Release Notes

0.2.44 (2026-04-07)

v2-actions: Add unified Actions -> Action -> Inputs -> Input hierarchy for models and tools
- Models now expose model.actions["run"].inputs with model.inputs as a shorthand
- Tools now expose tool.actions["action_name"].inputs for action discovery and parameter configuration
- Inputs support schema-aware validation, defaults, reset behavior, and dot/bracket access patterns
rlm: Add Recursive Language Model support to the SDK
- New RLM resource enables long-context analysis through an iterative REPL sandbox
- Supports orchestrator/worker model separation, structured context handling, and per-run REPL logs
- Available on the v2 surface and mirrored into the legacy v1 module layer
agent: Add v2 agent duplication support
- Agents can now be duplicated directly from the SDK
- Supports duplicating referenced subagents for independent copies when needed

agent: Improve polling ergonomics for v2 runs
- Agent.poll() and Agent.sync_poll() now accept either a full poll URL or a bare execution ID
- AgentRunResult.execution_id exposes the execution ID directly for follow-up polling and debugging
- Polling now consistently targets the supported /sdk/agents/{execution_id}/result endpoint
tool: Tighten action scoping and validation
- allowed_actions validation is stricter and better aligned with available integration actions
- Single-action tools can infer the action automatically in more cases
- Tool updates preserve scoped action configuration more reliably
integration: Improve action/input discovery and connection handling
uploads: Improve permanent upload S3 path generation
readme/docs: Refresh SDK landing experience and docs

v2: Improve snake_case and camelCase compatibility across resource payloads and serialization
agent: Fix polling behavior targeting unsupported run endpoints
tool/integration: Fix update and reconnect edge cases around integration-backed tools
tests: Stabilize benchmark, SQL tool, team agent, and v2 functional coverage