First impressions

[Kangaroux]

Using 0.70.0

I tried crush when I looked for a harness written in Go. I wanted something snappy and was curious if there was anything I was missing. So I gave it a go with some basic tasks.

Here's a list of the feedback I have in no particular order. I'm coming from Opencode, which I generally enjoy from its clean docs and UI, so some of the feedback will be biased towards that.

This is just my personal "needs work" criticism. Crush feels like it's on the right track.

• ctrl-d to close
• ctrl-c to clear (clears if there is text, pops up with quit modal if empty?)
• context tracking is broken (shows up as -9223372036854775808%. using llama.cpp locally. would like to see the raw token count as well)
• random text "spinner" is extremely distracting especially for someone with ADHD. would like an option for reduced animations
• better visual separation between subagent and primary. i sorta like having it displayed inline, but i want to clearly know where it starts and stops. different colored background and some horizontal lines would be nice
• would like to see context/execution time for subagents
• invoking skills are confusing. i miss having a /skills command to just see a list of them. (related: while asking the LLM about crush it told me i could say crush_stats. I think this is invoking a skill? where is this coming from?)
• option to hide grep/view results (knowing what it's grepping/viewing is fine, i don't care about the contents).
• can't undo messages (!!). this one is a big deal breaker. in opencode if i don't like the direction the llm is headed i can just undo and edit my msg. or sometimes i say something that i want to take back
• hardcoded and very opinionated system prompt. including things like tool calling instructions is fine. but don't assume what users want. for example **BE AUTONOMOUS**: Don't ask questions. this is definitely NOT something I want in my system prompt
• no global AGENTS.md support by default (~/.config/crush/AGENTS.md. had to add ~/.config/crush/AGENTS.md to context_paths)
• no onboarding route for local llm users (wasted time trying to find a custom provider option in the initial setup, eventually ctrl+f the readme for "ollama" and found what i needed).
• docs are severely lacking
  • no config reference
  • install cmds are strange (written in Go but npm install is the 2nd option?? this immediately put me off. binaries link feels buried)
  • would love a concise quickstart guide.
  • table of contents would help
  • consider splitting the docs into multiple pages (doc viewers with the sidebar for major topics are great)

[dcu]

I just found this, recently created PRs to fix some of these issues. but I'm curious how would the undo messages work?

[Kangaroux]

In Opencode, every message you send is like a checkpoint for the context. If I send a msg, the agent responds, and I realize it went off track, I would use undo. It would remove the agent and my msg from the chat history, and behave as though that part of the conversation never happened

Sometimes you can just tell the llm "sorry ignore that last message" but it still bloats and potentially poisons the context.

OC takes it a step further by allowing redos and even snapshotting the git changes/undoing them but that adds a ton of complexity to the feature + it's sometimes undesirable