Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 327 Agent status: 1 with issues Updated Oct 7, 2026
AgentGid

DSPy changelog: what's new each month

Every stable DSPy release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

September 2026

1 release: 3.4.0

September brought major updates to DSPy's language model interface, introducing OpenAI-style message calls and a new custom engine system that gives users fine-grained control over model connections and behavior.

Action needed
  • Client configuration (apikey, apibase, timeout) must now be set on custom engine objects, not on LM construction or calls.
  • Custom engines require importable classes and JSON-compatible serialization; unsafe loading requires allowunsafelmstate=True flag.
  • Native n answers behavior changed: separate sequential requests are now used, potentially affecting billing and retry logic.

Highlights

  • OpenAI-style lm(messages=[...]) calls now supported, including SDK message objects.
  • Custom engine interface allows you to implement your own BaseLM.forward() and aforward() integrations.
  • Native n answers now use separate sequential requests, with explicit control over billing and retry behavior.
  • Timeout handling improved with native bounds on individual transport waits rather than total generation time.
  • Streaming and caching behavior clarified: streams are not retried after emitting chunks, and incomplete responses are not cached.

New

  • OpenAI-style lm(messages=[...]) calls with SDK message objects.
  • Custom engine objects that own their own connections and client settings.
  • LegacyEngine and AsyncLegacyEngine as transition tools for custom integrations.
  • Native n answers via separate sequential requests.
  • Stream chunk emission prevents retry behavior.
  • Rate-limit error preservation of retry hints, request IDs, and headers.

Improved

  • Client settings (apikey, apibase, timeout) now live on engine objects rather than LM construction.
  • Timeout behavior now bounds individual transport waits instead of total generation time.
  • Incomplete responses are no longer incorrectly cached as successful results.
  • Rate-limit errors preserve metadata from the original request for better debugging.

Fixed

  • Streams no longer attempt retries after emitting a chunk.
  • Incomplete responses no longer cached as successful completions.
  • Custom engines can properly save and restore state via JSON-compatible dumpstate() and loadstate() methods.

August 2026

2 releases: 3.3.0 → 3.3.1

August brought major improvements to DSPy's tool calling and message handling architecture, with native multi-turn support and better structured history management. Security and reliability also received attention with sandbox isolation fixes.

Highlights

  • Native multi-turn tool call support now replays prior calls and results as structured messages instead of flattening them into text.
  • Parallel tool calls are preserved with IDs in both native and non-native modes for clearer tracking.
  • Each conversation turn is now stored as structured messages in dspy.History, enabling better prompt caching across stable prefixes.
  • Custom LM implementations can now work with a single typed request-response path instead of guessing provider-specific formats.
  • Sandbox isolation was significantly strengthened to prevent desynchronization, file collisions, and host-tool identity mutations.

New

  • Parallel tool calls support with per-call tracking by ID.
  • Multi-turn native tool call support with message replay.
  • Structured message-based turn storage in dspy.History.
  • Image.fromurl() now downloads resources and returns embedded data URIs.
  • Diagnostic event tracking for sandbox-to-host tool calls and interpreter lifecycle.

Improved

  • LiteLLM transitions to optional compatibility fallback instead of required core dependency.
  • Custom LM authors now implement one typed LMRequest to LMResponse path.
  • Custom LMs can translate between DSPy types and their own provider or inference stack.
  • Adapters can now depend on DSPy's representation of messages, multimodal content, tool calls, reasoning, citations, and metadata.
  • Prompt caching effectiveness improved through stable message prefixes in structured history.

Fixed

  • Unsolicited sandbox diagnostics no longer desynchronize JSON-RPC replies.
  • Request IDs are now unpredictable to prevent execution manipulation.
  • Mounted files with distinct host paths no longer silently collide at the same guest path.
  • Guest code cannot change host-tool identity through JavaScript global mutation.
  • Sandbox runtime files are protected and Deno cache access is revoked after execution.

May 2026

1 release: 3.2.1

May brought fixes for streaming and caching behavior, along with documentation updates to better support production deployments and community use cases.

Highlights

  • Fixed custom headers forwarding in async streaming LM calls to LiteLLM
  • Fixed per-call caching=False behavior for both sync and async embedding calls
  • Removed upper bound constraint on litellm dependency
  • Reorganized documentation to better highlight production use cases and deployment guidance

Improved

  • Usecase page updated with clearer guidelines for community contributions
  • Production use-cases documentation copy revised
  • Deployment moved into technical documentation tabs for better visibility

Fixed

  • Async streaming LM calls now correctly forward custom headers to LiteLLM
  • Embedder per-call caching=False now honored for sync and async calls
  • MkDocs admonition rendering in Deployment and Observability documentation
  • Duplicate-word typos in documentation, source code, and tests

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. DSPy: pricing, features and alternatives · All changelogs