pwn 0.5.656 → 0.5.658

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: ba4560cb04bae940cc23704a47b54f60a33a47587531101f340beb12897b12e9
-  data.tar.gz: 7635dbe7cb031343f1df74c430b22c1a9db3b2a23d3990812573e0785dd4aaa1
+  metadata.gz: 84155e448029c5d41ebd8dda5e4c00147a3283ebf8258c42f2fdf7cb3f159616
+  data.tar.gz: ceecb743a07ddf4d0a2f45434f1b7816e859754762d60f8932f9dd1f4a83d80a
 SHA512:
-  metadata.gz: e21ee1d6b4e66d692b6ad5896f1ec0f87426258854911d9dc7759d567dd3b472ab8dc99409ca33e7412afc7bab00d0174d56fa9cc979074b0192cc2bc43e0459
-  data.tar.gz: fdbed596d420d497b1874c1f5473c513327813b2625b2b9bedf7f21d5dbe51da26598c6e5a2082ebc7d2735f4f56e69a3f8feb9f99d2883899230c39b02e6b09
+  metadata.gz: 5cb1bf7948fa027253fa5e2aeb2204d917bf51ab4ebf2710073b8ca3c4fa07f53012a3e8f98e5db5f3dabda050b405119dadf78b7a5583476da35fd11af5c5c7
+  data.tar.gz: 07e6832ea6946949d3042f62b03930d5e6f882dd8f246449f314c994459b4f16ce4e32700ab4467b55d44e01f5ba35dbc3788f742d44a7fd79b36026b2d61ccf

data/lib/pwn/ai/agent/loop.rb

@@ -146,12 +146,13 @@ module PWN
           # is "iteration budget exhausted" ×N; more headroom only produces more
           # empty terminal failures for ORM/PRM/DPO.
           if budget_exhaustion_hot?
-            # P17 + P28 — finish-under-N while budget fingerprints dominate, but
-            # do NOT collapse multi-step remote work to 8 (that re-creates the
-            # polite handoff / early empty-final failure mode P28 fixed).
-            # Local/ollama stays harsh at 8; remote engines keep the W3 runway.
-            hot_cap = active_engine == :ollama ? 8 : 40
-            n = [n, hot_cap].min
+            # P17 — always 8 for ALL engines while budget fingerprints dominate.
+            # P28 remote runway (40) applies only to W3 overconf above; stacking
+            # a 40 hot-cap on top re-opens multi-dozen-iter thrash and is the
+            # 2026-08-04 regression (22 exhaust hits / day). Finish-under-8 is
+            # the sustained drop; incomplete_final? keeps multi-step autonomy
+            # inside that budget without polite handoffs.
+            n = [n, 8].min
           end
           n
         rescue StandardError
@@ -634,7 +635,9 @@ module PWN
             # nothing the user (or ORM) can use.
             # P17 deepen — when budget_hot, force text-only on the LAST TWO
             # iters so a final tool_calls batch cannot burn the terminal slot.
-            text_only_iters = budget_exhaustion_hot? ? 2 : 1
+            # P17 deepen³ — under hot, force text-only on last THREE of the
+            # 8-iter cap so a late tool binge cannot burn every salvage slot.
+            text_only_iters = budget_exhaustion_hot? ? 3 : 1
             last_iter = (i >= max_iters - text_only_iters)
             if last_iter
               tag = i >= max_iters - 1 ? 'FINAL ITERATION' : 'PENULTIMATE — wrap up'
@@ -830,6 +833,7 @@ module PWN
 
               P28 autonomy: incomplete-final detector refuses mid-goal handoffs;
               W3 overconf max_iters_cap is 40 on remote engines (8 on ollama).
+              P17 budget-hot always caps max_iters to 8 for ALL engines (not only ollama).
 
               #{self}.authors
           USAGE

data/lib/pwn/ai/agent/tools/shell.rb

@@ -6,6 +6,13 @@ require 'pwn/ai/agent/registry'
 
 # Run a shell command on the pwn host. Lifted from the bash branch of the
 # legacy :pwn_ai_hook (repl.rb).
+#
+# Timeout path must NOT leave Open3's pipe-reader threads racing a closed
+# pipe — that surfaces as noisy:
+#   #<Thread:0x... open3.rb:664 run> terminated with exception
+#   (report_on_exception is true): IOError stream closed in another thread
+# Fix: own the popen3 lifecycle, process-group kill on timeout, and silence
+# report_on_exception on the reader threads before closing pipes.
 PWN::AI::Agent::Registry.register(
   name: 'shell',
   toolset: 'terminal',
@@ -29,16 +36,84 @@ PWN::AI::Agent::Registry.register(
     timeout = (args[:timeout] || 120).to_i
     raise ArgumentError, 'command is required' if cmd.strip.empty?
 
-    stdout = stderr = ''
-    status = nil
-    begin
-      Timeout.timeout(timeout) do
-        stdout, stderr, status = Open3.capture3(cmd)
+    stdout = +''
+    stderr = +''
+    exitstatus = nil
+    timed_out = false
+
+    # pgroup: true => child is session/process-group leader; kill(-pid)
+    # reaps the whole tree (shell + grandchildren) on timeout.
+    Open3.popen3(cmd, pgroup: true) do |stdin, out_io, err_io, wait_thr|
+      stdin.close
+      pid = wait_thr.pid
+
+      out_reader = Thread.new { out_io.read.to_s }
+      err_reader = Thread.new { err_io.read.to_s }
+      # Prevent Ruby from dumping IOError to $stderr when we close pipes
+      # out from under these readers on the timeout path.
+      out_reader.report_on_exception = false
+      err_reader.report_on_exception = false
+
+      begin
+        Timeout.timeout(timeout) do
+          # Wait for the child; readers drain pipes concurrently.
+          wait_thr.value
+        end
+      rescue Timeout::Error
+        timed_out = true
+        begin
+          Process.kill('TERM', -pid)
+        rescue Errno::ESRCH, Errno::EPERM
+          nil
+        end
+        # Brief grace, then hard kill the process group.
+        sleep 0.2
+        begin
+          Process.kill('KILL', -pid)
+        rescue Errno::ESRCH, Errno::EPERM
+          nil
+        end
+        begin
+          wait_thr.value
+        rescue StandardError
+          nil
+        end
+      end
+
+      # Ensure pipes are closed so readers unblock even if the child
+      # ignored signals or left descendants holding fds.
+      begin
+        out_io.close unless out_io.closed?
+      rescue StandardError
+        nil
+      end
+      begin
+        err_io.close unless err_io.closed?
+      rescue StandardError
+        nil
+      end
+
+      # Join readers; swallow the IOError that arrives when the pipe was
+      # closed mid-read (timeout path). report_on_exception is already off.
+      begin
+        stdout = out_reader.value.to_s
+      rescue StandardError
+        stdout = +''
+      end
+      begin
+        stderr = err_reader.value.to_s
+      rescue StandardError
+        stderr = +''
+      end
+
+      unless timed_out
+        status = wait_thr.value
+        exitstatus = status&.exitstatus
       end
-    rescue Timeout::Error
-      return { stdout: stdout, stderr: stderr, exit: nil, error: "timeout after #{timeout}s" }
     end
 
-    { stdout: stdout, stderr: stderr, exit: status&.exitstatus }
+    return { stdout: stdout, stderr: stderr, exit: nil, error: "timeout after #{timeout}s" } if timed_out
+
+    { stdout: stdout, stderr: stderr, exit: exitstatus }
   }
 )

data/lib/pwn/banner/white_rabbit.rb

@@ -15,7 +15,7 @@ module PWN
             / \     / \
            {   }   {   }
            {   {   }   }
-            \   \ /   /
+            \   \ /   /   R.I.P. Houdini
              \   Y   /
              .-"`"`"-.
            ,`         `.

data/lib/pwn/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module PWN
-  VERSION = '0.5.656'
+  VERSION = '0.5.658'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: pwn
 version: !ruby/object:Gem::Version
-  version: 0.5.656
+  version: 0.5.658
 platform: ruby
 authors:
 - 0day Inc.