I'm having mild success with how llamacpp works with thinking models. It's implemented `reasoning-budget` and `reasoning-budget-message`; I've paired the message with the agent, along with having a specific name for the agent. Not just "you're a researcher" but "your name is <ai agent>".
Using opencode then, this is the primary agent, and I've added dynamic context pruning plugin; so when the Qwen3.6-35B-A3B model starts getting confused as evidenced by the length of the reasoning-budget cut off, it injects a message like `<ai agent>, you've been thinking too much, compress the context and use a subagent to figure out the quandry`. So that gets injected every time the budget exceeds.
About half the time it works in either of the options, which is compressing/pruning to reduce the poisoned context or forming a specific subagent description of the problem and letting the subagent, hopefullly, work on a smalled scoped issue.
I think the next step is some kind of omniscient markdown file/folder combo which can be used to survive some of the microcosmic issues with the dumber MoE slices.
I'm having mild success with how llamacpp works with thinking models. It's implemented `reasoning-budget` and `reasoning-budget-message`; I've paired the message with the agent, along with having a specific name for the agent. Not just "you're a researcher" but "your name is <ai agent>".
Using opencode then, this is the primary agent, and I've added dynamic context pruning plugin; so when the Qwen3.6-35B-A3B model starts getting confused as evidenced by the length of the reasoning-budget cut off, it injects a message like `<ai agent>, you've been thinking too much, compress the context and use a subagent to figure out the quandry`. So that gets injected every time the budget exceeds.
About half the time it works in either of the options, which is compressing/pruning to reduce the poisoned context or forming a specific subagent description of the problem and letting the subagent, hopefullly, work on a smalled scoped issue.
I think the next step is some kind of omniscient markdown file/folder combo which can be used to survive some of the microcosmic issues with the dumber MoE slices.