@@ -5,6 +5,8 @@ All notable changes to CodeSetu will be documented here.
 Format: [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
 Versioning: [Semantic Versioning](https://semver.org/)
 
+For human-friendly, themed highlights see [Release Notes](docs/RELEASE_NOTES.md).
+
 ---
 
 ## [Unreleased]
 
@@ -22,6 +22,7 @@ This repository is a pnpm + Gradle monorepo organized as:
 - `skills/` — AI skill manifests (`SKILL.md` per skill) loaded by hosts at activation
 - `plugins/` — first-party plugins built on `@codesetu/plugin-sdk`
 - `docs/ARCHITECTURE.md` — full architecture, layout, and deployment notes
+- `docs/RELEASE_NOTES.md` — human-friendly highlights of what's new each release
 
 ## Agent Mode
 
 
@@ -0,0 +1,74 @@
+# CodeSetu Release Notes
+
+Human-friendly highlights of what's new, grouped by theme. For the complete,
+technical list of every change, see the [CHANGELOG](../CHANGELOG.md) — that
+remains the source of truth.
+
+Everything here is **bring-your-own-model** and works fully local / air-gapped:
+point CodeSetu at Sarvam, Ollama, vLLM, OpenRouter, or any OpenAI-compatible
+endpoint.
+
+---
+
+## 0.4.14 — Codebase-aware, more transparent, runs anywhere
+
+Shipped for **both** VS Code and JetBrains (IntelliJ / PyCharm / WebStorm), now
+aligned at the same version.
+
+### 🧠 `@workspace` — semantic codebase search
+
+Index your repo once and CodeSetu retrieves code by *meaning*, not just text.
+
+- Type `@workspace <question>` in chat, or let Agent Mode call the
+  `search_workspace` tool.
+- **Auto-builds on first use** and **auto re-indexes a few seconds after you
+  save** (incremental) — retrieval stays fresh with no manual step.
+- **Always-on retrieval** option: pull relevant code into *every* turn, not only
+  when you type `@workspace`.
+- Runs against any OpenAI-compatible embeddings endpoint (defaults to Ollama
+  `nomic-embed-text`). The index persists under `.codesetu/`.
+
+### 🤖 Agent Mode — now the default
+
+- Reads, edits, and runs commands in a tool-calling loop, with your approval and
+  a one-click **revert** of a turn's edits.
+- **Works with models that lack native tool-calling** (e.g. Gemma and many small
+  local models) via a prompted tool-calling fallback.
+- Read-only tools auto-run; file edits and shell commands ask first.
+
+### 🔍 "Context & activity" panel — see what CodeSetu is doing
+
+Every turn shows the **provider · model · endpoint**, whether Agent Mode is on,
+the **tools available**, and the **@workspace** outcome (retrieved files, or the
+reason it found nothing). A small **version badge** by the model picker makes it
+obvious which build is loaded.
+
+### ✂️ Sharper editing & context
+
+- **`/edit`** — rewrite a selection or file from an instruction, preview it as a
+  diff, and **accept/reject individual hunks**. Available as a command and as a
+  `/edit` slash in chat.
+- **`@folder`** pinning — pin a whole folder as context, not just single files.
+- **Multi-session chat history** — a switcher to jump between past
+  conversations; "New chat" starts a fresh one and keeps the old.
+
+### 🐳 On-prem Docker installer
+
+A turnkey self-hosted stack (`deploy/docker/`): one OpenAI-compatible server that
+serves **both** the chat/agent model and the `@workspace` embedding model, fully
+inside your network. Includes an **air-gapped** install path (save image + model
+volume, restore offline), a health check, a GPU toggle, and a documented vLLM
+alternative. The IDE extensions are unchanged — they just point at the endpoint.
+
+### 💡 Tips
+
+- For agent / coding tasks, prefer a tool-calling coder model such as
+  `qwen2.5-coder`. Reasoning-only models (which emit a long "thinking" stream
+  before the answer) can return an empty reply inside the tool loop.
+- To use a large context window with local Ollama, raise it on the server, e.g.
+  `OLLAMA_CONTEXT_LENGTH=131072 ollama serve` — the model's context window is a
+  server setting, separate from the per-reply output limit.
+
+---
+
+<!-- Add new releases above this line -->