References: GHSA-35p6-xmwp-9g52
Signed-off-by: Matteo Collina <hello@matteocollina.com>
Co-authored-by: Ulises Gascon <ulisesgascongonzalez@gmail.com>
Signed-off-by: Matteo Collina <hello@matteocollina.com>

* fix: keep idle validation on native timers

Signed-off-by: Matteo Collina <hello@matteocollina.com>

* test: wait for poisoned idle socket disconnect

Signed-off-by: Matteo Collina <hello@matteocollina.com>

---------

Signed-off-by: Matteo Collina <hello@matteocollina.com>
(cherry picked from commit c9fbe9d)

Signed-off-by: Matteo Collina <hello@matteocollina.com>
(cherry picked from commit 9290322)

Signed-off-by: Matteo Collina <hello@matteocollina.com>
(cherry picked from commit 5655ea4)

Signed-off-by: Matteo Collina <hello@matteocollina.com>
(cherry picked from commit cb105d7)

… routing (#5041)

The Socks5ProxyAgent stored a single Pool keyed implicitly to the first
origin it saw and reused it for every subsequent request. When a second
request targeted a different origin, it was dispatched through the
existing pool whose connect callback tunnelled to the original origin,
causing the request to reach the wrong host (and potentially leaking
headers/credentials intended for origin B to origin A).

Track pools in a Map keyed by origin so each origin gets its own pool
and SOCKS5 tunnel.

Signed-off-by: Nikita Skovoroda <chalkerx@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Nikita Skovoroda <chalkerx@gmail.com>
(cherry picked from commit a516f87)

(cherry picked from commit 1dec881)
Signed-off-by: Matteo Collina <hello@matteocollina.com>

…5099)

Assisted-by: openai:gpt-5.4

Signed-off-by: Kamat, Trivikram <16024985+trivikr@users.noreply.github.com>
(cherry picked from commit d76f932)

* fix: use configured connector in Socks5ProxyAgent

Assisted-by: openai:gpt-5.5
Signed-off-by: Kamat, Trivikram <16024985+trivikr@users.noreply.github.com>

* test: use t.after in Socks5ProxyAgent custom connector test

---------

Signed-off-by: Kamat, Trivikram <16024985+trivikr@users.noreply.github.com>
(cherry picked from commit 1a45226)

Assited-by: openai:gpt-5.5

Signed-off-by: Kamat, Trivikram <16024985+trivikr@users.noreply.github.com>
(cherry picked from commit e2bc508)

ProxyAgent silently dropped opts.requestTls when the configured proxy
URI was socks5:// (or socks://). Socks5ProxyAgent also lacked any
equivalent of requestTls, so even constructed directly it had no way to
configure target HTTPS TLS settings (ca, cert, key, rejectUnauthorized,
servername).

The inner connect callback called tls.connect with
...connectOpts.tls || {} but undici's internal connectOpts never
populates a .tls field, so the spread was a dead expression.

This change:
- Forwards opts.requestTls from ProxyAgent to Socks5ProxyAgent.
- Adds a requestTls constructor option on Socks5ProxyAgent stored in
  the kRequestTls symbol.
- Applies it on the inner tls.connect for the target HTTPS upgrade,
  with socket overriding (must point at the SOCKS5 tunnel) and
  servername defaulting to targetHost but allowing user override
  to match the HTTP-proxy path semantics.
- Documents the new option in Socks5ProxyAgent.md.
- Adds two regression tests covering both the direct Socks5ProxyAgent
  path and the ProxyAgent forwarding path.

Refs GHSA-vmh5-mc38-953g

(cherry picked from commit 42d4955)
Signed-off-by: Matteo Collina <hello@matteocollina.com>

Introduce the `maxFragments` option to limit the  number of fragments in
a message. Without this limit, a remote peer could exploit structural
wrapper overhead by sending many tiny fragments, consuming far more
memory than the payload itself and crashing the process due to OOM. When
the limit is hit, the connection is closed with close code 1008.

Signed-off-by: Luigi Pinca <luigipinca@gmail.com>
(cherry picked from commit 32dbf0b)

* feat: add configurable maxPayloadSize for WebSocket (#4955)

(cherry picked from commit bd91f86)
Signed-off-by: Matteo Collina <hello@matteocollina.com>

* test: fix flaky permessage-deflate limit timeout (#5229)

(cherry picked from commit 9d82667)
Signed-off-by: Matteo Collina <hello@matteocollina.com>

* fix(websocket): enforce max payload size across fragments

Account for previously received fragment bytes when checking WebSocket payload size limits, so fragmented messages cannot exceed maxPayloadSize by splitting the payload across frames.

Add coverage for cumulative fragmented payload size enforcement.

Co-authored-by: Matthew Aitken <maitken033380023@gmail.com>
(cherry picked from commit b4c287b)
Signed-off-by: Matteo Collina <hello@matteocollina.com>

* websocket: handle empty fragments and stream limits

Treat zero-byte frames as real fragments so fragmented messages can start
with an empty frame and empty continuations still count toward
maxFragments.

Pass dispatcher WebSocket limits through to WebSocketStream's parser, add
regression coverage for WebSocket and WebSocketStream fragment limits, make
the fragment close tests wait for both endpoints, and fix the Client docs
typo for maxFragments.

Co-authored-by: Ulises Gascon <ulisesgascongonzalez@gmail.com>
(cherry picked from commit c5ed787)
Signed-off-by: Matteo Collina <hello@matteocollina.com>

---------

Signed-off-by: Matteo Collina <hello@matteocollina.com>
Co-authored-by: Matthew Aitken <maitken033380023@gmail.com>
Co-authored-by: Luigi Pinca <luigipinca@gmail.com>
Co-authored-by: Ulises Gascon <ulisesgascongonzalez@gmail.com>

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>