fix(windows): give a starting kernel long enough to publish its ports

packages/coding-agent/CHANGELOG.md

@@ -12,6 +12,7 @@
 - Fixed Homebrew installs attempting to self-update their versioned Cellar keg instead of directing users to `brew upgrade prime-agent` ([#844](https://github.com/PrimeIntellect-ai/prime-agent/issues/844))
 - Fixed the agents view collapsing expanded subagent lists when returning from an opened agent ([ENG-5105](https://linear.app/primeintellect/issue/ENG-5105/keep-the-agents-view-state-persistent)).
 - Kept the subagent summary row visible and selectable while its list is expanded in the agents view, so pressing enter on it collapses the list again ([ENG-5105](https://linear.app/primeintellect/issue/ENG-5105/keep-the-agents-view-state-persistent)).
+- Fixed a slow IPython kernel start on Windows being reported as a broken kernel. The wait for resolved ports is 15s there, unchanged elsewhere.
 
 ## [0.7.1] - 2026-08-07
 

packages/coding-agent/src/core/kernel/index.ts

@@ -24,7 +24,16 @@ import {
 
 const DELIM = Buffer.from("<IDS|MSG>");
 const PROTOCOL_VERSION = "5.3";
-const PORTS_RESOLVE_TIMEOUT_MS = 5000;
+/**
+ * How long to wait for a starting kernel to publish its resolved ports.
+ *
+ * Windows pays a much slower interpreter start than Linux or macOS, and several
+ * kernels starting at once — a sharded test run, or a daemon bringing up
+ * sessions in parallel — pushes it past five seconds. The wait loop below exits
+ * early when the kernel dies, so a longer budget never delays a real failure;
+ * it only stops a slow start being reported as a broken one.
+ */
+const PORTS_RESOLVE_TIMEOUT_MS = process.platform === "win32" ? 15_000 : 5000;
 const READY_TIMEOUT_MS = 5000;
 // Loopback PUB/SUB subscription propagation is usually sub-ms, but keep a small guard before first execute.
 const IOPUB_SUBSCRIBE_DELAY_MS = 50;