fix(provider): consolidate leading system messages to prevent Jinja crash on compaction

[adi-IL]

Fixes #11

What was wrong

When running with OpenAI-compatible endpoints or local models (Ollama, vLLM, Qwen, M.O.G.-SEC, Mistral) and reaching auto-compaction or turns with volatile engagement state, the session failed permanently with an HTTP 500 Jinja template error:

Error: Jinja
Exception: System message must be at the beginning.

Root Cause

1. In packages/opencode/src/session/llm/request.ts, prepare appends volatile engagement context as a secondary system entry (system.push(input.volatileSystem)). When messages is constructed via ...system.map((x) => ({ role: "system", content: x })), multiple discrete { role: "system" } messages are emitted at the head of the message array.
2. Standard Jinja chat templates (e.g., Qwen, Llama, Mistral) iterate over messages and enforce that role == 'system' is only valid on loop.first (messages[0]). When the loop evaluates messages[1] (role == 'system'), not loop.first evaluates to true, raising System message must be at the beginning.
3. Similarly, if any chronological system update appeared later in the conversation history, it also triggered this template assertion.

What changed

1. System Message Normalization (packages/opencode/src/provider/transform.ts):

   • In normalizeMessages, all contiguous leading role: "system" messages are consolidated into a single leading role: "system" message (joining contents with \n\n) for OpenAI-compatible, Ollama, and non-Anthropic providers.
   • For Anthropic models (@ai-sdk/anthropic, @ai-sdk/google-vertex/anthropic), distinct leading system blocks are preserved so prompt caching continues to hit.
   • Any non-leading role: "system" messages appearing after user/assistant turns in history are lowered into user-compatible <system-update>\n...\n</system-update> blocks, preserving chronological context without violating chat template role ordering.

2. Middleware Transform Support (packages/opencode/src/session/llm.ts):

   • Updated transformParams in wrapLanguageModel to run ProviderTransform.message for both stream and generate request types.

3. Unit Tests (packages/opencode/test/provider/transform.test.ts):

   • Added unit tests for:
     • Consolidating multiple leading system messages for OpenAI-compatible models.
     • Lowering non-leading system messages to <system-update> blocks.
     • Preserving separate leading system blocks on Anthropic for prompt caching.
     • Single leading system messages and messages without system prompts.

Verification

• bun test test/provider/transform.test.ts: 295 passed (all new tests passing).
• bun test test/session/compaction.test.ts: 52 passed.
• bun turbo typecheck: 19/19 packages passed cleanly with zero type errors.