yamine 0.21.0 → 0.21.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 73d2fe319b2c8af2c4bf821285dc76af42705e532761a8b2351377696928a114
-  data.tar.gz: 5a92b4760142642deed0a5a0131536094ac553b79759924daa7cc2157dec90a6
+  metadata.gz: 8e920f5f6d29a92e0a6e81ca671bccf62e0a264ec453f9b5bc4d42d4e56980d3
+  data.tar.gz: cc06342930ee7c882cf30fd2818f1a60495c8a59e44ec6779cc666f569493ac0
 SHA512:
-  metadata.gz: 859e65d51d8e417cb1c747cf2d95a8336d4266fd57347d611ca3744baf4538dde86190fc5bc46023cbf28a0c12faceb7d2e3f84895afd62617e2c78e7fd1fc46
-  data.tar.gz: 8fc48c3d08d4ef3dfb5f579f41583ce3f315c90e896289279b072a5f9e6d66013eaa2b02f66eeb8ce3e983d2d4664287aa3e868e33606bc4f687f7975cd9c25a
+  metadata.gz: 4164dabf9bd5537023ad59d426a958a5e284a90f0226cc83e0533cf635edf0c65acc16dd311d58a88df83d7abdcdfb92034703ecda06d89af08bb5ec96c73635
+  data.tar.gz: 1e3af2b55e54885fe8aeed98856c4371eab280ec981cd2644ebb75aba0206752ae44307c0252f51188a778426efbb882ef2285c63f142def39f4e876f77302d6

data/CHANGELOG.md

@@ -2,6 +2,27 @@
 
 ## [Unreleased]
 
+## [0.21.1] — 2026-09-29
+
+### Fixed
+
+- **Claiming a per-worktree database no longer raises `NoMethodError`
+  on Ruby 3.2/3.3.** The claim stamps the state file with
+  `Time#iso8601`, but `lib/yamine/database.rb` never required the
+  stdlib `time` library, and on 3.2/3.3 nothing else loads it
+  implicitly — so every claim raised (34 unit errors on those rubies
+  in CI). 0.21.0 is affected. One `require "time"`, alongside the
+  stdlibs that file already pulls in.
+- **A closed WebSocket upgrade no longer leaks a thread on Linux.**
+  `pipe_both` tore each end down with `close` from the sibling
+  thread, and a close from another thread does not interrupt that
+  copy's blocking read on Linux (it does on macOS) — so every closed
+  upgrade parked a thread forever, `handle` never unwound, and they
+  accumulated in the root daemon. Teardown is now symmetric
+  `shutdown(2)`, which wakes the parked read with EOF; both sockets
+  still close exactly once, and no timeout was added to the
+  unbounded upgrade path.
+
 ## [0.21.0] — 2026-09-28
 
 ### Security

data/README.md

@@ -508,10 +508,11 @@ bundle exec rake test
   self-signed, and there is no buffering or rate limiting.
 
 
-The `yamine-apps` fixture fleet (sibling checkout) exercises
+The `test/fixtures/apps/` fixture fleet exercises
 detection, inference, and boot across Rails variants, Roda, Sinatra,
-bare Rack, Jekyll, compound Procfiles, and a monorepo. CI runs the
-fixture sweep automatically.
+bare Rack, Jekyll, compound Procfiles, and a monorepo. It ships with
+the repo and runs as part of the unit suite across the Ruby matrix —
+no sibling checkout, no separate CI job.
 
 ## License
 

data/lib/yamine/database.rb

@@ -4,6 +4,7 @@ require "digest"
 require "fileutils"
 require "json"
 require "open3"
+require "time"
 
 module Yamine
   # Per-worktree databases. Every worktree directory gets its own database

data/lib/yamine/proxy.rb

@@ -566,25 +566,51 @@ module Yamine
     # the other. Deliberately outside the idle bound: an idle websocket
     # is legitimate traffic (silence is the normal state), and framing
     # no longer applies once the connection is hijacked.
+    #
+    # Whichever direction reaches EOF first tears down BOTH ends, and it
+    # has to do that with shutdown before close. The sibling copy is
+    # parked in a blocking read on the socket being torn down, and a
+    # close from another thread does not interrupt that read on linux (it
+    # does on macOS) — so close-only left the parked copy alive forever
+    # and `t1.join` never returned: one leaked thread per closed upgrade,
+    # plus a `handle` that never unwound, in the root daemon. shutdown(2)
+    # changes the socket's kernel state instead of dropping the fd, so
+    # the parked read comes back with EOF at once — deterministic, and
+    # with no timeout added to a path that must stay unbounded.
     def pipe_both(client, backend)
       t1 = Thread.new do
         IO.copy_stream(client, backend)
       rescue IOError, SystemCallError
         nil
       ensure
-        backend.close rescue nil
+        teardown_half(backend)
       end
       t2 = Thread.new do
         IO.copy_stream(backend, client)
       rescue IOError, SystemCallError
         nil
       ensure
-        client.close rescue nil
+        teardown_half(client)
       end
       t1.join
       t2.join
     end
 
+    # End one end of a hijacked connection: unblock the sibling copy
+    # first, then close. Runs from an ensure, so it swallows everything
+    # — an exception escaping here would unwind into the acceptor.
+    def teardown_half(sock)
+      # An SSLSocket (the TLS client's socket) has no #shutdown; its
+      # kernel socket is one level down, and that is the fd the parked
+      # read is sitting on.
+      io = sock.respond_to?(:shutdown) ? sock : sock.to_io
+      io.shutdown(Socket::SHUT_RDWR)
+    rescue StandardError
+      nil
+    ensure
+      sock.close rescue nil
+    end
+
     # DNS-rebinding boundary: the proxy binds loopback, so any website
     # that tricks a browser into requesting 127.0.0.1 reaches us. A
     # foreign Host (outside our TLDs) gets a bare 404 that names nothing —

data/lib/yamine/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Yamine
-  VERSION = "0.21.0"
+  VERSION = "0.21.1"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: yamine
 version: !ruby/object:Gem::Version
-  version: 0.21.0
+  version: 0.21.1
 platform: ruby
 authors:
 - Kaka Ruto