fix(clients): normalize OpenAI-wire messages for Ollama native /api/chat

[antoinezambelli]

What

Fixes two Ollama proxy bugs (#111, #115) of the same class. OllamaClient runs
native passthrough, so the client's verbatim OpenAI-wire messages reach
Ollama's native /api/chat, whose schema is stricter than the OpenAI wire format:

• Ollama client doesn't handle messages.content as an array #115 — multi-part array content → 400 cannot unmarshal array ... of type string
• Ollama backend: native raw-message passthrough sends tool_call arguments as JSON strings → 400 on multi-turn tool sessions #111 — assistant tool_calls[].function.arguments as a JSON string →
  400 Value looks like object, but can't find closing '}' (every multi-turn
  tool session, from the 2nd turn on)

How

An Ollama-local outbound normalizer applied in both send and send_stream
before building the request body:

• flatten multi-part content arrays to text (text-only — images dropped,
  consistent with the existing converted path; full multi-modal support is Add multi-part / image content support (text-only today; proxy↔WorkflowRunner parity) #116)
• coerce tool_calls[].function.arguments from JSON string to dict, only when
  it decodes to an object (malformed/non-object payloads pass through untouched)

Native /api/chat is kept — switching to Ollama's /v1 compat shim would force
a rewrite of the native response parsing (thinking field, dict args, NDJSON).

The content-flatten primitive is extracted to clients/base.py and reused in
proxy/convert.py (behavior-preserving); base.decode_tool_args is reused for
argument coercion.

Testing

• Unit: normalizer + on-the-wire body for send/send_stream; arg edge cases,
  image-drop, no-mutation.
• Integration: proxy → real OllamaClient (mocked transport), asserting the
  raw-passthrough first-attempt body — the seam that let both bugs ship.
• Full suite: 1178 passed.
• Live end-to-end against a real Ollama server: both raw shapes 400 today; both
  succeed through forge's normalizer.

Known limitation

Images in multi-part content are dropped (text-only), matching prior behavior.
Real multi-modal support is tracked separately in #116.

Closes #111
Closes #115

🤖 Generated with Claude Code