feat(provider): crof.ai / kimi-k2.5 compatibility and non-streaming fallback fix

[thomwebb]

Problem

crof.ai's kimi-k2.5-lightning endpoint returned 500 errors and silently dropped agent responses when used through code-puppy. Three distinct issues were identified:

1. strict: true causes 500 on non-streaming requests

crof.ai's non-streaming endpoint rejects strict: true in tool function schemas with a 500 Failed to generate response. Only one tool (load_image_for_analysis) had strict: true set by pydantic-ai, but that was enough to kill every request.

	strict=false	strict=true
stream=False	✅	❌ 500
stream=True	✅	✅

2. Flaky SSE streaming

crof.ai's streaming transport intermittently fails with "Streamed response ended without content or tool calls" in the real runtime (DBOS + TUI + event handler). Raw HTTP tests pass, but the full CLI stack hits timing issues. Disabling streaming via streaming: false in models.json provides a stable single-shot path.

3. Silent agent output when streaming is disabled

When streaming: false is set for a model, the runtime correctly disables the stream handler — but DBOS's skip_fallback_render callback still returned True, preventing the non-streaming fallback renderer from running. Nobody printed the output.

Changes

code_puppy/model_factory.py

• Add _CompatChatModel (subclass of OpenAIChatModel) for providers needing compatibility shims. Activated when strict_tools: false or provider: "crof" is set in models.json.
• _map_tool_definition: strips strict key from tool schemas when strict_tools: false.
• _map_model_response: flattens tool calls into assistant text for providers that can't handle tool_calls in assistant messages.
• _map_user_message: degrades role=tool results into role=user messages for providers that reject tool-role messages.
• Only send parallel_tool_calls setting when the provider/model advertises support (prevents 500 on unknown fields).

code_puppy/agents/_runtime.py

• Add _model_allows_streaming(): reads streaming from models.json config. Models with "streaming": false skip the SSE path entirely.
• When streaming is disabled, clear _event_stream_handler on the pydantic agent so pydantic-ai uses model.request() instead of request_stream(). Restore after the run.
• Remove the 📡 Streaming disabled info emission (noisy, not actionable).

code_puppy/plugins/dbos_durable_exec/runtime.py

• skip_fallback_render() now checks whether streaming is actually active for the current model. When streaming is disabled, returns False so the core non-streaming fallback renderer can produce output. Previously it returned True whenever DBOS was launched, causing silent agent responses.

code_puppy/models.json

• Add streaming: false and strict_tools: false for crof-kimi-k2.5-lightning.

Tests

• tests/agents/test_streaming_retry.py: new test class TestModelAllowsStreaming covering _model_allows_streaming() for None, crof-kimi, and normal models. New class TestNonStreamingHandlerClear verifying _event_stream_handler is nuked on DBOSAgent-like objects.
• tests/test_crof_tool_result_compat.py: tests for _CompatChatModel behavior (strict stripping, tool result mapping, tool call flattening).
• tests/test_model_factory_coverage.py: updated to verify parallel_tool_calls is conditionally included based on provider support.

Testing

All changes verified with 1700+ passing tests and live e2e requests against crof.ai:

• strict_tools=false → no 500s across 10+ consecutive requests
• streaming=false + DBOS → fallback render fires, agent output visible
• Tool calls (create_file, list_files) complete and produce text responses
• parallel_tool_calls not sent for crof models