Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

0.1.0 - 2026-10-20

Added

The pycurl console script and python -m curlike: the drop-in curl (curlike.shim.main). main(argv=None, *, entry=None) takes the executable acting as this curl, whose directory is skipped when looking for the real curl and which hook wrappers receive as _HOOK_CURL.
curlike.main(), curlike.run_fallback() and curlike.shim.main() take ssl_context= / proxy_ssl_context=: a caller's ssl.SSLContext used as given for every https origin or an https:// proxy (mutual TLS, ciphers, protocol versions). The TLS flags it replaces are refused beside it (exit 2).
With no proxy variable set, the system's proxy configuration routes each URL on the command line (through syshttp: WinHTTP, CFNetwork, libproxy, PAC, WPAD); its redirect hops keep that decision, as with curl's own --proxy. The real curl is handed that proxy with --proxy (curl reads only the *_proxy variables). SYSHTTP_SYSTEM_PROXY=0 turns it off. Of a list (PROXY a; PROXY b; DIRECT) the first proxy that accepts a connection is used; with none and no DIRECT, exit 7 and real curl is not run.

Changed

curlike is a pure-Python curl and never runs the real one: pycurl and python -m curlike are curlike.main. The launcher that prefers the real curl -- hook wrappers, the agent proxy handed to curl as --proxy, the _CURL, _CURL_SSL_REVOKE and _PYTHON settings -- moved to the dotagents net overlay (lib/curlshim). New for such a launcher: curlike.caller_steering(), requested_url(), find_real_curl() and describe_curl().
CI installs syshttp at one pinned commit (the 0.1.0.dev7 floor) in every workflow, so a run is reproducible and the floors job tests the floor.
There is no hidden 30 s time limit: --connect-timeout bounds the connect only, -m is a deadline for the whole attempt (redirect hops and a trickling body included), and --timeout (not a curl option) is opt-in. -N is real, and the "Output written to FILE" line, which curl never prints, is gone.
The default request is curl's: User-Agent: curl/8.22.0 and Accept: /, and no Accept-Encoding: identity. Connection: close remains, the one difference, because each transfer is one request on one connection.
Two URLs, an unknown option and a missing option value are curl-style one- line usage errors (exit 2, with the --help hint) instead of argparse's 52-line usage dump.
The duho floor is 0.7.0, the first release with duho.Meta (the suite passes at it); -v is documented as printing the request as given, the caller's Authorization and Cookie included, with only the configured proxy's credential redacted.
The test run is warning-clean: the fake servers close the connections they open, and the test configuration turns every warning, ResourceWarning included, into an error.
Settings are read as syshttp 0.1.0.dev19, the new floor, reads them: every variable has a full name, read as <PREFIX><NAME> under each prefix of SYSHTTP_PREFIX, a :-separated list of literal prefixes (none by default: the names as written). Where the entries here write _NAME, the variable is SYSHTTP_NAME. SYSHTTP_PREFIX=AGENTS_ (it was AGENTS) reads AGENTS_SYSHTTP_PROXY, AGENTS_HTTPS_PROXY, AGENTS_NO_PROXY and AGENTS_HTTP_PROXY_AUTH and nothing unprefixed: the conventional proxy variables used to ignore the prefix, and now follow it. AGENTS_: reads the prefixed names first and falls back to the plain ones. CURL_CA_BUNDLE is curl's own variable and is read as curl names it, under no prefix.
_PYLIB is CURLIKE_PYLIB, a full name read under each prefix like the others; SYSHTTP_PYLIB is no longer read.
no_proxy is read before NO_PROXY, as curl does (it was the other way round, which showed only when both were set to different lists).
A proxy setting that is refused (exit 2) names the variable that was found, under its prefix: AGENTS_SYSHTTP_PROXY_TYPE for SYSHTTP_PREFIX=AGENTS_.
Cookies follow syshttp's stricter rules, which agree with curl 8.21 on the cases measured: a Set-Cookie header with a control character (the tab apart) is ignored whole, a -b file's row with one in its name or value is skipped, and -c writes no row with a control character or a tab in a field.
The syshttp floor is 0.1.0.dev19 (it was 0.1.0.dev14) and the CI workflows install its commit.

Removed

_CURL_COMPAT and the per-version answers behind it (curlike.compat): curlike answers as curl 8.22, the latest stable release, and follows new releases. find_real_curl() and describe_curl() are unchanged.
curlike.shim (main, default_entry, find_real_curl): the dotagents net overlay's lib/curlshim is the launcher now.
Routing comes from syshttp 0.1.0.dev6's proxy.proxy_map() over proxylib 2 (now a direct dependency): the *_PROXY variables are read per scheme, NO_PROXY takes CIDR entries, HTTP_PROXY_AUTH goes to the machine's proxy (a *_PROXY variable's or the system's) and _PROXY_AUTH to _PROXY only. Of a list the first proxy that answers is used; a lone proxy is not probed. The exit-7 message is no proxy of the list is reachable: host:port, ....
The package stands alone. It was previously the dotagents net overlay's curl shim (bin/curl.py) and its fallback (httplib.cli, python -m httplib). The fallback is now curlike (curlike.main), and the shim is curlike.shim. It runs on syshttp, the overlay's httplib as a package of its own.
Settings follow syshttp's prefix (SYSHTTP_PREFIX, default SYSHTTP): NET_CURL → SYSHTTP_CURL, NET_CURL_COMPAT → SYSHTTP_CURL_COMPAT, NET_CURL_SSL_REVOKE → SYSHTTP_CURL_SSL_REVOKE, AGENTS_PYTHON → SYSHTTP_PYTHON, AGENTS_PYLIB → SYSHTTP_PYLIB. The proxy and hook variables are syshttp's (SYSHTTP_PROXY, SYSHTTP_HOOKS_<KEY>_CURL). With SYSHTTP_PREFIX=AGENTS, the AGENTS_* names are read instead.
The fallback's TLS is the system's by default (syshttp's sslcontext system context): the operating system verifies certificates on Python 3.10+, and a server that asks for a client certificate gets one from the OS store, chosen as a browser chooses. --cert turns the store off: the named certificate is the one sent. --crlfile keeps an OpenSSL context, because the OS verifier ignores CRL flags.
A malformed proxy setting is reported as curl: ignoring the configured proxy: … (was ignoring the agent proxy).
curl -V from the fallback says curl 0.0.0-curlike-shim … -- curlike fallback, and %{json}'s curl_version says (curlike fallback).

Fixed

-k, --proxy-insecure and --doh-insecure build a plain ssl context. The OS verifier (truststore) on macOS still evaluated the chain under them and failed with 60.
--local-port with a range moves to the next port when connect() finds the port in use. Windows reports a port whose last connection to the same address is in TIME_WAIT there, not at bind(), and the fallback used to give up with exit 7.
-L follows redirects in curlike's own loop, as curl does: 50 hops by default (--max-redirs -1 is unlimited, 0 refuses the first; exit 47 past the limit); a redirect to ftp:// is exit 1 instead of a direct connection outside the proxy and every connection option, and file:// or gopher:// are exit 1; a -H Host goes only to the first request's host; -D and -i show every hop's headers; -e 'URL;auto' sets each hop's Referer to the page it came from; --location-trusted implies -L.
With the cookie engine (-b FILE or -c), the cookies a redirect hop sets are sent to the next hop and filed under that hop's host, not the first URL's.
--digest and --anyauth answer a server's 401 once, with one implementation shared with the proxy's: a rejected password is sent once (two requests, not seven) and the 401 and its body are the response; a challenge curlike cannot answer (Negotiate, Bearer) is the response too, not exit 2; algorithm=SHA-256 works on Python 3.9; quotes in the user name or realm are escaped.
The netrc is read as curl reads it: quoted words with backslash escapes, comments, a machine name in any case, the first matching entry in file order, a login with no password (an empty one, not None), and the login a user@ URL names. -u USER asks for the password, as curl does.
-c no longer stores a cookie whose Domain the host that set it does not domain-match, or a public suffix, and a cookie with no Path gets the directory of the request (RFC 6265, through syshttp 0.1.0.dev14, now the floor).
The body is streamed, not held in memory (a 200 MB download peaks near 40 MB, not 470 MB, and the first byte reaches stdout at once): --max-filesize stops a body of unknown length at the cap, --compressed decodes piece by piece, and -T reads its file as it sends it.
Response limits are curl's: headers over 300 KB are exit 56, over 5000 headers exit 100, a conflicting or negative Content-Length or a NUL in a header exit 8. http.client's limits and socket.getaddrinfo are no longer replaced for the length of a transfer (concurrent calls cannot leave them changed).
A 206 answering -C must start at the resume offset or it is exit 33 (the file is left alone), and --remove-on-error removes a file the failed transfer overwrote; with -C it is exit 2, as in curl.
--retry covers a name that does not resolve and, under --retry-all-errors, a body cut short; every failed attempt prints its error line; the warning reads "Retrying in N seconds" (curl 8.21); a Retry-After of 0 is none, and one past --retry-max-time ends the retries with curl's warning.
--crlfile loads only the CRLs of its file: a CA certificate inside it is no longer trusted (a file with no CRL is exit 82). The contexts curlike builds no longer apply Python 3.13's strict X.509 checks, so a leaf with no Authority Key Identifier verifies as it does with curl and on older Pythons.
--proxy-pinnedpubkey is checked on a proxy connection reopened after a 407, and --pinnedpubkey, -0, --interface, --local-port and the -w timings apply over --unix-socket (it now goes through the one connector); a CONNECT carries a single Host line on Python 3.12+; an over-long SOCKS5 user name, password or host name is exit 97; an unreadable --pinnedpubkey file no longer fails an http:// transfer; --doh-url with an unusable --cacert fails the lookup (77) instead of falling back to the OS store.
-H sends a repeated name once per value, replaces only a default of the same name, and -H 'Name:' removes Host, Connection, Accept-Encoding and Content- Type too, as the header always claimed. Header values go out as UTF-8 (they were refused or sent as latin-1), and response header bytes are written back unchanged by -i, -D and -w, with the status line's HTTP version; -w @file is read and written as bytes. A URL with a space or control character is exit 3 (not 8) and nothing is sent; non-ASCII characters in a URL are sent percent- encoded as UTF-8 (a host as IDNA) instead of exit 2.
-J reads filename= only at a parameter's start and gives a Windows name as curl does (no drive-relative C:evil out of --output-dir, no device names); -O with no name in the URL writes the last segment or curl_response (exit 23 before); a -c jar that cannot be written no longer ends in a traceback; --etag-save is created before the transfer (26 when it cannot be); -b is a file when it has no = (a missing one holds nothing, - is stdin) and -c - writes the jar to stdout; -d @file also drops NUL and an unreadable data, -H or -w file is curl's three lines; --data-urlencode and --url-query write a space as +; -d x --json y joins with no &; -F takes @"file;name" and type=...;charset=..., and no longer types .json as application/json; -T sends no Content-Type and adds no file name to a URL with a query; -I -X POST sends POST; %{json} writes "" for the addresses of a transfer that never connected; the connect failure reads host:port.
Messages follow curl 8.21's words where they had drifted: a bad chunk names the byte (0x7a), a bad gzip stream drops zlib's prefix, a refused proxy connection says "over proxy", a missing --cacert, --crlfile, --netrc-file or --proxy-cacert file is wrapped and followed by the option hint, an unreadable -d, -H or -w file wraps its path, a -T file that is not there prints curl's three lines, an unknown --proto name prints no first line, a proxy with no scheme is shown as http://..., and the glob error words match.
curlike.caller_steering() and requested_url() read argv with curl 8.23's own table of the options that take a value: those the fallback never declared (22 long ones in curl 8.21: --proxy1.0, --hsts, --rate, ..., and -P) no longer leave their value as the URL, and --expand-NAME is read as --NAME. A --proxy1.0, --preproxy or -K/--config (alone or inside a short cluster such as -sK) counts as the caller naming a proxy. caller_steering no longer needs duho.
The proxy-syntax error (exit 5) hides the whole userinfo as *** (http://***@host), whatever the password holds (/, ?, #, @); it used to print a password containing /, or its tail after an @.
Ctrl-C ends curlike.main() with exit 130 and nothing printed, as with curl, instead of a Python traceback.

Security

Answering a 401 (--digest, --anyauth) for an https URL through an HTTP proxy re-sent the request in clear inside the tunnel: the password or Digest answer, cookies and every -H value were readable by the proxy and the network, and the request then failed with exit 52. With -p, the second CONNECT went to the proxy's own address. The answer now goes over a CONNECT to the origin and the origin's TLS, as the first attempt did.
-O and --remote-name-all percent-decoded the URL's last path segment, so ..%2F..%2Fname (and ..%5C..%5Cname, C%3A%5C... on Windows) wrote outside the current or --output-dir directory, overwriting what was there. The file is named after the segment as written, as curl does (. and .. resolved, a backslash separating); on Windows : becomes _ and a device name such as CON gets a _.
--compressed decodes in pieces of at most 16 KiB per layer: a 1.7 KB body with Content-Encoding: gzip, gzip held about 2 GB in memory. A response naming more than 5 content encodings (curl's limit) is exit 61.
--proxy-anyauth and --proxy-digest answer only a 407 that the proxy sent. An origin's 407, from a host reached directly (--noproxy, NO_PROXY, a redirect hop that bypasses the proxy) or inside a tunnel, is the response, and the proxy credential is never sent to it.
A proxy URL with a path, query or fragment (-x, _PROXY, HTTP_PROXY) is the proxy, as in curl, instead of exit 2 with a message that printed the password in full. A refused proxy setting never shows the password or the -U credential in its message.
A -b jar file is split into rows at a newline only (syshttp 0.1.0.dev19): a value holding a vertical tab, a form feed or another character that str.splitlines ends a line at could end its row early and start a row for another domain. A row with a control character in its name or value is skipped, as curl skips it.