Starstruk Logo MarkStarstruk

v0.6.25 3.4.1

medium risk

stanfordnlp/dspySep 27, 2026 → Oct 12, 202616 days between releases

What's New

DSPy 3.4.1 makes `InstructionProposer` the default for GEPA and enables RLM to build DSPy sub-agents.

Breaking Changes

  • GEPA reflection requests now require the LM to return JSON.
  • RLM prompts now include a sub-agent section, invalidating caches from 3.4.0.
  • RLM now reserves input names `dspy`, `_dspy`, and `__dspy`.
  • Flex bridge constants moved from `dspy.predict.flex.bridge` to `dspy.primitives.facade`.
  • `dspy.ReActV2` signatures can no longer declare input fields named `history` or `tools`.

New Features

  • Structured `InstructionProposer` handles complex types (images, history) natively.
  • RLM sub-agents allow nested DSPy logic and tool usage within the REPL.
  • Tool argument defaults (Enums, Decimals, Paths) are stored in JSON form.
  • History replay preserves output-format reminders for better prompt caching efficiency.
  • DSPy now supports Python 3.15.

Improvements

  • Streaming, adapter serialization, and interpreter execution errors fixed.
  • Improvements to adapters, interpreters, and model metadata.
  • History replay preserves output-format reminders for better prompt caching efficiency.

Bug Fixes

  • Fixes for 3.4.0 import crash.
  • Fixes for streaming.
  • Fixes for adapter serialization.
  • Fixes for interpreter execution errors.

Upgrade Notes

GEPA reflection now requires JSON, so optimization results may differ. RLM prompts include a sub-agent section, invalidating caches from 3.4.0, and sub-agent calls count toward max_llm_calls. Programs saved on 3.4.0 with a null Score level no longer load.

The Two Releases

v0.6.25

DBX v0.6.25

Sep 27, 2026· 25 days ago

A major update adding unified MCP call history, data import background execution, MongoDB bulkWrite support, and various UI/UX improvements.

Full release notes
3.4.1

3.4.1

Oct 12, 2026· 9 days ago

DSPy 3.4.1 introduces a structured `InstructionProposer` as the default for GEPA and enables RLMs to build DSPy sub-agents. It adds Python 3.15 support and fixes a critical 3.4.0 import crash, alongside improvements to adapters, interpreters, and model metadata.

Full release notes