From c6f172cbdc23e1c481ed1b29856a313cadd3ebcc Mon Sep 17 00:00:00 2001 From: krypticmouse Date: Fri, 29 May 2026 02:59:35 +0000 Subject: [PATCH] docs(agents): add eval-backed model-capability guidance for OpenCodeAgent A 27B local model (Qwen3.5-27B via vLLM) passed a 7-task coding suite cleanly (create/edit/bug-fix/implement-to-pass-tests/multi-file, verified by running code + pytest); an 8B model was unreliable. Document so users pick a capable model for real coding work. Co-Authored-By: Claude Opus 4.7 (1M context) --- docs/user-guide/agents.md | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/docs/user-guide/agents.md b/docs/user-guide/agents.md index 8cec9dd6..f56b590a 100644 --- a/docs/user-guide/agents.md +++ b/docs/user-guide/agents.md @@ -430,6 +430,16 @@ jarvis ask --agent opencode "Refactor the parser to use a state machine" !!! tip "Pass-through providers" If the `engine` has no derivable base URL, pass `model` as `provider/model` (e.g. `ollama/llama3`) and opencode resolves it from its own configuration — no `opencode.json` is written. +!!! warning "Model capability matters" + opencode's agentic loop (planning + correct tool calls + multi-step + follow-through) needs a reasonably capable model. In testing, a **27B** + local model (Qwen3.5-27B served via vLLM) solved a 7-task coding suite + cleanly (create / edit / bug-fix / implement-to-pass-tests / multi-file, + verified by running the code and tests). An **8B** model (qwen3:8b) was + unreliable — malformed tool calls, syntactically broken code, and + half-finished tasks. Prefer a capable local model (or a cloud model) for + real coding work. + --- ## OperativeAgent