Ollama backend: native raw-message passthrough sends tool_call arguments as JSON strings → 400 on multi-turn tool sessions

[imwldev]

Summary

With --backend ollama (native capability), forge forwards the verbatim OpenAI
transcript to Ollama. Assistant history messages carry
tool_calls[].function.arguments as a JSON string (OpenAI wire format), but
Ollama's /api/chat requires them as an object (dict). Every multi-turn
tool session fails on the 2nd+ turn with:

400 {"error":"Value looks like object, but can't find closing '}' symbol"}

First turn succeeds (no tool history yet); as soon as an assistant tool_calls
turn enters the history, all subsequent requests 400. Symptom from the client
(opencode) side: the proxy returns a non-OpenAI error chunk and the client's
schema validation crashes.

Minimal repro (directly against Ollama, bypassing forge)

# args as STRING  -> 400
curl -s http://localhost:11434/api/chat -d '{"model":"qwen3-coder:30b","stream":false,
 "messages":[{"role":"user","content":"x"},
  {"role":"assistant","tool_calls":[{"function":{"name":"t","arguments":"{\"x\":\"1\"}"}}]},
  {"role":"tool","content":"ok"}],
 "tools":[{"type":"function","function":{"name":"t","parameters":{"type":"object","properties":{"x":{"type":"string"}}}}}]}'

# args as DICT    -> 200   (same request, arguments as object)

Root cause

• proxy/handler.py: in native_passthrough, raw_messages_for_backend = _raw_openai_messages(request_messages) forwards arguments unchanged.
• clients/ollama.py send() / send_stream() post these messages to /api/chat without coercing arguments str→dict.
• proxy/convert.py already does this coercion on the inbound path (if isinstance(args, str): args = json.loads(args)), but the raw-passthrough path bypasses it.

Suggested fix

Coerce tool_calls[].function.arguments from str→dict for the Ollama backend
(e.g. in OllamaClient.send/send_stream before building the request body, or
suppress raw passthrough for Ollama so the normal format="ollama" serialization
is used).

Env

forge-guardrails 0.7.5, Ollama 0.30.8, model qwen3-coder (custom), client: opencode (OpenAI protocol, streaming).

--- Above message is written by Opus4.8 and it fixed my local env automatically and confirm working. ---
I with this info will help to improve this project.