Auto runtime selection with plugin harness registry

[MervinPraison]

Problem

When runtime: auto is set on a model, there is no principled way to pick among plugin harnesses that support a given provider/model pair. Users must hard-code runtime ids or rely on agent-level cli_backend.

Current state

• CLI backend registry in praisonai/cli_backends/registry.py lists ids but does not score by supports(provider, model) or priority.
• Model router in praisonaiagents/llm/ selects LLM endpoints without runtime harness matching.
• Entry points: praisonai.cli_backends group; no parallel praisonai.runtimes auto-selection contract.

Proposed design

• Extend AgentRuntimeProtocol with supports(provider: str, model: str) -> bool and optional selection_priority() -> int (lower = preferred), mirroring harness selection patterns used elsewhere in the wrapper registry.
• auto resolution in model-scoped policy: filter registered runtimes by supports, sort by priority, fall back to praisonai default.
• Plugin packages (PraisonAI-Plugins) register runtimes via entry points with declared provider/model prefixes.
• Log chosen runtime id at debug level for observability.

Acceptance criteria

• auto selects highest-priority supporting runtime in unit tests with multiple registered harnesses.
• Tie-breaking and empty-match fallback to praisonai documented and tested.
• No import-time scan of heavy plugin deps; lazy load factories on selection only.

Layer

Primary: core (selection algorithm in praisonaiagents/runtime/resolve.py)

Secondary: plugins (harness registration), wrapper (entry point wiring)

3-way parity (Python / YAML / CLI)

Surface	Supported
Python	yes
YAML	yes — runtime: auto on model entries
CLI	yes — inherits from agent run model config

Priority

P2

Related modules

praisonaiagents/runtime/registry.py, praisonai/cli_backends/registry.py, praisonaiagents/llm/, praisonaiagents/plugins/

Related issues

• AgentRuntimeProtocol and built-in embedded praisonai runtime (AgentRuntimeProtocol and built-in embedded praisonai runtime #1934)
• Model-scoped runtime selection policy (Model-scoped runtime selection policy #1935)
• Runtime-scoped tool result middleware for plugin harnesses

[praisonai-triage-agent]

Claude finished @MervinPraison's task —— View job • claude/issue-1939-20260613-0842 • Create PR ➔

---

Auto runtime selection with plugin harness registry ✅

I've successfully implemented auto runtime selection with plugin harness registry following the PraisonAI architecture guidelines.

Completed Tasks:

• Setup git identity and authentication
• Read AGENTS.md to understand architecture rules
• Analyze existing codebase structure for runtime and registry components
• Create runtime directory in core SDK (praisonaiagents/)
• Extend AgentRuntimeProtocol with supports() and selection_priority() methods
• Implement auto resolution logic in runtime/resolve.py
• Update registry system for plugin harness registration
• Add entry point wiring for plugin packages
• Create unit tests for auto selection behavior
• Run linting and type checking
• Commit and push changes
• Create pull request

Implementation Summary:

Core Features Implemented:

• AgentRuntimeProtocol: New protocol with supports(provider, model) and selection_priority() methods
• Auto Resolution: Smart runtime selection based on provider/model support and priority ranking
• Plugin Registry: Entry point discovery for praisonai.runtimes group with thread-safe registration
• Built-in Runtime: Default praisonai runtime with universal support as fallback
• Comprehensive Tests: Unit tests covering auto selection, priority ordering, and edge cases

Key Files Created:

• praisonaiagents/runtime/protocols.py - Runtime protocol definition
• praisonaiagents/runtime/resolve.py - Auto-selection algorithm
• praisonaiagents/runtime/registry.py - Plugin harness registry
• praisonaiagents/runtime/builtin.py - Default praisonai runtime
• Comprehensive test suite with 15+ test cases

Usage Examples:

# Auto selection (picks best runtime based on provider/model)
runtime = resolve_runtime("openai", "gpt-4", mode="auto")

# Explicit selection
runtime = resolve_runtime("openai", "gpt-4", mode="praisonai") 

# Register custom runtime
register_runtime("custom", CustomRuntimeFactory())

Architecture Compliance:
✅ Core SDK protocols in praisonaiagents/ package
✅ No heavy dependencies in core implementation
✅ Lazy loading of plugin factories
✅ Thread-safe registry operations
✅ Proper fallback handling

---

Job run