feat: A2A agent inner loop framework (2/3)

[mdear]

A2A Agent Inner Loop Framework (2/3)

Merge order: This PR targets main but is intended to be merged after feature/local-docker-sandbox_1_of_3 (#198). It includes all B1 changes plus the incremental A2A agent layer below.

74 files changed (incremental from 1/3) — A2A inner loop strategy, backend registry, billing strategies, and comprehensive test coverage.

A2A Inner Loop

• A2AInnerLoop strategy with SSE streaming client
• CircuitBreaker (threshold=5) with automatic native fallback
• EventStreamAdapter mapping SSE events to agent runtime events
• ContextAdapter for conversation history parity with native loop
• ToolBridge for bidirectional tool registration between backends
• Agent factory wiring for inner loop strategy selection

A2A Adapter Server & Backends

• AdapterServer (FastAPI/uvicorn) running inside sandboxes on :18100
• Backend registry with simulate, copilot, claude_code, codex providers
• CopilotBackend using GitHub Copilot SDK (github-copilot-sdk>=0.1.25)
• 15 native sandbox tools bridged to Copilot CLI
• ClaudeCodeBackend and CodexBackend implementations
• Multimodal support across backends
• Sandbox start-services.sh A2A adapter startup section

A2A Billing

• Backend-aware billing in CreditUsageHandler (a2a: prefix routing)
• Three strategies: token_based, provider_reported, none
• Copilot premium-request multipliers (hot-configurable via env)
• AgentSettings: A2A billing strategy, multipliers, inner loop mode, timeout, fallback config

Events & Config

• EventGroup, EventType enums and ApplicationEvent model
• DelegationFallbackEvent, CompactionAuthorityEvent, CompactionSkippedEvent
• ModelUsageEvent A2A billing fields (billing_backend, is_user_key, premium_requests, model_multiplier)
• CompactionLock async context manager for cross-authority compaction guard
• Health endpoint: A2A inner loop mode and backend fields

Testing

• 26 new test files covering all A2A components
• Inner loop, tool bridge, agent factory, circuit breaker, event mapping
• Backend tests: copilot, claude_code, codex, adapter server, multimodal
• Credit usage handler A2A billing tests
• Unit: 5520 passed, 3 skipped, 0 failed
• E2E: 30 pass, 1 skip, 1 expected fail (chat A2A routing — added in 3/3)

Docs

• A2A billing model design doc with strategy comparison
• Copilot SDK integration assessment, inner loop parity assessment
• Tool bridge gap analysis, implementation handoff guide
• 17 design/impl/runtime/test docs

[mdear]

I am continuing testing on branch feature/a2a-chat-inner-loop_3_of_3 and will not introduce churn on this PR unless requested by the team.

[mdear]

This PR is 2 out of a sequence of 3 PRs:

#198 — #198: Docker sandbox runtime, local deploy stack, session lifecycle, frontend, test overhaul (389 files)

#199 — #199: A2A inner loop strategy, backend registry, billing strategies, adapter server (74 incremental files)

#200 — #200: Chat A2A turn loop, council A2A routing, cross-authority compaction (16 incremental files)

Merge order: #198 → #199 → #200