diff --git a/.plugin/marketplace.json b/.plugin/marketplace.json
index 14acf5a..71e6f94 100644
--- a/.plugin/marketplace.json
+++ b/.plugin/marketplace.json
@@ -226,6 +226,19 @@
                 "cloud-native"
             ]
         },
+        {
+            "name": "adapt-yourself",
+            "source": "./adapt-yourself",
+            "description": "Decide how to make OpenHands adapt/improve itself persistently: skills, AGENTS.md, hooks, plugins, MCP servers, or code changes in OpenHands-CLI and software-agent-sdk.",
+            "category": "development",
+            "keywords": [
+                "openhands",
+                "customization",
+                "skills",
+                "hooks",
+                "plugins"
+            ]
+        },
         {
             "name": "notion",
             "source": "./notion",
diff --git a/skills/adapt-yourself/README.md b/skills/adapt-yourself/README.md
new file mode 100644
index 0000000..3996930
--- /dev/null
+++ b/skills/adapt-yourself/README.md
@@ -0,0 +1,8 @@
+# adapt-yourself
+
+This skill is a playbook for “self-modification” requests:
+
+- deciding whether the user actually needs code changes, vs.
+- adding a repo/user skill, always-on repo instructions (`AGENTS.md`), hooks (`.openhands/hooks.json`), plugins, or MCP tooling.
+
+The canonical instructions live in [`SKILL.md`](SKILL.md).
diff --git a/skills/adapt-yourself/SKILL.md b/skills/adapt-yourself/SKILL.md
new file mode 100644
index 0000000..0d6a2b3
--- /dev/null
+++ b/skills/adapt-yourself/SKILL.md
@@ -0,0 +1,152 @@
+---
+name: adapt-yourself
+description: Decide how to make OpenHands adapt/improve itself persistently (after restart) by choosing between repo/user skills, always-on instructions, hooks, plugins, MCP servers, or code changes in OpenHands-CLI and software-agent-sdk. Includes uv + uv tool workflows and restart/reload behavior.
+triggers:
+- modify yourself
+- change your behavior
+- self-modify
+- adapt yourself
+- self modification
+- hack openhands
+- .agents/skills
+- hooks.json
+- plugin.json
+---
+
+# Modify OpenHands (self-modification playbook)
+
+When a user says things like “modify yourself to do X” or “remember this next time”, **don’t jump to editing Python code**.
+OpenHands has multiple extensibility layers; choose the *smallest*, *safest*, and *most persistent* mechanism that matches the user’s intent.
+
+This skill is about **persistent** changes that should still apply after the user reloads or restarts the CLI / app. If the user wants behavior only for the current conversation, you don't need the mechanisms in this skill—just follow their instructions directly in this session.
+
+## 0) First: clarify what kind of change the user wants
+
+Clarify the **scope** and **type** of the change (skip questions the user already answered):
+
+1. **Persistence scope**: Should this change apply…
+   - for this *repository/workspace*?
+   - for *all repos on this machine* (user-global)?
+   - for *everyone* (contribution upstream / extensions registry)?
+2. **Change type**: Is it primarily…
+   - instructions/workflow (“when I ask for X, do Y”)?
+   - a safety/policy constraint (“never run rm -rf”, “ask before network calls”)?
+   - a new capability/tool integration (GitHub, Jira, internal APIs, etc.)?
+   - a bug fix / UI change in the CLI or SDK?
+3. **Where do they run OpenHands?**
+   - `uv tool install openhands` (tool environment)
+   - from a local clone (developer workflow)
+   - via a remote runtime / server
+
+Then follow the decision tree below.
+
+## 1) Prefer non-code extensibility (often enough)
+
+### 1.1 Write a skill (AgentSkills / progressive disclosure)
+
+Use when the user wants repeatable expertise/workflows, e.g.:
+- “When I say ‘release notes’, use this structure”
+- “Here’s how we do migrations in our company”
+- “Read the Daytona docs and learn how to use it”
+
+**Where to put it (in priority order):**
+
+1. **Project-local (recommended):** `<repo>/.agents/skills/<skill-name>/SKILL.md`
+2. **User-global:** `~/.agents/skills/<skill-name>/SKILL.md`
+
+Notes:
+- **Project skills override user + public skills** (same `name`). This is the simplest way to “patch” a public skill locally.
+- Add `triggers:` only when you really want automatic activation; prefer **distinctive phrases** to avoid injecting the skill on unrelated messages.
+
+See: [Templates](references/TEMPLATES.md#agentskillsskillmd-template).
+
+### 1.2 Add always-on repo instructions (AGENTS.md)
+
+Use when the user wants **always-applied** guidance for a repository (style, commands, guardrails, conventions):
+
+- Create or edit `<repo>/AGENTS.md` (or model-specific variants like `CLAUDE.md`, `GEMINI.md` if your environment supports them).
+- Put the repo rules there.
+
+This gets loaded at conversation start as “always-on context” (not progressive disclosure).
+
+### 1.3 Add hooks (policy/automation) via `.openhands/hooks.json`
+
+Use when the user wants **enforcement** or **automation** around tool usage, e.g.:
+- “Before any git commit, run tests.”
+- “Block any command that tries to access prod without approval.”
+- “Reject tool calls that write outside the repo.”
+
+Hooks are executed on lifecycle events (like `pre_tool_use`) and can **allow / block / modify** actions.
+
+See: [Hook templates](references/TEMPLATES.md#hookshooksjson-template) and [Hook architecture notes](references/ARCHITECTURE.md#hooks).
+
+### 1.4 Add tools via MCP (`.mcp.json`)
+
+Recommend MCP **only when there is a concrete MCP server to add**:
+
+- The user explicitly asks to “add an MCP server”, **or**
+- The user asks for a capability and you can point to an existing MCP server for it (e.g., from that project’s docs / an official MCP server repo).
+
+If there isn’t an MCP server available for what they want, don’t force MCP:
+- start with a **skill** (instructions/workflow), or
+- if they truly need a new tool, the real work item is to **build or install an MCP server** (or implement a native tool/code change).
+
+MCP server configuration location **depends on the runtime**:
+
+- **OpenHands-CLI** persists MCP servers in `~/.openhands/mcp.json` (or `$OPENHANDS_PERSISTENCE_DIR/mcp.json`).
+- **Software Agent SDK skills/plugins** use a **per-skill/per-plugin** `.mcp.json` file located inside the skill/plugin directory.
+
+See: [MCP template](references/TEMPLATES.md#mcp-template).
+
+### 1.5 Package it as a plugin (shareable bundle)
+
+Use when the change should be **portable** across repos or shared with others, and it includes more than just instructions:
+- skills + hooks + MCP config together
+- optional slash commands and/or specialized agents
+
+Plugins follow the Claude Code-compatible structure, with `.plugin/plugin.json` (or `.claude-plugin/plugin.json`).
+
+See: [Plugin template](references/TEMPLATES.md#plugin-template) and [Plugin architecture notes](references/ARCHITECTURE.md#plugins).
+
+## 2) When code changes are actually required
+
+Choose the correct codebase / package:
+
+- **CLI UX, TUI widgets, slash commands, config files** → `OpenHands/OpenHands-CLI`
+- **Agent runtime logic, skill loading, hooks system, plugin loader, event model** → `OpenHands/software-agent-sdk` → `openhands-sdk`
+- **Built-in tools (Terminal, File Editor, Task Tracker, etc.)** → `OpenHands/software-agent-sdk` → `openhands-tools`
+- **Workspace mounting / file sync / workspace abstraction** → `OpenHands/software-agent-sdk` → `openhands-workspace`
+
+If you’re not sure, start by reproducing the issue and locating the code path.
+
+## 3) uv development workflows (including `uv tool install openhands`)
+
+- If they run OpenHands from a **local clone**: use `uv sync` + `uv run openhands`.
+- If they installed OpenHands via **`uv tool install openhands`**:
+  - safest: `uv tool run --from /path/to/OpenHands-CLI openhands`
+  - persistent: `uv tool install --force --editable /path/to/OpenHands-CLI`
+- If they need to modify the **SDK** while running the CLI, remember: the CLI often **pins exact SDK/tool versions**. Prefer aligning versions; use overrides only for local development.
+
+Details: [Installation modes (`uv tool` vs `uv run`)](references/INSTALLATION_MODES.md).
+
+## 4) What requires a restart to take effect?
+
+- **Skills / AGENTS.md**: loaded at conversation start → usually **start a new conversation** to pick them up. Ask the user to fully restart OpenHands only if needed.
+- **Hooks (.openhands/hooks.json)**: loaded when the agent/conversation initializes → ask the user to restart (or create a fresh session) to guarantee reload.
+- **Plugins**: loaded when the conversation initializes → ask the user to restart (or create a fresh session) to guarantee reload.
+- **Editable Python code (uv tool `--editable` or local `uv run`)**: changes apply the next time the Python process starts → ask the user to restart the CLI/app.
+
+## 5) Verification checklist
+
+After making a “self-modification” change, guide the **user** through verification:
+
+1. Ask the user to start a **fresh conversation** (and restart OpenHands only if needed).
+2. If they’re using the **OpenHands CLI TUI**, ask them to type `/skills` to confirm the expected skills/hooks/MCPs are loaded.
+3. Ask the user to send a tiny, deterministic **test prompt** that should trigger the new behavior.
+4. If it’s a code change, run the smallest relevant test suite (or ask the user to run it) and confirm the behavior matches.
+
+## References
+
+- [Architecture: CLI vs SDK vs skills/hooks/plugins](references/ARCHITECTURE.md)
+- [Installation modes: `uv tool` vs `uv run` and version pinning](references/INSTALLATION_MODES.md)
+- [Templates: SKILL.md, hooks.json, plugin.json, .mcp.json](references/TEMPLATES.md)
diff --git a/skills/adapt-yourself/references/ARCHITECTURE.md b/skills/adapt-yourself/references/ARCHITECTURE.md
new file mode 100644
index 0000000..1dc140b
--- /dev/null
+++ b/skills/adapt-yourself/references/ARCHITECTURE.md
@@ -0,0 +1,101 @@
+# OpenHands extensibility architecture (CLI + SDK)
+
+OpenHands “behavior” can be modified at multiple layers. The key is to choose the layer that matches the user’s intent.
+
+## Components
+
+### OpenHands CLI (`OpenHands/OpenHands-CLI`)
+
+- Python package name: `openhands`
+- Entry points: `openhands`, `openhands-acp`
+- Responsibilities:
+  - Terminal UI (Textual)
+  - Loading config + wiring up an SDK `Agent`
+  - Choosing tools and building an `AgentContext`
+
+In the CLI codebase, look for where it creates the `AgentContext` and enables:
+- project skills
+- user skills
+- public skills
+
+### Software Agent SDK monorepo (`OpenHands/software-agent-sdk`)
+
+This repo is a **uv workspace** that produces multiple packages:
+
+- `openhands-sdk` (module: `openhands.sdk`)
+  - agent runtime primitives (Agent, Conversation, Event model)
+  - context loading (skills)
+  - hooks system
+  - plugin loader
+
+- `openhands-tools` (module: `openhands.tools`)
+  - built-in tools used by many OpenHands projects
+
+- `openhands-workspace` (module: `openhands.workspace`)
+  - workspace mounting, file sync abstractions
+
+- `openhands-agent-server`
+  - server-side runtime
+
+## Skills
+
+Skills are loaded from multiple locations; **project-local skills override user + public**.
+
+Common locations:
+
+- Project-local:
+  - `<repo>/.agents/skills/`
+
+- User-global:
+  - `~/.agents/skills/`
+
+- Public registry:
+  - `https://github.com/OpenHands/extensions` (cached locally under `~/.openhands/skills-cache/`)
+
+Skill types:
+
+- **AgentSkills format**: a directory with `SKILL.md` (progressive disclosure)
+- **Always-on instructions**: repo instruction files like `AGENTS.md` (loaded at conversation start)
+
+## Hooks
+
+Hooks are configured via JSON and executed on events (e.g. `pre_tool_use`).
+
+Typical locations (depending on project / SDK config):
+
+- `<repo>/.openhands/hooks.json`
+- `~/.openhands/hooks.json`
+
+Hook scripts receive an event payload on stdin as JSON and respond with a decision.
+
+## Plugins
+
+Plugins are bundles of:
+
+- skills
+- hooks
+- MCP config
+- optional commands and agents
+
+They follow the Claude Code-compatible directory structure:
+
+```
+my-plugin/
+├── .plugin/
+│   └── plugin.json
+├── skills/
+├── hooks/
+│   └── hooks.json
+├── .mcp.json
+└── README.md
+```
+
+Plugins can be:
+- loaded from a local path
+- fetched from GitHub/git URLs and cached under `~/.openhands/cache/plugins/`
+
+## MCP servers (tools)
+
+MCP servers provide tool endpoints to the agent.
+
+They are configured via `.mcp.json` (location depends on host project). Plugins can also provide MCP config.
diff --git a/skills/adapt-yourself/references/INSTALLATION_MODES.md b/skills/adapt-yourself/references/INSTALLATION_MODES.md
new file mode 100644
index 0000000..18b965b
--- /dev/null
+++ b/skills/adapt-yourself/references/INSTALLATION_MODES.md
@@ -0,0 +1,114 @@
+# Installation and development modes (uv)
+
+This reference is for making “persistent self-modifications” that apply after restart.
+
+## `uv run` (project environment) mode
+
+Use when you are developing from a git clone.
+
+- `uv sync` installs dependencies into a project-local `.venv/` (and/or uses the uv lockfile).
+- `uv run ...` runs inside that environment.
+
+Typical CLI dev flow:
+
+```bash
+cd OpenHands-CLI
+uv sync --group dev
+uv run openhands
+```
+
+Typical SDK dev flow:
+
+```bash
+cd software-agent-sdk
+uv sync --group dev
+uv run pytest
+```
+
+### Pros
+
+- Simple debugging (local sources)
+- Easy to run tests + linters
+- No global tool replacement needed
+
+### Cons
+
+- You must run `openhands` via `uv run openhands` (or ensure the venv is on PATH)
+
+## `uv tool install openhands` (tool environment) mode
+
+In this mode, `uv` installs the CLI into a dedicated tool environment and exposes a shim executable.
+
+### Inspect installed tools
+
+```bash
+uv tool list
+uv tool dir
+```
+
+If `openhands` is installed and not found on PATH, run:
+
+```bash
+uv tool update-shell
+```
+
+### Replace the tool with an editable local checkout
+
+This is the most common way to run a modified CLI *after restart*:
+
+```bash
+uv tool install --force --editable /path/to/OpenHands-CLI
+```
+
+After that, `openhands` should use the editable sources.
+
+### Run a local checkout without installing
+
+This is the safest way to test a change:
+
+```bash
+uv tool run --from /path/to/OpenHands-CLI openhands
+```
+
+## SDK vs CLI version pins
+
+The CLI often pins exact versions of:
+
+- `openhands-sdk`
+- `openhands-tools`
+- `openhands-workspace`
+
+If you want to use a local SDK checkout *together with the CLI*, you must handle that pinning.
+
+### What usually works
+
+1. Check the versions the CLI requires:
+   - `OpenHands-CLI/pyproject.toml`
+2. Check out the matching tag/commit in `software-agent-sdk`.
+3. Install the matching SDK packages in editable mode into the environment that runs the CLI.
+
+### What commonly fails
+
+Trying to install an SDK checkout at version `X` into an environment where the CLI requires `openhands-sdk==Y`.
+
+### Using `--overrides` for development
+
+As a last resort for local development, you can override pinned deps.
+
+Create a file (e.g. `overrides.txt`) containing a direct URL or local path spec:
+
+```text
+openhands-sdk @ file:///absolute/path/to/software-agent-sdk/openhands-sdk
+openhands-tools @ file:///absolute/path/to/software-agent-sdk/openhands-tools
+openhands-workspace @ file:///absolute/path/to/software-agent-sdk/openhands-workspace
+```
+
+Then install the CLI tool using that overrides file:
+
+```bash
+uv tool install --force --editable /path/to/OpenHands-CLI --overrides overrides.txt
+```
+
+Notes:
+- This is intentionally “break glass”; it can produce a mismatched runtime.
+- Prefer aligning versions whenever possible.
diff --git a/skills/adapt-yourself/references/TEMPLATES.md b/skills/adapt-yourself/references/TEMPLATES.md
new file mode 100644
index 0000000..0937d65
--- /dev/null
+++ b/skills/adapt-yourself/references/TEMPLATES.md
@@ -0,0 +1,156 @@
+# Templates
+
+Use these as starting points when implementing a “self-modification” without changing OpenHands code.
+
+## AgentSkills (`SKILL.md`) template
+
+If you decide an Agentskill is appropriate for user's request to remember or adapt to their needs, you can create it. First, create a directory:
+
+```
+<repo>/.agents/skills/<skill-name>/
+└── SKILL.md
+```
+
+Example `SKILL.md`:
+
+```md
+---
+name: my-team-workflow
+description: Our team’s workflow for X. Use when the user asks about X.
+triggers:
+- team x workflow
+- how do we do x
+---
+
+# Our workflow for X
+
+## When to use
+
+Use this workflow when...
+
+## Steps
+
+1. ...
+2. ...
+
+## Verification
+
+- Run: `...`
+```
+
+Notes:
+- Keep triggers distinctive.
+- Keep the file focused; put deep details in `references/`.
+
+## Always-on repo instructions (`AGENTS.md`) template
+
+Create `<repo>/AGENTS.md`:
+
+```md
+# Project rules
+
+## Build / run
+- `make install`
+- `make test`
+
+## Code style
+- Ruff formatting
+
+## Safety
+- Never run destructive commands without asking
+```
+
+## Hooks (`.openhands/hooks.json`) template
+
+If you decide a Hook is appropriate for addressing the user's request to adapt for next time, you can create a hook. Create or use `<repo>/.openhands/hooks.json`:
+
+```json
+{
+  "pre_tool_use": [
+    {
+      "command": "bash .openhands/hooks/block-dangerous.sh",
+      "matchers": {
+        "tool_name": "terminal"
+      }
+    }
+  ]
+}
+```
+
+Then create the script `<repo>/.openhands/hooks/block-dangerous.sh`:
+
+```bash
+#!/usr/bin/env bash
+
+# stdin: HookEvent JSON
+payload="$(cat)"
+
+# Example: block `rm -rf ...` for Terminal tool calls.
+# Requires: jq
+cmd="$(echo "$payload" | jq -r '.tool_input.command // ""')"
+
+if echo "$cmd" | grep -Eq '(^|[[:space:]])rm[[:space:]]+-rf([[:space:]]|$)'; then
+  # HookExecutor parses JSON on stdout; "decision" must be "allow" or "deny".
+  jq -n --arg reason "Blocked destructive command" '{"decision":"deny","reason":$reason}'
+  exit 0
+fi
+
+jq -n '{"decision":"allow"}'
+```
+
+Notes:
+- Prefer parsing the structured fields (like `.tool_input.command`) rather than grepping raw JSON.
+- The exact `tool_input` shape is tool-specific; inspect the SDK `HookEvent` schema and your tool inputs.
+
+## Plugin template
+
+Plugin structure:
+
+```
+my-plugin/
+├── .plugin/
+│   └── plugin.json
+├── skills/
+│   └── my-skill/
+│       └── SKILL.md
+├── hooks/
+│   └── hooks.json
+├── .mcp.json
+└── README.md
+```
+
+Minimal `.plugin/plugin.json` (fields vary by plugin schema):
+
+```json
+{
+  "name": "my-plugin",
+  "version": "0.1.0",
+  "description": "My custom OpenHands behavior",
+  "author": {
+    "name": "Example",
+    "email": "dev@example.com"
+  }
+}
+```
+
+## MCP template (`.mcp.json`)
+
+Minimal example:
+
+```json
+{
+  "mcpServers": {
+    "my-server": {
+      "command": "python",
+      "args": ["-m", "my_mcp_server"],
+      "env": {
+        "MY_SERVER_TOKEN": "$MY_SERVER_TOKEN"
+      }
+    }
+  }
+}
+```
+
+Notes:
+- Never hardcode real secrets in files.
+- Prefer environment variables.