v0.13.0

Highlights

Programmatic Tool Calling (PTC) — LLMs can write Python code that orchestrates multiple tool calls in a single block, with full permission enforcement and execution tracing.

registry = ToolRegistry()
registry.register(search)
registry.register(summarize)
registry.ptc.enable()

# LLM generates: tool_use("programmatic_tool_call", {code: "..."})

Added

• registry.ptc controller — always-present sub-object for PTC lifecycle management
  • registry.ptc.enable(timeout=30) / registry.ptc.disable()
  • registry.ptc.enabled / registry.ptc.last_invocation_id
  • Runtime injection: registry.ptc.enable(runtime=custom_runtime)
  • Two-layer isolation: ThreadBackend (outer) + IpcSubprocessRuntime (inner)
• registry.invoke(tool_name, kwargs) — single-tool execution with full pipeline (permissions, logging)
• Invocation tracking — invocation_id on all execution log entries
  • tr_bat_ (batch execute_tool_calls), tr_ptc_ (PTC), tr_sig_ (single invoke)
  • Query: log.get_entries(invocation_id="tr_ptc_...")
• ToolMetadata.force_thread — forces ThreadBackend for tools needing main-process access
• runtimes/ bridge layer — ToolProjection, DirectProjection, PtcController for codecell integration
• codecell>=0.2.1 as optional [ptc] dependency

Changed

• Unified execution helpers — _check_tool_access() and _log_tool_result() shared by invoke() and execute_tool_calls()
• runtimes/ delegates code execution to the codecell package

Fixed

• MCP/OpenAPI wrappers survive cloudpickle serialization via __getstate__/__setstate__
• PTC tool forced to ThreadBackend to prevent registry pickle issues

Install

pip install toolregistry==0.13.0

# With PTC support:
pip install toolregistry[ptc]==0.13.0

Full Changelog: v0.12.0...v0.13.0