trace-codex mvp

[Andrew Kent (realark)]

initial trace codex. supports:

• llm spans
• tool spans
• thinking/reasoning
• subagents
• permission requests
• suspend/resume
• compaction

kitchen sink trace: https://www.braintrust.dev/app/Braintrust%20SDKs/p/andrew-coding-agent/trace?object_type=project_logs&object_id=ebabe4b7-25b2-4884-8d85-55fa6e2e6014&r=ed6f0665-4fd9-4f7b-ad8e-e928ec8421b1&s=ed6f0665-4fd9-4f7b-ad8e-e928ec8421b1

[Abhijeet Prasad (AbhiPrasad)]

It took me a while to review this because I wanted to take some time to land on what I think is ideal here. But I feel like that would block a lot, so I'm approving to get this shipped and landed.

As a follow up though, I would vastly prefer if we adopted more of a state machine architecture for this.

class CodexEventProcessor implements EventProcessor {
  private state: CodexState;
  private runtime: EffectRunner;

  async process(event: EnqueueEvent) {
    const domainEvents = await this.expandHookIntoDomainEvents(event);

    for (const domainEvent of domainEvents) {
      const result = reduceCodexState(this.state, domainEvent);
      this.state = result.state;
      await this.runtime.run(result.effects);
    }
  }

  async flush() {
    const result = reduceCodexState(this.state, { type: "FlushRequested" });
    this.state = result.state;
    await this.runtime.run(result.effects);
  }
}

This gives us a bunch of benefits:

1. tests are easier to reason about
2. snapshots in tests are nicer
3. side effects are easier to track

the state machines themselves can be hierarchal

   CodexSessionMachine
   ├── Root lifecycle
   ├── Main ScopeMachine
   │   ├── TurnMachine(s)
   │   │   ├── LlmMachine
   │   │   └── ToolMachine(s)
   │   └── CompactionMachine(s)
   └── Subagent ScopeMachine(s)
       ├── TurnMachine(s)
       ├── LlmMachine
       └── ToolMachine(s)