Context

The ExecutionLoop protocol (defined in M3, #124) enables pluggable loop architectures. The MVP ships with ReAct (Loop 1). This issue implements the second loop: Plan-and-Execute — a two-phase approach where the agent first creates a full plan, then executes each step sequentially.

Acceptance Criteria

Plan-and-Execute Loop Implementation

• Implements the ExecutionLoop protocol
• Phase 1 (Plan): agent creates a structured plan with numbered steps
• Phase 2 (Execute): agent executes each step sequentially, checking results against expectations
• Re-planning: if a step fails or produces unexpected results, agent can revise remaining steps
• max_replans configurable (default: 3) to prevent infinite revision loops
• Plan visible in agent output for observability

Auto-Selection by Complexity (Future)

• Complexity scoring function (input length, tool count, estimated turns)
• Automatic loop selection: simple tasks → ReAct, complex tasks → Plan-and-Execute
• Override via agent config or per-task setting

Testing

• Unit tests with scripted FakeProvider (deterministic plan + execution)
• Integration test: complex multi-step task completed via plan-and-execute
• Comparison test: same task via ReAct vs Plan-and-Execute (verify both complete)

Dependencies

• Define ExecutionLoop protocol and implement ReAct loop (DESIGN_SPEC §6.5) #124 — ExecutionLoop protocol + ReAct (must exist first)
• Implement agent engine core with ExecutionLoop protocol integration (DESIGN_SPEC §3.1, §6.1, §6.5) #11 — Agent engine core

Design Spec Reference

• §6.5 — Agent Execution Loop (Loop 2: Plan-and-Execute)