Relay liking the new context / compact management in this version. Had 3 hours session running without issue! super impressive. This was a killer for me before with some long horizon tasks I do with local Qwen 3.8 27b. Big, big! QoL for me.
Really enjoy using DSH and have switch over from OpenCode. I've struggled with long sessions and context collapse. Running local Qwen 3.8 27b NVFP4 with a context of 200k on a 5090. The backend llama.cpp is fine, switch session and llama.cpp keeps working as expected.
Is there anyway to keep a session after this happens?
Compaction could not produce a useful summary. The conversation is unchanged; the attempt is recorded in the session log.
My work around is to switch my model at the back down to kv q8 256k ctx, then use a normal prompt to do a summary i can use to pickup work in a new session. That gives me something I can then start a new context session with, but its a pain when you have long running prompt.
This is an music track I wrote then used AI Tools to render into music. Its inspired the 86 Movie one of my vivid childhood memory's. I read rules, but if this isn't appropriate for here I under.
Transformers the Movie 1986 G1 Music The Battle for Autobot City (Metroplex).
Track 3 - City Battle ('86 Metal Female Vocialist) [AI] [Transformers] [G1]
The biggest issue I'm finding with LLMs is bad output. The OpenCode harness AGENT.MD file is limited to just LLM output/thinking and if the model makes a mistake OpenCode plays along. I'd like a hybrid approached without having write my own Python UI/Backend as it would take me at least 30 minutes using AI!
The output of the LLM is like variables, then OpenCode harness can run logical operations on those with Classic Code control: IF/THEN/ELSE/WHILE/GOTO/<CallFunction>... etc
The output and the decision would still be done by LLM, but the OpenCode would ensure an IF This = That or IF This != That. etc. Rules are 100% followed as its outside of LLM thinking. Its like doing a sub-agent call to ask the LLM a question and it decides an answer but we have binary control over what happens next with the Thinking part to ensure it follow rules to the letter. AGENT.MD files are more guild lines if you feel like it little AI.
I love opencode, its amazing stuff. But also I want to be able to show the context I'm using an image of the problem for a UI element. What are people doing to work around that type of issue. I normally switch into something like OpenWebUI get it to text out the image and try to feed that into the OpenCode context. but its an annoying workflow
Done 4, raids 2 yesterday and 2 today. No Rogues on roofs. Is something changed? I know Icebreaker has a ton. But that's a one way trip and would sooner not waste a ton of cash just to get head-eyes by Knight.
Since the event not been able to find any and need them for Lightkeeper & Ragman questlines.