Enum AgentCommand 

pub enum AgentCommand {
    SendMessage {
        text: String,
        images: Vec<ImagePart>,
        image_markers: Vec<usize>,
    },
    Cancel,
    ApproveTool,
    ApproveToolAlways,
    DenyTool,
    ReloadConfig(Config),
    ChangeDir(String),
    AppendInput(String),
    ClearConversation,
    SetMessages(Vec<Message>),
    SetPlanMode(bool),
    Compact {
        prompt: Option<String>,
    },
    Remember {
        content: String,
        global: bool,
    },
    Forget {
        keyword: String,
    },
    ShowMemory,
    Background {
        task: String,
    },
    RefreshContextStats,
    ReloadHooks,
    SyncMessages,
    Shutdown,
}

Commands sent FROM the UI TO the agent loop.

Variants

SendMessage

User sent a message (may include attached file content and/or images). image_markers[i] is the [Image #N] number printed for images[i] at paste time. Round-tripped through AgentEvent::RestorePendingImages so that on VL preprocess failure the TUI can re-attach images with their ORIGINAL markers — otherwise an UP-recalled [Image #5] text wouldn’t match a freshly-renumbered restored image. Empty when the caller has no images (slash commands, queued text from streaming, CLI single-shot).

Fields

text: String

images: Vec<ImagePart>

image_markers: Vec<usize>

Cancel

Cancel current operation.

ApproveTool

Approve a pending tool call.

ApproveToolAlways

Approve and always allow this tool for the session.

DenyTool

Deny a pending tool call.

ReloadConfig(Config)

Reload config from TUI (the single source of truth for in-memory config, including ephemeral OAuth providers). Switches to the new default provider.

ChangeDir(String)

Change working directory.

AppendInput(String)

Append input during streaming — queued and injected before next LLM call.

ClearConversation

Clear conversation history.

SetMessages(Vec<Message>)

Set messages from a resumed session.

SetPlanMode(bool)

Set plan mode (read-only exploration, no edits).

Compact

Manually compact conversation history. prompt is accepted for forward-compat with an eventual LLM-backed summarize-with-instruction path; currently unused — this is the mechanical path only.

Fields

prompt: Option<String>

Remember

Fields

content: String

global: bool

Forget

Fields

keyword: String

ShowMemory

Background

Run a one-shot task in an isolated background context (read-only-ish tool subset, independent conversation, capped turns + timeout). Result is returned via AgentEvent::BackgroundComplete.

Fields

task: String

RefreshContextStats

Recompute and re-emit a rich ContextStats snapshot. /context sends this before rendering so the user never sees a stale cache — the cache is only refreshed on LLM round-trips, so between turns (or after out-of-turn mutations like inject_post_compress_state) the snapshot can lag the actual conversation state.

ReloadHooks

Rebuild the hook executor from disk after a /plugin install|uninstall or other change to plugin state. Cheap (just re-reads JSON files); does NOT touch provider/model state, unlike ReloadConfig.

SyncMessages

Request a snapshot of the current conversation messages. The agent responds with AgentEvent::MessagesSync carrying conversation.messages. Used by the TUI before /bg to ensure the session has up-to-date message history even when a turn is still in progress (e.g. waiting for tool approval).

Shutdown

Shutdown the agent.

Trait Implementations

impl Debug for AgentCommand

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.