diff --git a/.github/prompts/endgame.prompt.md b/.github/prompts/endgame.prompt.md
deleted file mode 100644
index f9d9a969..00000000
--- a/.github/prompts/endgame.prompt.md
+++ /dev/null
@@ -1,158 +0,0 @@
-# Endgame Verification Prompt
-
-## Instructions
-
-**⚠️ CRITICAL - YOU MUST FOLLOW THESE RULES:**
-1. **Do NOT run `gh pr view` or `gh issue view` for individual tasks** - only run it once for the main endgame issue
-2. **Do NOT research or analyze any task yourself**
-3. **Do NOT create any verification files yourself**
-4. **IMMEDIATELY call `runSubagent` for each task** after parsing the issue - no delays, no research
-
-Your workflow is ONLY:
-1. Create initial todo: "Fetch endgame issue and parse tasks"
-2. Fetch the endgame issue (ONE `gh issue view` call)
-3. Parse to get task list, then ADD a todo item for each task found
-4. Create output directory
-5. Call `runSubagent` for each task - mark todo complete when subagent returns
-
-### Steps
-
-1. **Create initial todo list**:
-   Use `manage_todo_list` to create the first todo:
-   - Todo 1: "Fetch endgame issue" (mark as in-progress)
-
-2. **Ask the user** for the following information:
-   - The GitHub endgame issue link (e.g., `https://github.com/microsoft/copilot-for-eclipse/issues/XXXX`)
-   - The user's GitHub account name
-
-3. **Fetch the endgame issue** (this is the ONLY `gh issue view` you should run):
-   ```shell
-   gh issue view <issue_number> --repo microsoft/copilot-for-eclipse
-   ```
-   Parse the issue body to find all tasks (checkboxes) assigned to the specified user.
-   Extract the task title and any linked PR/issue URL as plain text.
-   **STOP - do NOT fetch any of the linked PRs or issues.**
-
-4. **Update todo list with all tasks found**:
-   Use `manage_todo_list` to:
-   - Mark "Fetch endgame issue" as completed
-   - ADD a new todo for each task found (e.g., "Task 1: <title>", "Task 2: <title>", etc.)
-   - All new task todos should be "not-started"
-
-5. **Create the output directory**:
-   ```shell
-   mkdir -p .github/endgame/<issue_number>
-   ```
-
-6. **For each task, mark todo as in-progress, then call `runSubagent`**:
-   
-   For each task:
-   1. Update todo list - mark that task as "in-progress"
-   2. Call `runSubagent` tool with:
-      - **description**: "Endgame: <short_task_title>"
-      - **prompt**: The template below filled in with only the info you extracted
-   3. When subagent returns, mark that task's todo as "completed"
-   
-   **Subagent prompt template**:
-   
-   ---
-   ## Task Details
-   - Task Number: <N>
-   - Task Title: <task_title>
-   - Assignee: <username>
-   - Related Issue/PR: <link_if_available> (NOT YET FETCHED - you must fetch this)
-   
-   ## Your Mission
-   YOU (the subagent) must research this task AND create the verification file.
-   
-   ### Step 1: Research the Task
-   - If there is a linked PR/issue, fetch it using `gh pr view` or `gh issue view`
-   - Understand what feature/fix needs to be verified
-   - **If anything is ambiguous or unclear after reading the PR/issue, ask the user
-     for clarification before proceeding. Keep asking until you have enough
-     information to write concrete, accurate test steps.**
-   
-   ### Step 2: Create Verification File (YOU must create this file)
-   Create the file: `.github/endgame/<issue_number>/<N>_<task_slug>.md`
-   
-   Use **exactly** this format (modelled on the project test-plan style under
-   `com.microsoft.copilot.eclipse.swtbot.test/test-plans/`):
-   
-   ```markdown
-   # <Feature / Task Title>
-   
-   ## Overview
-   <1-3 sentences: what is being verified and why it matters.>
-   
-   Entry points:
-   - <Primary UI/menu path to reach the feature>
-   
-   Not exercised:
-   - <Known out-of-scope items, or omit this block if none>
-   
-   ---
-   
-   ## Prerequisites
-   
-   - Eclipse IDE with the GitHub Copilot for Eclipse plugin installed and activated.
-   - <Additional prerequisite specific to this feature>
-   
-   ---
-   
-   ## 1. <Scenario group name, e.g. "Happy-path verification">
-   
-   ### TC-001: <Specific test case title>
-   
-   **Type:** `Happy Path`
-   **Priority:** `P0`
-   
-   #### Preconditions
-   - <Specific state required before starting these steps>
-   
-   #### Steps
-   1. <Detailed, concrete step>
-   2. <Next step>
-   3. <Continue as needed>
-   
-   #### Expected Result
-   - <Observable outcome that proves the feature works>
-   - <Second outcome if needed>
-   
-   #### Key Screenshots
-   - [ ] **<Label>** -- <What to capture in this screenshot>
-   
-   #### Notes on failure modes
-   - <Common failure symptom> -- <Likely cause and where to look>
-   ```
-   
-   Guidelines for filling in the template:
-   - **Overview**: 1-3 sentences max. Do not repeat what is already covered by
-     Prerequisites or the Steps.
-   - **Type** values: `Happy Path`, `Negative`, `Edge Case`, `Regression`.
-   - **Priority** values: `P0` (must-pass), `P1` (high), `P2` (medium).
-   - Add additional `### TC-NNN` blocks (3-digit zero-padded: TC-002, TC-003, …) if multiple distinct scenarios are needed.
-   - Omit `#### Notes on failure modes` if there are no obvious failure modes to call out.
-   - If you are still unsure about any step after researching, **ask the user**
-     before writing that step -- do not guess.
-   
-   ### Step 3: Return Summary
-   Return ONLY:
-   - File path created
-   - One-line summary of what needs to be verified
-   ---
-
-7. **After all subagents complete**, provide:
-   - Summary table of all generated verification files
-   - Total tasks processed
-   - Any tasks that could not be processed (with reasons)
-
----
-
-## Notes
-
-- **NEVER run `gh pr view` or `gh issue view` on task links** - only on the main endgame issue
-- **NEVER analyze or research tasks yourself** - immediately delegate to subagents
-- Each subagent runs independently
-- Subagents should be concise - create the file and return a brief summary
-- If a task is unclear after asking the user, note the outstanding question in the verification file
-- Use slugified task titles for filenames (lowercase, hyphens, no special chars)
diff --git a/.github/skills/endgame/SKILL.md b/.github/skills/endgame/SKILL.md
new file mode 100644
index 00000000..2aec5de6
--- /dev/null
+++ b/.github/skills/endgame/SKILL.md
@@ -0,0 +1,173 @@
+---
+name: endgame
+description: Orchestrate endgame verification for a GitHub milestone issue. Fetches the issue, parses assigned tasks, delegates each task to a subagent that researches the linked PR/issue and writes a test plan, and saves every plan to test-plans/ following the project's standard test-plan format.
+---
+
+# Endgame Verification Skill
+
+Use this skill to run an endgame verification pass for a GitHub milestone issue.
+Each task in the issue is delegated to a subagent that researches the linked
+PR/issue and writes a test plan under `test-plans/`.
+
+## Workflow
+
+**⚠️ CRITICAL — YOU MUST FOLLOW THESE RULES:**
+
+1. **Do NOT run `gh pr view` or `gh issue view` for individual tasks** — only
+   run it once for the main endgame issue.
+2. **Do NOT research or analyse any task yourself.**
+3. **Do NOT create any test-plan files yourself.**
+4. **IMMEDIATELY call `runSubagent` for each task** after parsing the issue —
+   no delays, no research.
+
+---
+
+### Step 1 — Ask the user for inputs
+
+Ask for:
+- The GitHub endgame issue URL
+  (e.g. `https://github.com/microsoft/copilot-for-eclipse/issues/XXXX`)
+- The user's GitHub account name
+
+---
+
+### Step 2 — Fetch the endgame issue (ONE call only)
+
+```shell
+gh issue view <issue_number> --repo microsoft/copilot-for-eclipse --json body --jq '.body'
+```
+
+Parse the body to find **all tasks (checkboxes) assigned to the specified
+user**. Extract each task's title and any linked PR/issue URL as plain text.
+
+**STOP — do NOT fetch any of the linked PRs or issues.**
+
+---
+
+### Step 3 — Process each task via subagent
+
+For each task, call `runSubagent` with:
+- **description**: `"Endgame: <short_task_title>"`
+- **prompt**: the template below, filled in with only the info you extracted.
+
+#### Subagent prompt template
+
+---
+## Task Details
+- Task Number: \<N\>
+- Task Title: \<task_title\>
+- Assignee: \<username\>
+- Related Issue/PR: \<link_if_available\> (NOT YET FETCHED — you must fetch this)
+
+## Your Mission
+
+YOU (the subagent) must research this task AND create the test-plan file.
+
+### Step 1: Research the task
+
+- Fetch the linked PR or issue with `gh pr view` / `gh issue view`.
+- Understand what feature or fix needs to be verified.
+- **If anything is ambiguous or unclear after reading the PR/issue, ask the
+  user for clarification before proceeding. Keep asking until you have enough
+  information to write concrete, accurate test steps.**
+
+### Step 2: Create the test-plan file
+
+Determine a short `<feature-slug>` (lowercase, hyphens, no special chars)
+derived from the task title (e.g. `chat-history-restore`).
+
+Create the directory and file:
+
+```
+test-plans/<feature-slug>/<feature-slug>.md
+```
+
+Use **exactly** this format, matching the project's existing test plans
+(see `test-plans/thinking-persistence/thinking-persistence.md` for a live
+example):
+
+```markdown
+# <Feature / Task Title>
+
+## Overview
+<1–3 sentences: what is being verified and why it matters.>
+
+---
+
+## Test Cases
+
+### TC-001: <Specific test case title>
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- <Specific state required before starting these steps>
+
+#### Steps
+1. <Detailed, concrete step>
+2. <Next step>
+3. <Continue as needed>
+
+#### Expected Result
+- <Observable outcome that proves the feature works>
+
+#### 📸 Key Screenshots
+- [ ] **<Label>** — <What to capture>
+
+---
+
+### TC-002: <Next scenario if needed>
+<!-- Repeat TC block for each distinct scenario. Use TC-003, TC-004, … -->
+
+---
+
+## Screenshots Checklist
+> Consolidated list of all key screenshot moments.
+
+- [ ] `TC-001` <Label>
+- [ ] `TC-002` <Label>
+```
+
+Guidelines:
+- **Overview**: 1–3 sentences max. Do not repeat what is already covered by
+  Preconditions or Steps.
+- **Type** values: `Happy Path`, `Negative`, `Edge Case`, `Regression`.
+- **Priority** values: `P0` (must-pass), `P1` (high), `P2` (medium).
+- Use 3-digit zero-padded TC numbers: TC-001, TC-002, TC-003, …
+- Omit `📸 Key Screenshots` within a TC block if there are no meaningful
+  screenshots to capture for that case.
+- Keep the `## Screenshots Checklist` section at the end — list every
+  screenshot across all TC blocks.
+- If you are still unsure about any step after researching, **ask the user**
+  before writing that step — do not guess.
+
+### Step 3: Return a brief summary
+
+Return ONLY:
+- File path created (e.g. `test-plans/chat-history-restore/chat-history-restore.md`)
+- One-line summary of what needs to be verified
+
+---
+
+### Step 4 — Final summary
+
+After all subagents complete, provide:
+- A table of all generated test-plan files
+- Total tasks processed
+- Any tasks that could not be processed (with reasons)
+
+---
+
+## Notes
+
+- **NEVER run `gh pr view` or `gh issue view` on task links yourself** — only
+  on the main endgame issue. Subagents handle the individual links.
+- **NEVER analyse or research tasks yourself** — immediately delegate to
+  subagents.
+- Each subagent runs independently.
+- Subagents should be concise — create the file and return a brief summary.
+- If a task is still unclear after the user is asked, note the outstanding
+  question inside the test-plan file.
+- Use slugified task titles for the feature slug and directory name
+  (lowercase, hyphens, no special characters).
diff --git a/com.microsoft.copilot.eclipse.swtbot.test/probe-scripts/endgame-chat-open-001.json b/com.microsoft.copilot.eclipse.swtbot.test/probe-scripts/endgame-chat-open-001.json
new file mode 100644
index 00000000..982e6ba2
--- /dev/null
+++ b/com.microsoft.copilot.eclipse.swtbot.test/probe-scripts/endgame-chat-open-001.json
@@ -0,0 +1,31 @@
+[
+  { "action": "waitForIdle" },
+  { "action": "screenshot", "id": "01-startup" },
+
+  { "action": "showView", "idRef": "com.microsoft.copilot.eclipse.ui.chat.ChatView" },
+  { "action": "waitForIdle" },
+  { "action": "screenshot", "id": "02-chat-open" },
+
+  { "action": "assertExists",
+    "locator": { "by": "viewId", "id": "com.microsoft.copilot.eclipse.ui.chat.ChatView" } },
+
+  { "action": "waitFor",
+    "locator": { "by": "styledText" },
+    "timeoutSec": 30 },
+
+  { "action": "dumpUi", "id": "03-chat-widgets" },
+  { "action": "screenshot", "id": "03-chat-ready" },
+
+  { "action": "waitForMethod",
+    "locator": { "by": "widgetId", "value": "model-picker" },
+    "method": "getSelectedItemId",
+    "timeoutSec": 60 },
+
+  { "action": "screenshot", "id": "04-model-loaded" },
+  { "action": "assertExists",
+    "locator": { "by": "widgetId", "value": "model-picker" } },
+
+  { "action": "screenshot", "id": "05-final" },
+  { "action": "assertExists",
+    "locator": { "by": "styledText" } }
+]
diff --git a/test-plans/cls-session-persistence/cls-session-persistence.md b/test-plans/cls-session-persistence/cls-session-persistence.md
new file mode 100644
index 00000000..454c5b8b
--- /dev/null
+++ b/test-plans/cls-session-persistence/cls-session-persistence.md
@@ -0,0 +1,197 @@
+# CLS Session Persistence and Restoration for Conversation History
+
+## Overview
+This feature integrates CLS (Copilot Language Server) server-side session persistence so that when a conversation is restored the CLS receives the original `conversationId` and `restoreToTurnId`, enabling full context continuation rather than relying solely on IDE-side turn history. It also configures `<user home>/.copilot/eclipse` as the transcript directory (matching IntelliJ's convention).
+
+> **Path note:** On macOS/Linux this directory is `~/.copilot/eclipse`; on Windows it is `%USERPROFILE%\.copilot\eclipse`.
+
+---
+
+## Test Cases
+
+### TC-001: Conversation history displays correctly after session switch
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- Copilot Chat is open
+- At least one conversation with multiple turns exists in history
+
+#### Steps
+1. Create a conversation with at least 3 user–Copilot turn pairs
+2. Open chat history and switch to a different conversation
+3. Open chat history again and select the original conversation
+
+#### Expected Result
+- All user and Copilot turns are displayed in the correct order
+- Each turn shows its full content (no missing or truncated messages)
+- Turn pairing is correct: each user message is followed by its corresponding Copilot response
+
+#### 📸 Key Screenshots
+- [ ] **Before switch** — Full conversation with all turns visible
+- [ ] **After restore** — Same conversation restored with all turns in correct order
+
+---
+
+### TC-002: Conversation history persists across Eclipse restarts
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- Eclipse is open with Copilot Chat functional
+- The `<user home>/.copilot/eclipse` directory does not exist (or is empty) before the test
+
+#### Steps
+1. Open Copilot Chat and send at least 2 messages; wait for responses
+2. Close Eclipse completely (File → Exit or equivalent)
+3. Verify that the `<user home>/.copilot/eclipse` directory exists and contains transcript files
+4. Reopen Eclipse and open Copilot Chat
+5. Open chat history and select the conversation from before the restart
+
+#### Expected Result
+- The conversation appears in chat history with all turns intact
+- Continuing the conversation produces contextually aware responses (CLS uses transcript for context)
+
+#### 📸 Key Screenshots
+- [ ] **Transcript directory** — `<user home>/.copilot/eclipse` folder visible with transcript files
+- [ ] **After restart** — Conversation restored in chat history after Eclipse reopen
+
+---
+
+### TC-003: New conversation starts without interference from restored state
+
+**Type:** `Regression`
+**Priority:** `P0`
+
+#### Preconditions
+- At least one existing conversation in chat history
+
+#### Steps
+1. Open chat history and restore an existing conversation
+2. Click "New Chat" to start a fresh conversation
+3. Send a message unrelated to any prior conversation
+4. Verify Copilot responds without referencing prior conversations
+
+#### Expected Result
+- New conversation has a clean context; no cross-contamination from the restored conversation
+- CLS receives a new `conversationId` (no `restoreToTurnId` for a brand-new conversation)
+
+---
+
+### TC-004: Transcript directory created on first startup
+
+**Type:** `Happy Path`
+**Priority:** `P1`
+
+#### Preconditions
+- `<user home>/.copilot/eclipse` directory does not exist
+
+#### Steps
+1. Start Eclipse with the Copilot plugin installed
+2. Wait for the Copilot language server to finish initializing (status bar shows Copilot ready)
+3. Check the file system for `<user home>/.copilot/eclipse`
+
+#### Expected Result
+- `<user home>/.copilot/eclipse` directory is created automatically on startup
+- The path mirrors IntelliJ's `<user home>/.copilot/jb` convention (different subdirectory, same parent)
+
+#### 📸 Key Screenshots
+- [ ] **Directory exists** — File explorer showing `<user home>/.copilot/eclipse` after startup
+
+---
+
+### TC-005: Partially completed conversation restores to the last completed turn
+
+**Type:** `Edge Case`
+**Priority:** `P1`
+
+#### Preconditions
+- Copilot Chat is open in Agent mode
+
+#### Steps
+1. Send a message and wait for Copilot to begin responding
+2. Click Cancel while the response is still streaming
+3. Open chat history, switch to a different conversation
+4. Open chat history again and select the original conversation
+5. Send a new follow-up message
+
+#### Expected Result
+- The restored conversation shows the partial/cancelled turn in the UI
+- The `restoreToTurnId` sent to CLS corresponds to the last *fully completed* turn, not the cancelled one
+- Copilot's next response is coherent with the conversation up to the last completed turn
+
+#### 📸 Key Screenshots
+- [ ] **After cancel** — Conversation with cancelled/partial turn visible
+- [ ] **After restore + new message** — New response coherent with completed context
+
+---
+
+### TC-006: No 400 Bad Request after restoring a tool-call conversation
+
+**Type:** `Regression`
+**Priority:** `P0`
+
+#### Preconditions
+- Copilot Chat is open in Agent mode
+- A workspace with at least one source file is open
+
+#### Steps
+1. Send a message that triggers at least one tool call and gets a full assistant response (e.g. "list the files in this project and summarise what each one does")
+2. Wait for the full multi-turn exchange to complete (tool calls executed, final assistant reply shown)
+3. Restart Eclipse IDE
+4. Open Copilot Chat and restore the previous conversation from history
+5. Send a new follow-up message in the restored conversation
+
+#### Expected Result
+- The follow-up message is sent without any 400 Bad Request error
+- The assistant replies successfully on the first attempt (no error banner)
+
+#### 📸 Key Screenshots
+- [ ] **Completed original turn** — Chat showing the tool-call exchange and final assistant reply before restart
+- [ ] **After restore + reply** — Restored history with successful follow-up reply and no error
+
+---
+
+### TC-007: Restored history shows no duplicated user messages
+
+**Type:** `Regression`
+**Priority:** `P0`
+
+#### Preconditions
+- Copilot Chat is open in Agent mode
+- A workspace with at least one source file is open
+
+#### Steps
+1. Send a message that triggers at least one tool call (e.g. "read the contents of README.md and explain it")
+2. Wait for the full response including the assistant's explanation
+3. Note the exact assistant response text
+4. Restart Eclipse IDE
+5. Open Copilot Chat and restore the conversation from history
+6. Inspect the restored conversation turns
+
+#### Expected Result
+- The assistant's previous response text does not appear as a user message in the restored conversation
+- Each turn shows the correct role (user prompt → tool call → assistant reply), with no duplicated content
+
+#### 📸 Key Screenshots
+- [ ] **Original conversation** — Chat showing the original tool-call turn and assistant reply
+- [ ] **Restored conversation** — Restored history with correctly attributed messages and no duplicates
+
+---
+
+## Screenshots Checklist
+> Consolidated list of all key screenshot moments.
+
+- [ ] `TC-001` Full conversation before session switch
+- [ ] `TC-001` Same conversation after restore, all turns in correct order
+- [ ] `TC-002` `<user home>/.copilot/eclipse` directory with transcript files
+- [ ] `TC-002` Conversation restored in history after Eclipse restart
+- [ ] `TC-004` `<user home>/.copilot/eclipse` directory created on first startup
+- [ ] `TC-005` Conversation with cancelled turn visible
+- [ ] `TC-005` New response coherent with completed context after restore
+- [ ] `TC-006` Completed tool-call turn before restart
+- [ ] `TC-006` Restored history with successful follow-up and no 400 error
+- [ ] `TC-007` Original conversation showing assistant reply
+- [ ] `TC-007` Restored conversation with correctly attributed messages and no duplicates
diff --git a/test-plans/skills-prompt-files/skills-prompt-files.md b/test-plans/skills-prompt-files/skills-prompt-files.md
new file mode 100644
index 00000000..b88cb673
--- /dev/null
+++ b/test-plans/skills-prompt-files/skills-prompt-files.md
@@ -0,0 +1,234 @@
+# Support Skills and Prompt Files
+
+## Overview
+This feature adds support for agent skills (`SKILL.md`) and custom prompt files (`.prompt.md`) in Agent Mode. Skills and prompts are discovered from workspace projects and user-global directories, exposed as slash commands in the chat input, and controlled by an **Enable Skills** toggle in `Window → Preferences → Copilot → Chat`.
+
+---
+
+## Test Cases
+
+### TC-001: Enable Skills toggle appears in Chat preferences
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- Eclipse is running with the Copilot plugin installed
+- A valid GitHub Copilot subscription is active (client-preview feature flag must be on)
+
+#### Steps
+1. Open **Window → Preferences → Copilot → Chat**
+2. Observe the preference page content
+
+#### Expected Result
+- An **Enable Skills** checkbox is visible
+- A note below it reads "Controls whether agent skills can be used to enrich chat context."
+- The checkbox is checked by default (enabled)
+
+#### 📸 Key Screenshots
+- [ ] **Preferences page** — Chat preferences showing the Enable Skills toggle (checked)
+
+---
+
+### TC-002: Workspace-scoped skill appears as slash command in Agent mode
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- Eclipse is running with a project open in the workspace
+- Copilot chat is open in **Agent** mode
+- **Enable Skills** is checked in preferences
+
+#### Steps
+1. Create the directory `.github/skills/my-skill/` inside any open project
+2. Create `SKILL.md` inside that directory with the following content:
+   ```markdown
+   ---
+   name: My Skill
+   description: A test skill for verification
+   ---
+   This skill provides test context.
+   ```
+3. In the Copilot chat input (Agent mode), type `/` and observe the slash command popup
+
+#### Expected Result
+- `my-skill` appears in the slash command list with description "A test skill for verification"
+- The display name shown is `My Skill` (from `shortDescription`/name front matter)
+
+#### 📸 Key Screenshots
+- [ ] **Slash command popup** — `/my-skill` visible in the Agent mode autocomplete list
+
+---
+
+### TC-003: Prompt file appears as slash command in Agent mode
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- Eclipse is running with a project open in the workspace
+- Copilot chat is open in **Agent** mode
+- **Enable Skills** is checked in preferences
+
+#### Steps
+1. Create a `.github/` directory (or any supported location) inside the project
+2. Create a file named `review.prompt.md` in the project with some markdown content
+3. In the Copilot chat input (Agent mode), type `/` and observe the slash command popup
+
+#### Expected Result
+- `review` (or the prompt file's ID) appears in the slash command list
+- The template is listed alongside built-in slash commands
+
+#### 📸 Key Screenshots
+- [ ] **Slash command popup** — Custom prompt file visible as slash command in Agent mode
+
+---
+
+### TC-004: Skills do NOT appear in Ask mode
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- A workspace-scoped `SKILL.md` exists (as created in TC-002)
+- **Enable Skills** is checked in preferences
+
+#### Steps
+1. Switch the Copilot chat to **Ask** mode
+2. In the chat input, type `/` and observe the slash command popup
+
+#### Expected Result
+- The skill (`my-skill`) does **not** appear in the Ask mode slash command list
+- Only built-in `chat-panel`-scoped commands are shown
+
+#### 📸 Key Screenshots
+- [ ] **Ask mode popup** — Skill absent from the slash command list in Ask mode
+
+---
+
+### TC-005: Disabling Skills hides skills from slash command list
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- A workspace-scoped `SKILL.md` exists and was visible in Agent mode (TC-002 completed)
+
+#### Steps
+1. Open **Window → Preferences → Copilot → Chat**
+2. Uncheck **Enable Skills** and click **Apply and Close**
+3. In the Copilot chat (Agent mode), type `/` and observe the slash command popup
+
+#### Expected Result
+- The skill is no longer listed in the slash command popup
+- Built-in slash commands (e.g. `/fix`, `/explain`) still appear if scoped for Agent mode
+
+#### 📸 Key Screenshots
+- [ ] **Preferences page** — Enable Skills unchecked
+- [ ] **Slash command popup** — Skill absent after disabling
+
+---
+
+### TC-006: Auto-refresh when SKILL.md is added or removed
+
+**Type:** `Happy Path`
+**Priority:** `P1`
+
+#### Preconditions
+- Eclipse is running with a project open
+- Copilot chat is open in Agent mode
+- **Enable Skills** is checked
+
+#### Steps
+1. In the chat input, type `/` — note that the skill does not yet exist
+2. In Eclipse's Project Explorer, create `.github/skills/refresh-skill/SKILL.md` with a description
+3. Without restarting, return to the chat input and type `/` again
+4. Then delete the `SKILL.md` file and type `/` once more
+
+#### Expected Result
+- After adding: `refresh-skill` appears in the slash command list without restarting Eclipse
+- After deleting: `refresh-skill` disappears from the slash command list without restarting Eclipse
+
+#### 📸 Key Screenshots
+- [ ] **Before add** — Slash command popup without the skill
+- [ ] **After add** — Slash command popup with the new skill
+- [ ] **After delete** — Slash command popup with skill removed
+
+---
+
+### TC-007: User-global skill discovery
+
+**Type:** `Happy Path`
+**Priority:** `P1`
+
+#### Preconditions
+- Copilot chat is open in Agent mode
+- **Enable Skills** is checked
+
+#### Steps
+1. Create `~/.copilot/skills/global-skill/SKILL.md` with a name and description in YAML front matter
+2. Restart Eclipse (or wait for auto-refresh)
+3. In the Copilot chat input (Agent mode), type `/`
+
+#### Expected Result
+- `global-skill` appears in the slash command list, available regardless of which workspace is open
+
+#### 📸 Key Screenshots
+- [ ] **Slash command popup** — User-global skill visible in Agent mode
+
+---
+
+### TC-008: Skill prefix filtering in slash command autocomplete
+
+**Type:** `Happy Path`
+**Priority:** `P1`
+
+#### Preconditions
+- At least two skills are available in Agent mode (e.g., `my-skill` and `other-skill`)
+
+#### Steps
+1. In the Copilot chat input (Agent mode), type `/my`
+2. Observe the filtered slash command popup
+
+#### Expected Result
+- Only commands matching "my" are shown (e.g., `my-skill` visible, `other-skill` absent)
+
+#### 📸 Key Screenshots
+- [ ] **Filtered popup** — Autocomplete showing only matching skills for typed prefix
+
+---
+
+### TC-009: Skills settings persist across Eclipse restarts
+
+**Type:** `Regression`
+**Priority:** `P1`
+
+#### Preconditions
+- **Enable Skills** preference is visible
+
+#### Steps
+1. Open **Window → Preferences → Copilot → Chat** and uncheck **Enable Skills**
+2. Click **Apply and Close**
+3. Restart Eclipse
+4. Open **Window → Preferences → Copilot → Chat**
+
+#### Expected Result
+- **Enable Skills** remains unchecked after restart, confirming the preference is persisted
+
+---
+
+## Screenshots Checklist
+> Consolidated list of all key screenshot moments.
+
+- [ ] `TC-001` Chat preferences showing the Enable Skills toggle (checked)
+- [ ] `TC-002` Slash command popup with `/my-skill` visible in Agent mode
+- [ ] `TC-003` Custom prompt file visible as slash command in Agent mode
+- [ ] `TC-004` Ask mode popup without skill in slash command list
+- [ ] `TC-005` Enable Skills unchecked in preferences
+- [ ] `TC-005` Slash command popup without skill after disabling
+- [ ] `TC-006` Slash command popup before skill is added
+- [ ] `TC-006` Slash command popup after skill is added
+- [ ] `TC-006` Slash command popup after skill is deleted
+- [ ] `TC-007` User-global skill visible in Agent mode slash command popup
+- [ ] `TC-008` Filtered autocomplete showing only matching skills for typed prefix
diff --git a/test-plans/subagent-restoration/subagent-restoration.md b/test-plans/subagent-restoration/subagent-restoration.md
new file mode 100644
index 00000000..cbf0f70e
--- /dev/null
+++ b/test-plans/subagent-restoration/subagent-restoration.md
@@ -0,0 +1,113 @@
+# Subagent Turns Appear as Separate Assistant Messages After Restoration
+
+## Overview
+When a subagent executes during an agent-mode conversation, its turns were previously persisted as independent `CopilotTurnData` entries. On restoration from history, those turns rendered as standalone assistant messages instead of being nested inside the parent agent turn.
+
+---
+
+## Test Cases
+
+### TC-001: Subagent turn restores nested inside parent agent turn
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- Copilot chat is open in Agent mode
+- A prompt that reliably triggers subagent execution is available (e.g. a task that invokes `run_subagent`)
+
+#### Steps
+1. Send a message in Agent mode that causes a subagent to execute (e.g. "analyze all Java files in the workspace using a subagent")
+2. Wait for the subagent to complete and the full response to finish streaming
+3. Note the structure of the response: the subagent output should appear nested/indented inside the parent agent turn, **not** as a separate assistant message below it
+4. Open chat history (the history panel), select a different conversation or click "New Chat"
+5. Open chat history again and select the original conversation to restore it
+
+#### Expected Result
+- The restored conversation shows the subagent output nested inside the parent agent turn, matching the structure visible before the session switch
+- No additional standalone assistant message appears below the parent turn for the subagent execution
+
+#### 📸 Key Screenshots
+- [ ] **Before switch** — Full conversation view showing subagent output nested inside parent turn
+- [ ] **After restore** — Same conversation restored with subagent output still nested, no orphaned message
+
+---
+
+### TC-002: Multiple subagent turns all restore nested correctly
+
+**Type:** `Happy Path`
+**Priority:** `P0`
+
+#### Preconditions
+- Copilot chat is open in Agent mode
+- A prompt that triggers multiple sequential subagent invocations is available
+
+#### Steps
+1. Send a message that causes two or more subagent invocations in a single parent turn
+2. Wait for all subagents to complete
+3. Verify each subagent's output is shown nested inside the parent turn (no orphaned messages)
+4. Switch to a different conversation via chat history, then switch back to restore the original
+
+#### Expected Result
+- All subagent turns restore nested inside the parent agent turn
+- The number of visible assistant messages in the turn matches what was shown before the session switch
+- No orphaned standalone assistant messages appear for any subagent
+
+#### 📸 Key Screenshots
+- [ ] **Before switch** — Multiple subagent outputs nested inside one parent turn
+- [ ] **After restore** — All subagent outputs still nested, none appear as separate messages
+
+---
+
+### TC-003: Conversation with no subagent is unaffected
+
+**Type:** `Regression`
+**Priority:** `P0`
+
+#### Preconditions
+- Copilot chat is open in Agent mode
+
+#### Steps
+1. Send a message that produces a normal (non-subagent) agent response
+2. Wait for the response to complete
+3. Switch to a different conversation via chat history, then switch back to restore the original
+
+#### Expected Result
+- The restored conversation looks identical to before the session switch
+- No duplicate or extra assistant messages appear
+
+---
+
+### TC-004: Partially-streamed subagent turn (cancelled) restores correctly
+
+**Type:** `Edge Case`
+**Priority:** `P1`
+
+#### Preconditions
+- Copilot chat is open in Agent mode
+
+#### Steps
+1. Send a message that triggers a subagent execution
+2. While the subagent is still streaming (before it completes), click the Cancel button
+3. Observe the state of the conversation — the parent turn and any partial subagent output
+4. Switch to a different conversation via chat history, then switch back
+
+#### Expected Result
+- The restored conversation reflects the cancelled state
+- The partial subagent output (if any) is shown nested inside the parent turn, not as a separate assistant message
+
+#### 📸 Key Screenshots
+- [ ] **After cancel** — Conversation showing cancelled subagent output nested in parent turn
+- [ ] **After restore** — Same nested structure restored after session switch
+
+---
+
+## Screenshots Checklist
+> Consolidated list of all key screenshot moments.
+
+- [ ] `TC-001` Subagent nested in parent turn before switch
+- [ ] `TC-001` Subagent nested in parent turn after restore
+- [ ] `TC-002` Multiple subagents nested before switch
+- [ ] `TC-002` Multiple subagents nested after restore
+- [ ] `TC-004` Cancelled subagent nested after cancel
+- [ ] `TC-004` Cancelled subagent nested after restore