Ten commits sit on k that are not upstream. All of them fix real defects found by running an 88k-line production bash script (kenji-cli) and its 98-suite BATS corpus under brush. This tracks getting them into reubeno/brush and in what order.
Verified against reubeno/brush@3a1de7ad (2026-08-21). k is 8 commits behind upstream at time of writing.
Before sending anything
Strip the (#N) suffixes from commit subjects. They are this fork's PR numbers, and they collide with unrelated real PRs upstream:
| our commit says |
reubeno/brush#N actually is |
| heredoc deadlock (#2) |
Bump prettydiff 0.6.4 → 0.7.0 |
| procsub deadlock (#5) |
Fix deny error |
| real-time signals (#15) |
Implement more builtins |
| Windows quoting (#17) |
Bump serde 1.0.201 → 1.0.202 |
GitHub renders these as links to those closed PRs. Left as-is, every PR opens with a confusing wrong reference.
Also rebase onto current main first, and re-run the suite per PR — applying cleanly is not the same as passing.
Status against current upstream
Every bug below is still live upstream; none was independently fixed.
setup_open_file_with_contents still does writer.write_all(bytes) before any reader exists- there is no
brush-core/src/sys/windows/commands.rs
- no real-time signal support in
sys/unix/signal.rs
Cherry-pick results onto 3a1de7ad: all ten apply, with conflicts only where a commit depends on an earlier one in its own chain (#2→#3, #13→#14→#15). Those chains apply cleanly in order. No conflicts with upstream drift.
Every functional commit carries compat tests with a bash oracle.
Order
Wave 1 — docs
|
|
docs(core): clarify process substitution task detachment |
10 lines |
docs(core): document builtin worker invariant |
6 lines |
Zero risk, and they show the code was read closely rather than patched blindly. Cheap first contact that makes the later reviews easier.
Wave 2 — deadlocks (highest value)
-
fix(core): avoid deadlock on here-documents larger than the pipe buffer
A hang, not a wrong answer. The entire here-document is written into a pipe before any reader exists, so the write only completes while the payload fits in the buffer. Bisected to an exact threshold: fine at 4096 bytes, hangs at 4160. Linux masks it via F_SETPIPE_SZ; Windows keeps the 4 KiB default. Also repairs two latent Linux gaps — the ?-propagated fcntl fails outright for payloads above /proc/sys/fs/pipe-max-size and for zero-length payloads.
-
fix(parser): honor backslash-newline line continuations in here-documents
Must follow (1) — same here.yaml. heredoc_escape_sequence handled \$, \` and \\ but not \<newline>, so continuations were emitted verbatim instead of joining lines.
-
fix(builtins): avoid deadlock when a builtin reads a process substitution
Independent of the two above. Another hang.
Deadlocks first because they are severe, self-contained, and each ships a bash-oracle test. Easiest "yes" a maintainer can give.
Wave 3 — signals, strictly in order
fix: report signals ignored at shell entryfix: preserve ignored signals at shell entryfix: support Linux real-time signals — 319 lines; send only after the first two land
Wave 4 — Windows argument quoting
fix(core): quote arguments passed to external commands on Windows
Highest review burden: ~300 lines of new platform code a maintainer may not be able to exercise locally. Kept last for that reason, not because it matters least — the bug is severe and silent. Arguments containing braces or escaped parens were corrupted en route to an external process, so grep -oE '[0-9]{3}' matched 23 instead of 123 and still exited 0. Wrong answers with a success status are worse than crashes.
Not upstreaming as a PR
chore: check out all text files with LF via .gitattributes — the only commit here that is not a bug fix. Adding a .gitattributes changes every contributor's working copy, which is a maintainer policy decision rather than a patch to drop on them. Raised as an issue upstream instead so they can decide. Stays on k regardless.
Caveat
The verification above confirms these apply to current upstream and that the bugs are still present. Each PR still needs a build and full-suite run after rebasing.
Ten commits sit on
kthat are not upstream. All of them fix real defects found by running an 88k-line production bash script (kenji-cli) and its 98-suite BATS corpus under brush. This tracks getting them intoreubeno/brushand in what order.Verified against
reubeno/brush@3a1de7ad(2026-08-21).kis 8 commits behind upstream at time of writing.Before sending anything
Strip the
(#N)suffixes from commit subjects. They are this fork's PR numbers, and they collide with unrelated real PRs upstream:reubeno/brush#Nactually isGitHub renders these as links to those closed PRs. Left as-is, every PR opens with a confusing wrong reference.
Also rebase onto current
mainfirst, and re-run the suite per PR — applying cleanly is not the same as passing.Status against current upstream
Every bug below is still live upstream; none was independently fixed.
setup_open_file_with_contentsstill doeswriter.write_all(bytes)before any reader existsbrush-core/src/sys/windows/commands.rssys/unix/signal.rsCherry-pick results onto
3a1de7ad: all ten apply, with conflicts only where a commit depends on an earlier one in its own chain (#2→#3,#13→#14→#15). Those chains apply cleanly in order. No conflicts with upstream drift.Every functional commit carries compat tests with a bash oracle.
Order
Wave 1 — docs
docs(core): clarify process substitution task detachmentdocs(core): document builtin worker invariantZero risk, and they show the code was read closely rather than patched blindly. Cheap first contact that makes the later reviews easier.
Wave 2 — deadlocks (highest value)
fix(core): avoid deadlock on here-documents larger than the pipe bufferA hang, not a wrong answer. The entire here-document is written into a pipe before any reader exists, so the write only completes while the payload fits in the buffer. Bisected to an exact threshold: fine at 4096 bytes, hangs at 4160. Linux masks it via
F_SETPIPE_SZ; Windows keeps the 4 KiB default. Also repairs two latent Linux gaps — the?-propagated fcntl fails outright for payloads above/proc/sys/fs/pipe-max-sizeand for zero-length payloads.fix(parser): honor backslash-newline line continuations in here-documentsMust follow (1) — same
here.yaml.heredoc_escape_sequencehandled\$,\`and\\but not\<newline>, so continuations were emitted verbatim instead of joining lines.fix(builtins): avoid deadlock when a builtin reads a process substitutionIndependent of the two above. Another hang.
Deadlocks first because they are severe, self-contained, and each ships a bash-oracle test. Easiest "yes" a maintainer can give.
Wave 3 — signals, strictly in order
fix: report signals ignored at shell entryfix: preserve ignored signals at shell entryfix: support Linux real-time signals— 319 lines; send only after the first two landWave 4 — Windows argument quoting
fix(core): quote arguments passed to external commands on WindowsHighest review burden: ~300 lines of new platform code a maintainer may not be able to exercise locally. Kept last for that reason, not because it matters least — the bug is severe and silent. Arguments containing braces or escaped parens were corrupted en route to an external process, so
grep -oE '[0-9]{3}'matched23instead of123and still exited 0. Wrong answers with a success status are worse than crashes.Not upstreaming as a PR
chore: check out all text files with LF via .gitattributes— the only commit here that is not a bug fix. Adding a.gitattributeschanges every contributor's working copy, which is a maintainer policy decision rather than a patch to drop on them. Raised as an issue upstream instead so they can decide. Stays onkregardless.Caveat
The verification above confirms these apply to current upstream and that the bugs are still present. Each PR still needs a build and full-suite run after rebasing.