Changelog

All notable changes to this project. Format follows Keep a Changelog; this module aims for Semantic Versioning.

Unreleased

Documentation and help, the metadata drift that made them inaccurate, and the two defects that stopped a Windows build from being usable.

Added

• README.md ### Windows: the PowerShell recipe ($env:CC instead of CC=…), which compilers work and what they link against, the .dll naming convention, and what a successful install prints.

• Build.rakumod MAIN, so raku -I. Build.rakumod — the command the README documents — actually runs. Previously only zef install . reached Build.build.

• docs/API.md — every public method with arguments and defaults, one table per module, plus the exception hierarchy and the environment variables.

• examples/serve.raku — the full server with a real mainline: --port, --host, --verbose, --demo, and a generated --help. Routes cover path parameters (/users/:id), a wildcard (/files/*), JSON (/health) and a body echo (POST /echo).

• examples/client.raku — the pooled client as a script: --url, --method, --data, --type, --timeout, --verbose, and --self-test, which starts a server in the same process, issues a GET and a POST, and exits non-zero on failure. No network required.

• examples/registry-watch.raku — the connection-registry leak as an executable check: four points, expected versus observed counts, ok / FAIL per line.

• CHANGELOG.md.

Changed

• Build no longer hard-codes zig cc as the default compiler. It tries zig cc, cc, gcc, clang in that order and keeps the first candidate that actually produces the library, because a bare zef install . on a machine without zig died in Building while an installed gcc sat there unused — and because probing with --version is not enough: WSL inherits the Windows PATH, so a zig.exe sitting there answers happily and then fails to build an ELF target. CC still overrides the search, and an unusable CC now fails immediately with the command it ran.
• README.md restructured: table of contents, an install section that names the library search order, a quick start, an examples table, a troubleshooting section built from failures that actually occurred, and pointers into the benchmark post-mortems.
• The test command is now prove -e 'raku -Ilib' t/*.rakutest, which needs no extra distribution; prove6 is mentioned as an alternative for those who have it installed.

Fixed

• examples/serve.raku and examples/client.raku only accepted --option=value for the numeric options. MAIN binds a space-separated value only to a Str option, so --port 8080, --demo 3 and --timeout 5 arrived as stray positional arguments and the scripts printed their usage block instead of running. The numeric options are now declared Str and validated by int-option / num-option: both spellings work, and a value that is not a number is reported by option name instead of falling through to the usage block. README.md and both script headers describe the two forms.
• README.md carried the AIGC content label twice; one copy is kept at the top.
• META6.json listed 18 of the 22 provided modules. HTTP::Core::Bytes, HTTP::Core::Codec::Content, HTTP::Core::Codec::Raku and HTTP::Core::Transport::TLS were missing.
• META6.json resources named libraries/http_wire.so, which is the Windows stem on a Unix platform; the build actually emits libhttp_wire.so and libhttp_wire.dylib. Both are now declared with their real names — and, as it turned out, naming files at all was the mistake; see the resource-key fix below.
• META6.json support.source pointed at raku-http-core, the repository is raku-HTTP-Core.
• META6.json did not declare Compress::Zlib, which HTTP::Core::Codec::Content loads at the top of the file. Machines that happened to have it installed (zef install . on this one) hid the omission; on a clean one zef staged the distribution and every use HTTP::Core afterwards died on the missing module.
• Build named every Unix kernel http_wire.dll. The platform test was written as given $*DISTRO.is-win { when True { … } default { … } }, and when True in Raku matches any value, False included, so the Windows branch won on every platform: an ELF or Mach-O image came out with a .dll name, and the binding went looking for a file whose name contradicted its contents. It is an if now.
• HTTP::Core::Native resolved its library path in a constant at compile time. zef install . compiles inside a staging repository and deletes it afterwards, so Testing passed and the very next use HTTP::Core died with Cannot locate native library 'http_wire' (error 0x7e) pointing at a directory that no longer existed. Resolution is now deferred to the first call (is native(&_find-lib)), which is also what makes library-path() honest after installation.
• HTTP::Core::Native._find-lib searched the resource table before the working tree, and matched resource keys by prefix — so libraries/http_wire resolved to the http_wire.pdb that zig cc drops beside the DLL, and the binding tried to dlopen a debug database. The order is now %*ENV<HTTP_CORE_LIB> → working tree → resource table → loader, and every candidate is checked against the platform library suffixes.
• META6.json resources listed the three build outputs by full name, which cannot describe a distribution built on a fourth platform and did not match what zef stages (resources/<hash>.dll). It now declares the platform-independent key libraries/http_wire; zef appends the extension.
• t/14-serialise-threads.rakutest compared the C fast path against the Raku builder byte for byte, Date header included, so a clock tick between the two payloads failed the test roughly once in a hundred runs. Date is normalised on both sides before the comparison.
• The Distribution::Resource deprecation notice is now explained rather than left hanging. Every path accessor on a resource object (.absolute, .Str, .IO, .resolve) is deprecated in Rakudo 2026.09 in favour of "using %?RESOURCES<key> directly", which cannot work: the value is a resource object and NativeCall needs a string. Both would-be replacements were tried and rejected — $?DISTRIBUTION.content('resources/libraries/…') returns Any for an installed distribution, and .open merely reaches the same deprecated Str from inside. The one call stays, README.md Troubleshooting gained the reasoning plus how to mute the notice deliberately (RAKUDO_NO_DEPRECATIONS=1, verified on 2026.09) instead of editing code that works.

0.1.0

Added

• C kernel (src/http_wire.c) exported through nine functions and bound with NativeCall, no shim: request head parsing, flat arena out-parameters, request and response writing.
• HTTP::Core facade: parse-request, encode-request, parse-response, decode-response, response-extent, core-version, library-path, assert-abi.
• Codec for all four directions and all four body framings (no-body, content-length, chunked, until-close), with a string-building oracle (Codec::Raku) the C path is tested against byte for byte.
• Router with :params and * wildcards, 405 with an Allow header.
• HTTP::Core::Server: keep-alive, pipelining, partial-body dispatch, max-head / max-body limits, connection accounting.
• HTTP::Core::Client with redirects, per-request deadlines and a connection pool (HTTP::Core::Pool) that reports reuse metrics.
• Transports: TCP, TLS, and an in-memory transport for tests.
• Test suite: 17 files, 101 tests.

Fixed

• Connection registry leak: !drop used a Str key against an Int transport id, threw inside the frame callback, and so never deregistered; the key was also a recyclable socket address. The registry is now keyed on a monotonic Connection.serial (55c8bae, cc2c5e2, c8d9af9), with t/17-registry-release.rakutest as the regression test and bench/RESULTS-CRO.md as the write-up.
• Lost wakeup in the connection pool (5998387).