Repository navigation
Impossible to edit Tasks prompt while waiting for the workspace to spin up #1084
Description
Imagine a scenario where the user makes a silly error in the original prompt, but clicks the "start task" ⬆ button. If that mistake was made the user will have to wait up to several minutes for the workspace to be created with no way of editing the prompt.
In addition, the original prompt is not visible on screen while the workspace is being spun up:
I think it would be super useful to show the prompt window there and say something like "you may still fine tune your prompt while the workspace is being started". This could save me a ton of waiting time.
Activity
johnstcn commentedon Oct 21, 2025
This is going to be tricky. At this point, the prompt is essentially set in stone. We would have to modify where the prompt comes from -- instead of from the provisioner, it would have to come from the Coder API (maybe from a new CLI command
coder exp task prompt "${TASK_ID}"?)mtojek commentedon Oct 22, 2025
If you look at how ChatGPT works, there is "Stop" button which lets you send another prompt. Would that be an option? I mean, start another tasks and kill the old one in the background.
DanielleMaywood commentedon Oct 22, 2025
I'm unable to give a complexity to this issue as the complexity depends on the solution we're looking at here. Giving some nice UI way of stopping, changing the prompt, and restarting, would probably be a 1 but changing in-flight would be way larger.
johnstcn commentedon Oct 22, 2025
This would be much simpler. 👍
bartekgatzcoder commentedon Oct 22, 2025
Are the operations issued through a "send" command somehow queued on our side? Could we check the queue for any changes in the precise moment when the Agentinc module is kicked off?
DanielleMaywood commentedon Oct 22, 2025
@bartekgatzcoder The prompt is made available to the workspace in terraform. By the time you've started a workspace build it is too late to change it.
bartekgatzcoder commentedon Oct 22, 2025
But that is exactly my point. Could Agentic AI pick the prompt from the queue, instead of the terraform? If the initial prompt was also queued (Assuming there is a queue behind the "send" action), this could add the desired flexibility of editing the prompt / overwriting it while the workspace starts.
johnstcn commentedon Oct 22, 2025
It would need fundamental changes to how this is all wired together and would likely need an RFC, or at least a proof-of-concept.
mtojek commentedon Oct 23, 2025
@bartekgatzcoder That queue behind send are
provisioner_jobs, and under proper setup, that queuing should last a couple of seconds, unless provisioners are busy. I'm not convinced if this is worth it.Actually, this is still my favourite option.
DanielleMaywood commentedon Oct 30, 2025
The following information is not strictly relevant to this issue, but useful context:
During the bug bash yesterday, we discovered a limitation (at least on Linux) with how we
currently pass prompts through to the agent.
On linux, environment variables are length-limited to
(PAGE_SIZE * 32)1. Typically the page size is 4KiB, meaning an environment variable is limited to 128KiB of data in a typical setup.We might investigate an alternative method of providing the task to a task workspace given this limitation, and this issue could benefit from that.
Footnotes
https://github.com/torvalds/linux/blob/e53642b87a4f4b03a8d7e5f8507fc3cd0c595ea6/include/uapi/linux/binfmts.h#L15 ↩
feat(coderd): add task prompt modification endpoint (#20811)
feat(coderd): add task prompt modification endpoint (#20811)
feat(site): allow modifying task prompts for starting tasks (#20812)
feat(site): allow modifying task prompts for starting tasks (coder#20812