[nicksosnovsky]

Problem

Every request fails with litellm.BadRequestError: OpenAIException - System message must be at the beginning (HTTP 400) when using OpenAI-compatible proxies backed by litellm/vLLM (e.g. qwen via a litellm proxy).

Root cause

The experimental.chat.system.transform hook appended the memory/goal/rules block as a second system message:

if (block) output.system.push(block)

opencode sends every element of output.system as a separate {role: "system"} message. Some servers reject more than one system message even when they are all at the beginning.

Repro

curl https://api-llm.novacom.by/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "qwen3.6-27b",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "system", "content": "Extra block."},
    {"role": "user", "content": "hi"}
  ]
}'

-> 400 System message must be at the beginning. Removing the second system message fixes it.

Fix

Append the block to the existing first system message instead of pushing a new one:

if (output.system.length > 0) output.system[0] += "\n\n" + block
else output.system.push(block)

Verification

• Direct API call with the merged system message succeeds.
• bun test bulba.test.ts: 45 pass / 1 fail (the failing search_memory test also fails on clean master - Windows environment, unrelated to this change).
• The test suite calls the hook with an empty system: [] array, which still works via the push fallback.

[sehaxe]

Reviewed against opencode source + the live plugin.

Verdict: correct diagnosis, correct fix — merged (adapted).

Confirmed the root cause in opencode's session/llm/request.ts: the system prompt starts as system = [header], and after the transform hook opencode only coalesces when system.length > 2. Pushing a 2nd element yields exactly two system messages, which litellm/vLLM reject. Your curl repro against api-llm.novacom.by matches this exactly.

Applied to master with two changes:

1. Reconciled a conflict — the branch was based on an older master. It used systemBlock() with no arg, which would have reverted the sessionID threading (systemBlock(input?.sessionID)) that the role-resume fix depends on. Kept the sessionID arg, applied your merge logic on top.
2. Added the missing test — all existing transform tests initialize output.system = [], so they exercised the push fallback and never the merge path. Added a test asserting system.length === 1 when starting from a non-empty header.

Minor note (no action): the else output.system.push(block) branch is dead in production — opencode always passes a non-empty system (it holds the header). It's kept only so tests with system: [] still exercise the injection. Fine to keep.

60 tests green. Thank you for the fix — it directly affects the corp LiteLLM proxy.

[sehaxe]

Merged to master (adapted, with sessionID conflict reconciled + test added). Thanks!