Feature request: Support `#tool:todo` and `#tool:vscode/askQuestions` when delegating session to Copilot CLI

[akilin]

Feature request

Make #tool:todo and #tool:vscode/askQuestions available when a Copilot Chat session target is set to Copilot CLI.

I find these tools very useful, as they greatly improve UX when working with Copilot in VS Code.
I have been using both of them a lot when writing custom agents/prompts/skills.
If I want to run multiple sessions at the same time by delegating sessions to Copilot CLI (in a separate git worktree) - those tools are not available. And my agents/prompts/skills stop working.

Usually, this leads to the model rejecting the task, stating that the requested tools are not available.
Sometimes the model attempts to replicate the tools by asking questions in text form and keeping an internal todo list.
Both result in a greatly degraded user experience.

I feel this change would make Copilot CLI a more complete session target for agentic work inside VS Code.
It would also improve consistency for custom agents/prompts/skills that rely on these tools.

[bamurtaugh]

Related (we could perhaps track here or in a follow-up, @DonJayamanne @meganrogge please let me know what you'd prefer) - it'd be great for CLI sub-agents to have the ability to use the ask tool.

Here's some further insight, drafted with Copilot:

The Copilot CLI's task tool (used to spawn sub-agents within a CLI session) also doesn't expose ask_user / vscode/askQuestions to those sub-agents. So even when a CLI session itself can ask the user clarifying questions at the top level, anything it delegates to a sub-agent loses that capability and has to either (a) embed the question in the sub-agent's final report and have the parent re-launch with clarified context, or (b) make assumptions and hope.

This is symmetric to the issue VS Code's native runSubagent had (#294949), which was resolved so that VS Code sub-agents can use vscode/askQuestions. Bringing the CLI path to parity — both at the top level (this issue) and inside task-spawned sub-agents — would let custom agents/prompts/skills work consistently regardless of session target.

[akilin]

I will add additional details, as I am observing some strange behaviour with a new scenario I tested.
This might not be a feature request, but a bug ticket.

I initially tested 2 scenarios, using a custom agent that explicitly references #tool:todo and #tool:vscode/askQuestions in its markdown file (full agent file can be found at the end of message in text form).

1. custom agent + local execution = everything great
2. custom agent + delegation to Copilot CLI = failure

Scenario 2 failing is what led me to believe that Copilot CLI does not have access to tools, and why I created the feature request originally.

I have recently tested a new scenario that shows that Copilot CLI does have access to those tools when you use the default agent. Which makes me think this is possibly a bug.
New scenario: use the default agent, set delegation to Copilot CLI, reuse the main body from the custom agent as a prompt for the default agent.
Result: everything works as expected.

Scenario 1

I have a custom agent that has specific instructions inside of its .md file that tell it to use the vscode/askQuestions tool to gather 3 tasks and todo tool to store the gathered tasks.
I select this agent from the dropdown, set session to local and typed go as a message.
Result: everything works perfectly.

Scenario 2

Same agent, same prompt as Scenario 1, but execution is delegated to Copilot CLI.
Result: tools are not available, nicely integrated UX is broken, and the model reverts to asking questions in text form

Scenaio 3

Default agent (not custom), execution delegate to Copilot CLI.
Prompt: Copied main part of body from custom agent. Explicit references to tools to use.
Result: works great, session delegated to Copilot CLI, and it can use askQuestions tool and todo tool.

Custom agent used

Below is a text for the custom agent used in my tests.
I have also attempted adding tools: ['vscode/askQuestions', 'todo'] to frontmatter, but it had no effect on results (agent worked fine in local mode, without this. so i did nto think the tools are the issue. but decided to check just in case).

---
name: task-collector
description: "Use when user wants to input tasks for you to complete. Collect exactly three tasks from the user and write them to #tool:todo."
argument-hint: "Run me"
user-invocable: true

handoffs: 
- label: Complete tasks
  agent: task-distributor
  prompt: 'Go'
  send: true
---

You are a focused task intake agent.
Your job is to collect exactly three tasks from the user via #tool:vscode/askQuestions and write them to #tool:todo .

## Workflow
1. Ask the user for three tasks.
2. Validate that each answer contains a usable task title.
3. Write the tasks to #tool:todo

[akilin]

Scenario 1 confirms that custom agents can use both tools (vscode/askQuestions & todo).
Scenario 3 confirms that sessions delegated to Copilot CLI can use both tools.
But when you combine the custom agent and Copilot CLI delegated execution (scenario 2) - things break.

[akilin]

I am not sure how ticket type conversions are usually handled in this repo, so I re-created this feature request as a bug, just in case:
#314010

Feel free to keep whichever you feel is better suited, and close the other one.

[DonJayamanne]

Thank you for filing this issue, this will be a tricky one to address, but a valid one.
Claude and Copilot CLI have their own set of tools, and are not necessarily the same name as those in VS Code,
All three are completely different agent harnesses (agent run times), with their own set of tools
& thats what you are running into.

If you didn't use explicity names like #tool:vscode/askQuestions, and used a more lose term like ask user question tool, that would work.

however thats merely a work around,
I'll see what we can do about this.

[akilin]

All three are completely different agent harnesses (agent run times), with their own set of tools
& thats what you are running into.
If you didn't use explicity names like #tool:vscode/askQuestions, and used a more lose term like ask user question tool, that would work.

I am not 100% sure that this is what is happening.
If you look at scenario 3, where I take the custom agents body and paste it as a prompt message - it works.
So exactly the same set of instructions, when executed by Copilot CLI works perfectly if provided as chat message for the default agent, but fails if provided as part of custom agent definition.

It seems this is not about tools not being available or having different names (or at least not only about this).
In logs I do see, that internal tools are named differently.
Given the prompt at the end of my message:

• local execution will use tool vscode/askQuestions and todo.
• Copilot CLI when using default agent, given exactly same prompt will use ask_user and sql for storing the todo list. While the tools are different, we still get a nice UI component when asking questions + it still shows a UI for task list added to todo list.
• Copilot CLI when given the same set of instructions via custom agents body will fail to use any of the tools above. No additional UI displayed for questions or task list, just text.

So while the tools are named differently, the Copilot CLI is still able to connect the dots and use the correct tools when instructions are given via prompt message, but fails when same instructions are provided via a custom agents body.

Prompt:

You are a focused task intake agent.
Your job is to collect exactly three tasks from the user via #tool:vscode/askQuestions and write them to #tool:todo .

[DonJayamanne]

oh got it thank you for the clarification. will have another look