From 805ddac61c608d9ec9acdb898880014239b40a34 Mon Sep 17 00:00:00 2001 From: zipg Date: Fri, 4 Sep 2026 16:45:28 +0800 Subject: [PATCH] fix(ai): send Agens-compatible follow-up tool-call requests --- crates/dbx-core/src/ai.rs | 126 +++++++++++++++++++++++++++++++------- 1 file changed, 103 insertions(+), 23 deletions(-) diff --git a/crates/dbx-core/src/ai.rs b/crates/dbx-core/src/ai.rs index 510d43f155..0a7263e8f9 100644 --- a/crates/dbx-core/src/ai.rs +++ b/crates/dbx-core/src/ai.rs @@ -777,6 +777,17 @@ fn ensure_anthropic_version_prefix(endpoint: &str) -> String { } } +/// Temporary Agens compatibility workaround. Agens' OpenAI-compatible gateway +/// currently uses a strict Responses input deserializer for agent follow-up +/// requests. Keep these tweaks scoped to that endpoint so other providers +/// retain the standard request shape. +/// +/// TODO(Agens): Remove `is_agens_endpoint` and all Agens-specific branches +/// below if api.agnes-ai.cn is discontinued or no longer needs this workaround. +fn is_agens_endpoint(config: &AiConfig) -> bool { + config.endpoint.to_ascii_lowercase().contains("agnes-ai.cn") +} + pub fn resolve_endpoint(config: &AiConfig) -> String { let ep = config.endpoint.trim().trim_end_matches('/'); if matches!(config.provider, AiProvider::Gemini) { @@ -1455,7 +1466,16 @@ pub fn build_responses_input(system_prompt: &str, messages: &[AiMessage]) -> ser json!(input) } +#[cfg(test)] fn build_responses_input_with_tools(system_prompt: &str, messages: &[AiMessage]) -> serde_json::Value { + build_responses_input_with_tools_variant(system_prompt, messages, false) +} + +fn build_responses_input_with_tools_variant( + system_prompt: &str, + messages: &[AiMessage], + include_function_call_metadata: bool, +) -> serde_json::Value { let mut input = Vec::new(); if !system_prompt.is_empty() { input.push(json!({ @@ -1482,12 +1502,20 @@ fn build_responses_input_with_tools(system_prompt: &str, messages: &[AiMessage]) })); } for tool_call in &message.tool_calls { - input.push(json!({ + let mut item = json!({ "type": "function_call", "call_id": tool_call.id, "name": tool_call.name, "arguments": tool_call.arguments.to_string(), - })); + }); + if include_function_call_metadata { + // OpenAI commonly returns these fields on output items; + // Agens' strict input enum requires them when replaying a + // function call in the next turn. + item["id"] = json!(format!("fc_{}", tool_call.id)); + item["status"] = json!("completed"); + } + input.push(item); } continue; } @@ -2179,9 +2207,17 @@ fn build_openai_chat_messages( system_prompt: &str, messages: &[AiMessage], ) -> Vec { + let agens_endpoint = is_agens_endpoint(config); let mut output = vec![json!({ "role": "system", "content": system_prompt })]; output.extend(messages.iter().map(|message| { - let mut item = json!({ "role": message.role, "content": openai_message_content(message) }); + let content = if agens_endpoint && message.role == "assistant" && !message.tool_calls.is_empty() { + // The canonical Chat Completions representation uses null content + // for an assistant turn that only contains tool calls. + serde_json::Value::Null + } else { + openai_message_content(message) + }; + let mut item = json!({ "role": message.role, "content": content }); if message.role == "tool" { if let Some(tool_call_id) = message.tool_call_id.as_ref() { item["tool_call_id"] = json!(tool_call_id); @@ -4101,7 +4137,11 @@ async fn stream_responses_with_tools( let mut body = json!({ "model": request.config.model, - "input": build_responses_input_with_tools(&request.system_prompt, &request.messages), + "input": build_responses_input_with_tools_variant( + &request.system_prompt, + &request.messages, + is_agens_endpoint(&request.config), + ), "max_output_tokens": responses_max_output_tokens(request.max_tokens, &request.config), "tools": tool_json, "tool_choice": "auto", @@ -4559,25 +4599,25 @@ mod tests { use super::{ append_gemini_model_parts, apply_chat_completion_thinking_toggle, build_ai_http_client, build_gemini_contents, - build_openai_chat_messages, build_responses_input_with_tools, call_claude, call_openai_compatible, - classify_error, claude_headers, claude_system_prompt, complete, decorate_chat_completion_body, - drain_next_stream_line, emit_gemini_tool_call_parts, emit_responses_function_call_item, format_transport_error, - gemini_text, is_kimi_model, is_retryable_error, list_models_core, maybe_bearer_headers, maybe_tag_retry_after, - measure_first_stream_chunk, merge_global_max_retries, minimax_stream_semantics, - ollama_selected_model_tool_support, openai_message_content, openai_response_text, openai_stream_reasoning, - openai_stream_text, parse_dynamic_effort_capability, parse_gemini_model_list_response, - parse_model_list_response, parse_retry_after, parse_retry_after_secs, provider_requires_api_key, - resolve_endpoint, resolve_gemini_stream_endpoint, resolve_model_effort_core, resolve_model_list_endpoint, - resolve_ollama_show_endpoint, responses_function_tool, responses_max_output_tokens, responses_stream_text, - responses_text, responses_token_usage, retain_ollama_completion_models, retry_after_secs, - set_chat_completion_token_limit, stream, stream_claude, stream_claude_with_tools, stream_data_payload, - stream_error, stream_openai_with_tools, stream_with_tools, test_connection_core, uses_anthropic_messages_api, - validate_config, validate_model_list_config, with_retry, with_stream_retry, AiApiStyle, AiAssistantMode, - AiAuthMethod, AiCapabilitySource, AiChatSelectionState, AiCompletionRequest, AiConfig, AiEffortCapability, - AiEffortOption, AiEffortSelection, AiInlineImage, AiMessage, AiModelInfo, AiProvider, AiReasoningLevel, - MiniMaxStreamDelta, MiniMaxStreamState, MiniMaxTextAccumulator, StreamToolEvent, StreamingToolCallAccumulator, - ToolCallRef, AUTHORIZATION, CLAUDE_DEFAULT_SYSTEM, CONTENT_TYPE, MINIMAX_REASONING_DETAILS_PAYLOAD_KEY, - TEST_PROMPT, + build_openai_chat_messages, build_responses_input_with_tools, build_responses_input_with_tools_variant, + call_claude, call_openai_compatible, classify_error, claude_headers, claude_system_prompt, complete, + decorate_chat_completion_body, drain_next_stream_line, emit_gemini_tool_call_parts, + emit_responses_function_call_item, format_transport_error, gemini_text, is_agens_endpoint, is_kimi_model, + is_retryable_error, list_models_core, maybe_bearer_headers, maybe_tag_retry_after, measure_first_stream_chunk, + merge_global_max_retries, minimax_stream_semantics, ollama_selected_model_tool_support, openai_message_content, + openai_response_text, openai_stream_reasoning, openai_stream_text, parse_dynamic_effort_capability, + parse_gemini_model_list_response, parse_model_list_response, parse_retry_after, parse_retry_after_secs, + provider_requires_api_key, resolve_endpoint, resolve_gemini_stream_endpoint, resolve_model_effort_core, + resolve_model_list_endpoint, resolve_ollama_show_endpoint, responses_function_tool, + responses_max_output_tokens, responses_stream_text, responses_text, responses_token_usage, + retain_ollama_completion_models, retry_after_secs, set_chat_completion_token_limit, stream, stream_claude, + stream_claude_with_tools, stream_data_payload, stream_error, stream_openai_with_tools, stream_with_tools, + test_connection_core, uses_anthropic_messages_api, validate_config, validate_model_list_config, with_retry, + with_stream_retry, AiApiStyle, AiAssistantMode, AiAuthMethod, AiCapabilitySource, AiChatSelectionState, + AiCompletionRequest, AiConfig, AiEffortCapability, AiEffortOption, AiEffortSelection, AiInlineImage, AiMessage, + AiModelInfo, AiProvider, AiReasoningLevel, MiniMaxStreamDelta, MiniMaxStreamState, MiniMaxTextAccumulator, + StreamToolEvent, StreamingToolCallAccumulator, ToolCallRef, AUTHORIZATION, CLAUDE_DEFAULT_SYSTEM, CONTENT_TYPE, + MINIMAX_REASONING_DETAILS_PAYLOAD_KEY, TEST_PROMPT, }; #[test] @@ -6655,6 +6695,46 @@ mod tests { assert!(tool_json.get("function").is_none()); } + #[test] + fn agens_responses_variant_adds_replay_metadata_without_changing_default() { + let messages = [AiMessage { + role: "assistant".to_string(), + content: String::new(), + images: Vec::new(), + tool_call_id: None, + tool_calls: vec![ToolCallRef { + id: "call_1".to_string(), + name: "list_tables".to_string(), + arguments: serde_json::json!({}), + provider_payload: None, + }], + }]; + + let standard = build_responses_input_with_tools("", &messages); + assert!(standard[0].get("id").is_none()); + assert!(standard[0].get("status").is_none()); + + let agens = build_responses_input_with_tools_variant("", &messages, true); + assert_eq!(agens[0]["id"], "fc_call_1"); + assert_eq!(agens[0]["status"], "completed"); + } + + #[test] + fn only_agens_endpoints_enable_compatibility_variant() { + let agens: AiConfig = serde_json::from_value(serde_json::json!({ + "provider": "custom", + "apiKey": "key", + "endpoint": "https://api.agnes-ai.cn/v1", + "model": "agnes-2.5-flash", + "apiStyle": "responses", + })) + .unwrap(); + let other = AiConfig { endpoint: "https://api.openai.com/v1".to_string(), ..agens.clone() }; + + assert!(is_agens_endpoint(&agens)); + assert!(!is_agens_endpoint(&other)); + } + #[test] fn responses_tool_done_item_can_supply_complete_function_call() { let mut accumulator = StreamingToolCallAccumulator::new();