feat(awk): rewrite awk to match gawk 5.2

[chaliy]

Requested by Михайло · project thread

What changed

awk (and gawk/mawk/nawk) is rewritten to behave like gawk 5.2, which is Debian's awk.

• Language. All of POSIX awk is supported. The gawk extensions agents use are supported too: gensub, asort/asorti, patsplit, strftime/mktime/systime, BEGINFILE/ENDFILE, switch, arrays of arrays, PROCINFO["sorted_in"], FIELDWIDTHS/FPAT, IGNORECASE, regex RS with RT, typeof and the bit functions.
• Commands work. system(), print | cmd and cmd | getline run in the sandbox shell through an execution-plan driver, the same mechanism make uses. Output stays in order.
• gawk output. Number printing, printf (%c, %i, *, %n$, inf/nan), rand()'s sequence, regex matching (POSIX leftmost-longest) and substr/getline/typeof edge cases all follow gawk.
• gawk diagnostics. Syntax errors print awk: cmd. line:N: with a caret and exit 1. Runtime errors print located fatal: messages and exit 2. Usage goes to stderr with exit 1.
• Options. -f (repeatable), -e/--source, --, -F t, -v with escapes, --csv, and gawk's no-op flags are accepted.
• Limits. Every existing resource cap is kept with the same messages. split() and record splitting now stop at the memory cap instead of first building a huge vector.

Why

pseudo-linus runs a Debian-oracle bench, and on it bashkit's awk matched only 51.5% of 260 cases. It was the weakest common text tool there. The old design could not reach gawk: it was a sync evaluator that hopped threads for I/O, arrays were stored as "name[key]" variables, function locals were broken, and commands were unsupported.

Before / After

Debian-oracle differential run (gawk 5.2.1 goldens, 260 awk cases):

before: awk | 260 | 50.8% strict | 51.5% lenient
after:  awk | 260 | 92.7% strict | 93.8% lenient

• All remaining failures but one invoke mawk by name and expect its own dialect (-W options, messages, substr, division by zero; L-AWK-004).
• The one gawk case left is unsorted for (k in a) order, which follows gawk's hash tables (L-AWK-001). That is a follow-up.

$ awk 'BEGIN { print "before"; r = system("echo inside; exit 3"); print "after", r }'
before:  before / after          (command silently not run, empty status)
after:   before / inside / after 3

$ printf 'b\na\n' | awk '{ print | "sort" } END { close("sort"); print "done" }'
before:  awk: pipe output redirection (|) is not supported   (exit 2)
after:   a / b / done

$ awk 'BEGIN { print ( }'
after:   awk: cmd. line:1: BEGIN { print ( }
         awk: cmd. line:1:                 ^ syntax error      (exit 1)

• Tests: bashkit lib 2837 and integration 1346 pass. awk spec 144/144, including new cases for system, pipes, cmd | getline, syntax errors and substr.
• Stack: the depth-64 recursion test needed about 3 MiB of debug-build stack in the first cut. Splitting each expression and statement kind into its own boxed future brought it under 512 KiB.

Risk

• Medium. This is a full rewrite of a heavily used builtin. It is covered by the existing unit, spec and security suites plus the oracle run.
• Intentional output changes toward gawk. A missing input file now exits 2. --csv no longer sets OFS to ,. substr(s, 0, n) keeps n characters. Redirected getline leaves NR alone. -0 prints 0. Non-ASCII identifiers are now a syntax error.
• Commands through Builtin::execute. Direct calls have no shell, so commands report 127 with a notice (L-AWK-003).

Checklist

• Tests added or updated (unit tests per module, spec cases, updated security/regression tests where gawk differs)
• Backward compatibility considered (gawk fidelity wins; caps and their messages unchanged; knowledge, limitations L-AWK-001..004 and TM-DOS-128 updated)

https://claude.ai/code/session_01VscpEzxtqpCKx5eRMpywNj

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	bashkit	8bbe6de	Commit Preview URL Branch Preview URL	Oct 07 2026, 06:55 PM