Documentation
¶
Index ¶
- Constants
- func BuildSystemMessage(ctx context.Context, toolreg *tools.Registry, sysPrompt, toolsPrompt string) (llm.Message, []llm.ToolDefinition)
- func ConvertMCPToolsToLLMTools(tools []mcps.ToolDescriptor) []llm.ToolDefinition
- func StreamChat(ctx context.Context, loop *AgentLoop, messages []llm.Message, ...) (string, error)
- func ThisMoment() string
- type AgentLoop
- type AgentLoopConfig
- type AgentLoopOption
- type ChatExecutor
- type StreamCallbacks
- type ToolExecutor
- func (e *ToolExecutor) ExecuteToolCallLoop(ctx context.Context, messages []llm.Message, tools []llm.ToolDefinition, ...) (string, []llm.ToolCall, *llm.Usage, error)
- func (e *ToolExecutor) ExecuteToolCalls(ctx context.Context, messages []llm.Message, toolCalls []llm.ToolCall, ...) ([]*llm.Event, []llm.Message, bool)
Constants ¶
const ( // DefaultSystemMsg is the fallback system prompt when no preset is configured. DefaultSystemMsg = "" /* 198-byte string literal not displayed */ // DefaultToolsMsg is the fallback tool-usage prompt when no preset is configured. DefaultToolsMsg = "" /* 712-byte string literal not displayed */ )
Variables ¶
This section is empty.
Functions ¶
func BuildSystemMessage ¶
func BuildSystemMessage(ctx context.Context, toolreg *tools.Registry, sysPrompt, toolsPrompt string) (llm.Message, []llm.ToolDefinition)
BuildSystemMessage constructs a system message from the tool registry and prompts. Returns the system message and the resolved tool definitions.
func ConvertMCPToolsToLLMTools ¶
func ConvertMCPToolsToLLMTools(tools []mcps.ToolDescriptor) []llm.ToolDefinition
ConvertMCPToolsToLLMTools converts MCP tool descriptors to LLM tool definitions.
func StreamChat ¶
func StreamChat(ctx context.Context, loop *AgentLoop, messages []llm.Message, tools []llm.ToolDefinition, cb StreamCallbacks) (string, error)
StreamChat runs the agent loop in streaming mode, invoking callbacks for each delta and think event. Returns the full accumulated answer text.
func ThisMoment ¶
func ThisMoment() string
ThisMoment returns a timestamp string in Chinese 时辰 format.
Types ¶
type AgentLoop ¶
type AgentLoop struct {
// contains filtered or unexported fields
}
AgentLoop encapsulates the full agent loop: LLM call + tool execution + iteration.
func NewAgentLoop ¶
func NewAgentLoop(cfg AgentLoopConfig, opts ...AgentLoopOption) *AgentLoop
NewAgentLoop creates a new AgentLoop with the given config and options. Defaults MaxLoop to 5 when unset or <= 0.
func (*AgentLoop) Run ¶
func (al *AgentLoop) Run(ctx context.Context, messages []llm.Message, tools []llm.ToolDefinition) iter.Seq2[*llm.Event, error]
Run executes the streaming agent loop, returning an iter.Seq2 that yields events from StreamChat and tool execution. The caller consumes events via `for event, err := range loop.Run(ctx, messages, tools)`.
The iterator runs synchronously within the caller's goroutine. It stops when:
- LLM returns no tool calls (final answer)
- all executed tools signal terminate=true
- maxLoop iterations are exceeded (yields error event)
- StreamChat returns an error (yields error event)
- the caller breaks out of the range loop
func (*AgentLoop) RunNonStreaming ¶
func (al *AgentLoop) RunNonStreaming(ctx context.Context, messages []llm.Message, tools []llm.ToolDefinition) (string, error)
RunNonStreaming executes the non-streaming agent loop, returning the final answer string. It calls Chat, collects tool calls, executes tools, and loops until no tool calls remain or all tools signal terminate.
type AgentLoopConfig ¶
type AgentLoopConfig struct {
LLM llm.Client // LLM client for Chat and StreamChat
ToolExec *ToolExecutor // tool executor (same package)
MaxLoop int // safety net: max loop iterations, default 5
}
AgentLoopConfig holds configuration for AgentLoop.
type AgentLoopOption ¶
type AgentLoopOption func(*AgentLoopConfig)
AgentLoopOption is a functional option for AgentLoopConfig.
func WithMaxLoop ¶
func WithMaxLoop(n int) AgentLoopOption
WithMaxLoop sets the maximum number of loop iterations before safety-net termination.
type ChatExecutor ¶
type ChatExecutor func(ctx context.Context, messages []llm.Message, tools []llm.ToolDefinition) (string, []llm.ToolCall, *llm.Usage, error)
ChatExecutor 定义聊天执行函数类型，支持流式/非流式
type StreamCallbacks ¶
StreamCallbacks holds callbacks for streaming agent responses.
type ToolExecutor ¶
type ToolExecutor struct {
// BeforeToolCall 在每个工具调用前执行。返回 block=true 时阻止工具执行，
// reason 作为 error 内容写入 ToolResult。可选，nil 时跳过。
BeforeToolCall func(ctx context.Context, name string, params map[string]any) (block bool, reason string)
// AfterToolCall 在每个工具成功执行后调用，可修改 result。
// 返回的 map 会替代原 result 用于后续格式化。可选，nil 时跳过。
AfterToolCall func(ctx context.Context, name string, result map[string]any) map[string]any
// contains filtered or unexported fields
}
ToolExecutor 封装工具调用循环逻辑
func NewToolExecutor ¶
func NewToolExecutor(toolreg *tools.Registry) *ToolExecutor
NewToolExecutor 创建 ToolExecutor
func (*ToolExecutor) ExecuteToolCallLoop ¶
func (e *ToolExecutor) ExecuteToolCallLoop( ctx context.Context, messages []llm.Message, tools []llm.ToolDefinition, exec ChatExecutor, ) (string, []llm.ToolCall, *llm.Usage, error)
ExecuteToolCallLoop 执行工具调用循环，直到无 tool calls
func (*ToolExecutor) ExecuteToolCalls ¶
func (e *ToolExecutor) ExecuteToolCalls(ctx context.Context, messages []llm.Message, toolCalls []llm.ToolCall, think string) ([]*llm.Event, []llm.Message, bool)
ExecuteToolCalls 执行单轮工具调用（并发），返回事件列表、更新后的消息和 allTerminate。 allTerminate 在本次所有成功执行的 tool 都标记 terminate 时为 true。
Source Files