Improving support for rich interactive CLIs in Copilot CLI - AZD

[rajeshkamal5050]

We are seeing several blocking usability issues when running a rich, interactive CLI through Copilot CLI.

Below are the some of the issues which we encountered when using azd CLI,

• Rich text and progress updates are garbled and hard to follow.

Copilot CLI

Native Experience

• Selection prompts/lists are not scrollable and only show 2–3 items at a time.

Copilot CLI

Native Experience

• Shell input/output appears to time out, even though work is still actively progressing in the background.

Copilot CLI

@ellismg @EvanBoyle can you help prioritize this?

@kristenwomack @spboyer @wbreza @tg-msft @Petermarcu

[ellismg]

I looked into what's going on here, and the problem is an unofrtunate interaction between the shell tool that Copilot uses, ANSI escape characters, and how the survey package used by azd for reading inputs works.

Right now when running programs that send ANSI escape sequences, our shell tool just disregards (instead of interpreting) the ANSI sequences. This is why you see garbled output above, for example, when rendering the progress bar azd is sending escape sequences to move the cursor, delete lines, and then rewrite output so it can create an animation. Since our shell tool drops all of these sequences when collecting output, you end up with garbled output.

But things get even worse, when azd goes to read input like the environment name, behind the scenes the survey pacakge ends up sending some ANSI DSR escape squences to both understand the size of the terminal and where the cursor is. From reading the survey source, I believe it does this because it wants to handle rendering the input as it comes in, so it needs to understand where the curosr is and how long the lines are so it can correctly "wrap" output as the user types it. Since our tool ignores all these escape sequences, we don't give back the answers that survey is looking for. Instead the model uses the write_bash tool to attempt to type an environment name, but this input is now unexpected by survey (it's still waiting for the answer to it's DSR sequences) and so the input is not correctly captured or echoed. The agent then gets confused and is unable to proceed.

The right fix here is for us to improve our terminal rendering, ideally leveraging something like @xterm/headless, but doing so is a bit of a lift on our side because it's really changing how the tool functions, so we need to take care in adopting this. A middle ground would be to interpret some subset of these ANSI escape sequences oursleves, but I'm worried that in the limit we'd just end up rebuilding a worse version of something like @xterm/headless, and so going down that path may end up costing us more with the bug tail than just thinking hard and adopting @xterm/headless.

[ellismg]

Also, an important thing for me to remember when we make improvments here and want to test with azd. The azd team recently landed this change: Azure/azure-dev#6633 to behave as if --no-prompt was passed when the CLI is run from copilot. So we need to make sure that when testing this we don't trigger this code path (this shipped in 1.23.3, so we need to use 1.23.2 or earlier to reproduce these issues, I found the best way to install this older version was to use the install-azd.sh script with --version parameter.