Agent Types

Core request/response types and message structures for agent execution. Serves as the data model layer between the orchestrator and the container runner.

Package: internal/agent

Types

AgentRequest

Input sent to the agent runner.

Field	Type	Description
`SessionID`	`string`	Persistent session ID for resuming context
`ForkSession`	`bool`	Fork session on first thread message (thread isolation)
`Messages`	`[]AgentMessage`	Conversation history (oldest-first)
`SystemPrompt`	`string`	Custom system prompt override
`ChannelID`	`string`	Chat channel or thread ID
`AuthorID`	`string`	ID of user who triggered the request
`DirPath`	`string`	Working directory for the container
`ParentDirPath`	`string`	Parent working directory (for worktree containers)
`PlanMode`	`bool`	Enable plan mode
`Prompt`	`string`	Current user prompt
`OnTurn`	`func(string)`	Callback for each assistant text turn during streaming
`OnToolUse`	`func(name, input string)`	Callback for each tool invocation
`OnActivity`	`func(activity, detail string)`	Callback for model detection and system events

AgentMessage

Single message in conversation history.

Field	Type	Description
`Role`	`string`	`"user"` or `"assistant"`
`Content`	`string`	Message text

AgentResponse

Output returned by the agent runner.

Field	Type	Description
`Response`	`string`	Final text response
`SessionID`	`string`	Updated session ID for resuming next turn
`Error`	`string`	Error message if run failed
`DurationMs`	`int`	Total execution time
`NumTurns`	`int`	Number of assistant turns
`StopReason`	`string`	Why the agent stopped (e.g. `"end_turn"`)
`Model`	`string`	Model used (e.g. `"claude-opus-4-6"`)

BuildPrompt

AgentRequest.BuildPrompt() returns the effective prompt for the request:

Session resume with prompt — returns only the new Prompt (avoids redundancy)
Session resume with messages — returns only the last message’s content
No session — returns all messages formatted as "role: content\n"

Callback-Driven Streaming

All callbacks are optional (nil-safe):

OnTurn — fired on each assistant text turn (used for streaming to chat)
OnToolUse — fired on each tool invocation (e.g. Bash, Read, Edit)
OnActivity — fired for model detection and subagent progress

Usage

Consumer	Purpose
`internal/container/runner.go`	Executes agent in Docker via `Runner.Run(AgentRequest) → AgentResponse`
`internal/orchestrator/orchestrator.go`	Builds requests from message history, executes with streaming
`internal/orchestrator/executor.go`	Executes scheduled tasks, creates threads on first turn

Related docs

Containers — Docker container lifecycle
Orchestrator — Message processing pipeline
Scheduling — Task execution

Architectural Quality