CLI prompt argument's quotes/newlines get corrupted before being sent as the user message to the LLM provider

Description

The CLI prompt argument (opencode run "...") is corrupted before being sent as the user message to the LLM provider, when it contains internal double-quotes and/or newlines. This appears to reliably break tool-calling for at least some local (Ollama, OpenAI-compatible) models, and may be a contributing factor in other reports of "model outputs tool calls as text instead of executing them" (e.g. #33959, #37543, #21181, #29757) — those reports show the same class of symptom, though I haven't verified they share this exact root cause.

Captured the real outgoing wire request via a transparent logging proxy in front of an OpenAI-compatible Ollama provider. The user message content opencode actually transmits is wrapped in a literal leading/trailing ", with every internal " backslash-escaped — but embedded newlines are left as raw literal newlines, not escaped as \n. This isn't valid JSON string content (real JSON.stringify would also escape \n); it looks like a hand-rolled "escape quotes for embedding" step that's missing newline handling. The system message (agent prompt config) in the same request was NOT corrupted — this is specific to how the CLI prompt argument becomes the user-turn message, not a general serialization bug.

Causal verification

Built matched payloads (real captured system prompt, same tool schema, temperature: 0.15, stream: false) and sent them directly to Ollama's native /api/chat, varying only the user message content:

Every real opencode run dispatch against this same prompt failed outright; the exact same prompt text sent uncorrupted via raw curl/direct API succeeds reliably. This also explains why "the model works fine when I test the API directly, but not through opencode" shows up in several other issues in this repo — testing the API directly naturally uses clean, uncorrupted prompt text.

Steps to reproduce

opencode run --agent <any> "text with \"internal quotes\" and
newlines"

Capture the outgoing request to the provider (e.g. via a logging reverse proxy in front of an OpenAI-compatible baseURL) and inspect the user message's content field — it will contain a stray wrapping ", escaped internal quotes, but raw unescaped newlines.

OpenCode version

1.18.18 (compiled Bun binary — could not locate readable source locally to pinpoint the exact code path; this may be in the CLI arg → message-part construction, or in provider message serialization)

Operating System

macOS