Change final answer function, add iPython support

What's Changed

• Add simple compaction capabilities, preliminary by @alexzhang13 in #110
• feat: add depth>1 recursive subcalls with limits and cost tracking by @rawwerks in #84
• Change asyncio.run_all() --> semaphore queue for sub-calls by @alexzhang13 in #131
• Add iPython REPL environment support by @alexzhang13 in #152
• New dict-based Final Answer format by @alexzhang13 in #162
• Add prime-rl-based RLM training harness and update the local REPL implementation. by @alexzhang13 in #165

Full Changelog: v0.1.1...v0.1.2

Depth > 1 support, compaction with offloaded history

What's Changed

• Add simple compaction capabilities, preliminary in #110
• feat: add depth>1 recursive subcalls with limits and cost tracking in #84

Full Changelog: v0.1.1...v0.1.1a

Added custom tools and small QoL changes.

What's Changed

• Add custom tool call passing in #106
• Add timeouts per client by in #108
• Ensure tool calls and reserved names aren't overwritten + add logger metadata in #109

Full Changelog: v0.1.0...v0.1.1

v0.1.0

Initial release on PyPi for RLMs

Initial Release

Initial Release

Working implementation of RLMs that supports local, Docker, Modal, and Prime Intellect sandboxes.

• Mainly communicates between lm_handler and the sandbox through either HTTP or sockets.
• Supports the sub-call llm_query and batch_llm_query for multiple asynchronous subclass.
• Supports persistent REPLs for RLM clients, meaning the REPL environment does not reset after every RLM call.
• Supports most major client providers (mainly tested with OpenAI completions SDK).