Crate aonyx_agent 

aonyx-agent

The agent loop and its inner subsystems.

Subsystems

• runner — the main AgentRunner::run(session, msg) loop.
• compaction — context-window pressure monitor + summarization triggers.
• classifier — routes user input to a prompt template (chitchat / task / recall / code / research).
• approval — gate around destructive tool calls.
• subagent — spawn isolated child agents with a whitelisted tool set (V2).

Loop sketch

loop {
  inject(skills_active(query, project));
  inject(memory_recall(query, k=10));
  chunks = llm.chat_stream(messages, tools).await;
  for tool_call in chunks.tool_calls() {
      approval.check(tool_call)?;
      result = tools.invoke(tool_call).await?;
      messages.push(result);
  }
  if chunks.no_tool_call() { break; }
}
post_turn::maybe_diary_append();
post_turn::maybe_kg_upsert();

Re-exports

pub use approval::ApprovalPolicy;

pub use runner::AgentRunner;

pub use runner::TurnEvent;

pub use runner::TurnResult;

Modules

approval

Approval gate.

classifier

Lightweight intent classifier.

compaction

Context-window pressure monitor + summarization.

runner

Main agent loop.

subagent

Subagent spawning.