diff --git a/CHANGELOG.md b/CHANGELOG.md
new file mode 100644
index 0000000..94165c4
--- /dev/null
+++ b/CHANGELOG.md
@@ -0,0 +1,82 @@
+# Changelog
+
+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).
+
+## [Unreleased]
+
+First stable release of codesearch — a Rust-based hybrid (vector + BM25 + AST)
+code search MCP server, optimised for AI coding agents working across many
+repositories.
+
+### Added
+
+- **Multi-repository serve mode** (`codesearch serve`): a long-running HTTP/SSE
+  process that holds many indexed repositories warm at the same time, with
+  per-project routing via `project=…`, group routing via `group=…`, and
+  cross-repository search using RRF fusion across project boundaries.
+- **Stdio proxy with auto-reconnect**: `codesearch mcp` (stdio mode) detects a
+  running `serve` process and proxies tool calls to it. The proxy now performs
+  client-side retries with a forced reconnect when it sees a transport-level
+  failure (broken TCP keep-alive, stale session 404, server restart, laptop
+  suspend) so MCP clients like Claude Desktop self-heal transparently. After a
+  serve restart the first call returns a clear "reconnecting" message and the
+  next call succeeds.
+- **MCP tool surface optimised for agents** to reduce grep-fallback behaviour:
+  - `search` (semantic / hybrid / lexical / pure-literal regex modes)
+  - `find` (definition / usages / imports / dependents)
+  - `explore` (file outline / similar chunks)
+  - `get_chunk` for cheap follow-up reads of a specific code chunk
+  - `status` (index / projects)
+- **Tree-sitter AST-aware chunking** for 9 languages: Rust, Python, JavaScript,
+  TypeScript, C, C++, C#, Go, Java.
+- **Persistent embedding cache** keyed on SHA-256 of chunk content, surviving
+  `--force` rebuilds and per-file re-indexes.
+- **Git worktree support**: when `.git` is a worktree marker file (not a
+  directory), the project root is correctly resolved to the worktree itself.
+- **Long UNC-path support** on Windows for repositories under `\\?\C:\…` paths.
+- **Repository groups** for cross-repo search across user-defined sets of
+  projects (e.g. all *.Aprimo* repos).
+
+### Changed
+
+- **Search quality**: re-tuned RRF fusion of the vector / BM25 / exact-identifier
+  signals so common tool names and exact strings are no longer drowned out by
+  semantic neighbours, reducing the rate at which agents fall back to external
+  grep.
+- **Idle eviction**: only refreshes a project's "last accessed" timestamp on a
+  direct query against that project, not on fan-out queries that touch the
+  index merely because they routed through the same group.
+- **TUI CPU%**: now normalised by core count.
+
+### Fixed
+
+- **Security**: validate `CODESEARCH_CONFIG` environment variable against a path
+  traversal pattern (CodeQL finding). Config path is now rejected if it contains
+  `..` segments, preventing a directory traversal via env var.
+- **Issue #30** ([LMDB resize crash on large repositories](https://github.com/flupkede/codesearch/issues/30)):
+  When the database grew beyond its initial allocation (`MDB_MAP_FULL`), the
+  resize failed with `"an environment is already opened with different options"`.
+  The fix closes and reopens the LMDB environment around the resize, allowing
+  codesearch to index large repositories (tested: 4400+ files, 89 MB) without
+  crashing.
+- File-change tracking and reaper visibility in `serve` mode.
+
+### Removed
+
+- Server-side transparent MCP session-reconnect middleware: replaced by the
+  client-side retry in the stdio proxy. The middleware could not reach
+  non-compliant remote MCP clients (their HTTP pool gives up at the TCP layer
+  before the request hits the server) and added a session-counter leak.
+
+### Known limitations
+
+- Remote MCP clients that do not handle 404 "Session not found" per the MCP
+  spec (e.g. OpenCode 1.14.x at the time of writing) need to be restarted after
+  a `codesearch serve` restart.
+- `codesearch serve` keeps one writer per database (LMDB invariant). Concurrent
+  reindex from a second process is rejected.
+
+[Unreleased]: https://github.com/flupkede/codesearch/compare/master...develop
diff --git a/Cargo.lock b/Cargo.lock
index 3ee4cdc..52d95a5 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -634,7 +634,7 @@ dependencies = [
 
 [[package]]
 name = "codesearch"
-version = "1.0.72"
+version = "1.0.74"
 dependencies = [
  "anyhow",
  "arroy",
diff --git a/Cargo.toml b/Cargo.toml
index 8867500..28076e2 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "codesearch"
-version = "1.0.72"
+version = "1.0.74"
 edition = "2021"
 authors = ["codesearch contributors"]
 license = "Apache-2.0"
diff --git a/src/serve/mod.rs b/src/serve/mod.rs
index d3dd6c4..e327c8e 100644
--- a/src/serve/mod.rs
+++ b/src/serve/mod.rs
@@ -1731,10 +1731,12 @@ pub async fn run_serve(
                 .map_err(std::io::Error::other)
         };
 
-    // Build session manager with extended keep_alive (default is 5 min which kills
-    // idle MCP sessions too aggressively). 30 minutes matches our repo idle eviction.
+    // Build session manager without keep_alive timeout. The default rmcp timeout
+    // (5 min) kills idle sessions too aggressively for a local long-running serve.
+    // We run single-user local, so abandoned sessions cost nothing — let TCP
+    // liveness determine when a session is truly dead.
     let mut session_manager = LocalSessionManager::default();
-    session_manager.session_config.keep_alive = Some(std::time::Duration::from_secs(30 * 60));
+    session_manager.session_config.keep_alive = None;
     let session_manager = Arc::new(session_manager);
     let config = StreamableHttpServerConfig::default();