diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7086bd2..9a27431 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,26 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.2.4] - 2026-06-18
+
+### Fixed
+
+- **A late first transcript flush under MCP/tool load no longer surfaces a
+  delivered turn as lost.** When MCP servers and tool schemas load under
+  contention, claude defers a fresh session's first `<id>.jsonl` write by 3–9 s —
+  past the fixed anchor window. `send()` then returned the `DELIVERY_UNCONFIRMED`
+  sentinel, and a reader that saw `wait()→completed` still read an empty
+  `messagesSince("0")` / `turnComplete === false`, even though the turn had
+  landed. claudemux now tolerates the late flush at both seams against one shared
+  budget: the write-side cursor anchor keeps polling while the transcript file is
+  addressable-but-absent, and a read-side settle waits the first records out
+  once. A present / fast-flush session is untouched (the anchor keeps its exact
+  window; the settle returns on the first poll). The settle is bounded on both
+  success and exhaustion, so a never-flushing session (a crashed/slow resume) is
+  never re-stalled. Verified live (claude 2.1.181): 1/20 → 0/20 sentinels under
+  20 concurrent heavy tool-MCP sessions. See
+  [ADR 0009](docs/decisions/0009-tolerate-late-first-transcript-flush.md).
+
 ## [0.2.3] - 2026-06-09
 
 ### Fixed
diff --git a/package.json b/package.json
index 08b5bca..3f683b8 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@wastedcode/claudemux",
-  "version": "0.2.3",
+  "version": "0.2.4",
   "publishConfig": {
     "access": "public"
   },