The Idea

Last updated: March 29, 2026

On March 27, 2026, antirez posted about agent harnesses:

One thing agent harnesses should be able to do is: to jump back in history trimming what follows, just injecting some self-steering text.

And the harder question:

I wonder if without explicit reinforcement learning for this kind of context saving, the model will be able to use it effectively.

That's the question pi-weaver tries to answer.

The name comes from Dota 2. Weaver's ultimate ability, Time Lapse, reverses position, health, and mana to 5 seconds ago. Everything that happened in those 5 seconds is undone: damage taken, mana spent, distance traveled. But the player's knowledge of what happened persists.

That's the exact mechanic we wanted. The agent reverses to a checkpoint. Conversation context rewinds. Structured state (the model's knowledge) is preserved. Everything between (the failed grep, the wrong edit, the misleading error) is gone.

It maps to programming primitives:

  • checkpoint = try:
  • time_lapse = raise
  • steering message = exception message
  • done = return

The bet was specific: give the model these tools and a good prompt, and it will use them without reinforcement learning. No fine-tuning, no RLHF on rewind behavior. Just tool descriptions and a cookbook of when to use them. Another bet on assumptions that might decay if models get RL for this.

Whether the bet paid off is more complicated than a yes or no. The model does use the tools: 17 time_lapse calls across 15 tasks. But using them well is harder. The polyglot-c-py session shows the model rewinding correctly from a broken rewrite, then immediately re-entering the same trap. The tool works. The judgment doesn't always.

antirez's question was the right one to ask.

What this isn't

This isn't a general self-correction framework. It doesn't do:

  • Automatic retries (the model decides when to retry)
  • Reflection or chain-of-thought forcing (no "think about what went wrong" injection)
  • Git-based checkpointing (file changes persist; only the conversation rewinds)
  • Phase gating or plan validation (no structure imposed)

It's three tools and a prompt. The model does the rest. That's the experiment.

What came before

pi-exec was the predecessor, a 1,656-line structured executor with rigid phases and gate calls. It worked well on controlled tasks but couldn't adapt mid-run. The architecture page covers why we threw it away.

The evaluation

We ran 15 Terminal-Bench 2.0 tasks with Sonnet 4.6, comparing plain pi against pi+weaver. Both scored 11/15, same pass rate, different tasks. The interesting question isn't "does it work" but "when does it help and when does it hurt." The task pages and analysis pages try to answer that.