member of the Team first needs to authorize it.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.