sleipnirtest

Documentation

Index

• func AssertTokenInvariant(t testing.TB, accountant *accounting.TokenAccountant, out sleipnir.RunOutput)
• func ByType[T sleipnir.Event](c *EventCollector) []T
• func MultiToolCallResponse(calls ...anyllm.ToolCall) *anyllm.ChatCompletion
• func StaticTool(name, result string) sleipnir.Tool
• func TextResponse(text string) *anyllm.ChatCompletion
• func ToolCallResponse(toolName, toolCallID string, args json.RawMessage) *anyllm.ChatCompletion
• type EventCollector
  • func NewEventCollector() *EventCollector
  • func (c *EventCollector) ByAgent(name string) []sleipnir.Event
  • func (c *EventCollector) CheckCompleted() error
  • func (c *EventCollector) Events() []sleipnir.Event
  • func (c *EventCollector) Send(e sleipnir.Event)
  • func (c *EventCollector) TokensFor(agentName string) sleipnir.Usage
  • func (c *EventCollector) ToolCalls() []sleipnir.ToolCallEvent
• type FakeMCPServer
  • func NewFakeMCPServer(t *testing.T) *FakeMCPServer
  • func (f *FakeMCPServer) AddRawTool(name, description string, ...)
  • func (f *FakeMCPServer) AddTool(name, description string, ...)
  • func (f *FakeMCPServer) Client() *mcp.ClientSession
  • func (f *FakeMCPServer) Start()
• type StubProvider
  • func NewStubProvider(t *testing.T, responses ...*anyllm.ChatCompletion) *StubProvider
  • func (s *StubProvider) Completion(_ context.Context, params anyllm.CompletionParams) (*anyllm.ChatCompletion, error)
  • func (s *StubProvider) CompletionStream(_ context.Context, _ anyllm.CompletionParams) (<-chan anyllm.ChatCompletionChunk, <-chan error)
  • func (s *StubProvider) Name() string
  • func (s *StubProvider) WhenMessageMatches(pattern *regexp.Regexp, resp *anyllm.ChatCompletion) *StubProvider
• type StubTool
  • func NewStubTool(name, desc string, schema map[string]any, ...) *StubTool
  • func (s *StubTool) Definition() sleipnir.ToolDefinition
  • func (s *StubTool) Invocations() []StubToolInvocation
  • func (s *StubTool) Invoke(ctx context.Context, args json.RawMessage) (sleipnir.ToolResult, error)
  • func (s *StubTool) InvokeCount() int
• type StubToolInvocation
• type TodoCollector
  • func NewTodoCollector() *TodoCollector
  • func (c *TodoCollector) All() map[string][]sleipnir.TodoItem
  • func (c *TodoCollector) ByToolCallID(id string) []sleipnir.TodoItem
  • func (c *TodoCollector) Send(e sleipnir.Event)

Functions

func AssertTokenInvariant

func AssertTokenInvariant(t testing.TB, accountant *accounting.TokenAccountant, out sleipnir.RunOutput)

AssertTokenInvariant verifies that TokenAccountant.Total().TotalTokens equals RunOutput.Usage.TotalTokens. Call immediately after Run returns.

func ByType

func ByType[T sleipnir.Event](c *EventCollector) []T

ByType returns all events in c that are of type T.

func MultiToolCallResponse

func MultiToolCallResponse(calls ...anyllm.ToolCall) *anyllm.ChatCompletion

MultiToolCallResponse returns a *ChatCompletion scripted to call multiple tools in a single turn.

func StaticTool

func StaticTool(name, result string) sleipnir.Tool

StaticTool returns a Tool with the given name that always returns result. Useful for registering tools with duplicate names to trigger ErrToolNameCollision.

func TextResponse

func TextResponse(text string) *anyllm.ChatCompletion

TextResponse returns a *ChatCompletion containing only assistant text and no tool calls. Covers the most common scripted scenario in harness tests.

func ToolCallResponse

func ToolCallResponse(toolName, toolCallID string, args json.RawMessage) *anyllm.ChatCompletion

ToolCallResponse returns a *ChatCompletion scripted to call one tool.

Types

type EventCollector

type EventCollector struct {
	// contains filtered or unexported fields
}

EventCollector implements Sink and accumulates all sent events. Safe for concurrent use.

func NewEventCollector

func NewEventCollector() *EventCollector

func (*EventCollector) ByAgent

func (c *EventCollector) ByAgent(name string) []sleipnir.Event

ByAgent returns all events where the AgentName field equals name. Uses reflection so it works across all event types.

func (*EventCollector) CheckCompleted

func (c *EventCollector) CheckCompleted() error

CheckCompleted returns an error if no AgentEndEvent with StopDone was received.

func (*EventCollector) Events

func (c *EventCollector) Events() []sleipnir.Event

Events returns a snapshot of all collected events.

func (*EventCollector) Send

func (c *EventCollector) Send(e sleipnir.Event)

func (*EventCollector) TokensFor

func (c *EventCollector) TokensFor(agentName string) sleipnir.Usage

TokensFor returns the Usage from the AgentEndEvent for the named agent. Returns zero Usage if no AgentEndEvent was received for that agent.

func (*EventCollector) ToolCalls

func (c *EventCollector) ToolCalls() []sleipnir.ToolCallEvent

ToolCalls returns all ToolCallEvents collected.

type FakeMCPServer

type FakeMCPServer struct {
	// contains filtered or unexported fields
}

FakeMCPServer is an in-process MCP server for testing. Register tools with AddTool, then call Start to connect a client. After Start returns, Client() provides a connected *mcp.ClientSession.

func NewFakeMCPServer

func NewFakeMCPServer(t *testing.T) *FakeMCPServer

NewFakeMCPServer creates a FakeMCPServer. Call AddTool to register tools, then call Start to connect. Cleanup is registered with t.Cleanup.

func (*FakeMCPServer) AddRawTool

func (f *FakeMCPServer) AddRawTool(name, description string, fn func(ctx context.Context, args map[string]any) (*mcp.CallToolResult, error))

AddRawTool registers a tool whose handler returns a raw *mcp.CallToolResult. Use this when you need to return multiple content items or non-text content. Must be called before Start.

func (*FakeMCPServer) AddTool

func (f *FakeMCPServer) AddTool(name, description string, fn func(ctx context.Context, args map[string]any) (string, error))

AddTool registers a low-level tool on the server. fn receives the raw arguments as a map[string]any and returns a text result. Must be called before Start.

func (*FakeMCPServer) Client

func (f *FakeMCPServer) Client() *mcp.ClientSession

Client returns the connected MCP client session. Panics if Start has not been called.

func (*FakeMCPServer) Start

func (f *FakeMCPServer) Start()

Start connects the server and client in-process and registers cleanup. Must be called before Client().

type StubProvider

type StubProvider struct {
	// contains filtered or unexported fields
}

func NewStubProvider

func NewStubProvider(t *testing.T, responses ...*anyllm.ChatCompletion) *StubProvider

func (*StubProvider) Completion

func (s *StubProvider) Completion(_ context.Context, params anyllm.CompletionParams) (*anyllm.ChatCompletion, error)

Completion implements anyllm.Provider

func (*StubProvider) CompletionStream

func (s *StubProvider) CompletionStream(_ context.Context, _ anyllm.CompletionParams) (<-chan anyllm.ChatCompletionChunk, <-chan error)

CompletionStream implements anyllm.Provider - not used in tests; panics with "not implemented"

func (*StubProvider) Name

func (s *StubProvider) Name() string

func (*StubProvider) WhenMessageMatches

func (s *StubProvider) WhenMessageMatches(pattern *regexp.Regexp, resp *anyllm.ChatCompletion) *StubProvider

WhenMessageMatches registers a response for calls where the last user message content matches pattern. Checked before the scripted sequence. Returns the receiver for chaining.

type StubTool

type StubTool struct {
	// contains filtered or unexported fields
}

StubTool is a Tool for use in tests. It records every invocation.

func NewStubTool

func NewStubTool(name, desc string, schema map[string]any,
	fn func(context.Context, json.RawMessage) (sleipnir.ToolResult, error)) *StubTool

func (*StubTool) Definition

func (s *StubTool) Definition() sleipnir.ToolDefinition

func (*StubTool) Invocations

func (s *StubTool) Invocations() []StubToolInvocation

func (*StubTool) Invoke

func (s *StubTool) Invoke(ctx context.Context, args json.RawMessage) (sleipnir.ToolResult, error)

func (*StubTool) InvokeCount

func (s *StubTool) InvokeCount() int

type StubToolInvocation

type StubToolInvocation struct {
	Args   json.RawMessage
	Result sleipnir.ToolResult
	Err    error
}

type TodoCollector

type TodoCollector struct {
	// contains filtered or unexported fields
}

TodoCollector implements Sink and accumulates TodoEvents. After a run, ByToolCallID returns the final todo list for each invocation.

func NewTodoCollector

func NewTodoCollector() *TodoCollector

func (*TodoCollector) All

func (c *TodoCollector) All() map[string][]sleipnir.TodoItem

func (*TodoCollector) ByToolCallID

func (c *TodoCollector) ByToolCallID(id string) []sleipnir.TodoItem

func (*TodoCollector) Send

func (c *TodoCollector) Send(e sleipnir.Event)