Why opencode/LLM still uses "Read" instead of codebase memory?

[DenisBY]

My opencode/LLM still continues to use "Read ..." instead of asking codebase memory. Why and how to fix it?

[DeusData]

Hey, thanks for bringing this up. I will dive into this more when I am back on my desk next week. In general we have seen the pattern, that Agents use there default behavior which is kinda hardcoded into themselves more than respecting additionally defined behvaior/skills. This also depends largely on which underlying LLMs you are using.

I will give a broader headsup to this when I am back :)

[DenisBY]

Yeah, I found that Opus constantly ignores rules set in AGENTS.md and during a session, but GPT-5.4 follows them. Sonnet told me, that Opus thinks that it smart enough to do everything on its own and therefore ignores the rules :D.

[esafak]

@DeusData Why does OpenCode only have AGENTS support while others have hooks? For OC the analog would be a plugin. For something that you want used often, AGENTS directives are too weak. Want me to open an issue?

[DeusData]

There is no clear "why". It is just a matter of "I did not have time to do it yet". Remember this project is still a couple months old and was built in my free time. In case you are interested of improving this, feel free to engage through an contribution :)

[DeusData]

Thank you for the earlier reports, and sorry the follow-up took so long. Discussion #2025 now has more recent observations on tool selection and hook delivery, including the distinction between graph context reaching the agent and visible MCP calls. My June statement about OpenCode integration is historical, not a current compatibility check. If this still affects you, the current client/model and CBM versions plus one example task will be more useful than counting Read calls alone.