feat: local Docker sandbox + A2A inner loop with pluggable backends

[mdear]

Summary

Local Docker sandbox runtime and A2A inner loop framework with pluggable backends for self-hosted deployments.

Local Docker Sandbox

• DockerSandbox provider with shell executor and port pool manager
• Orphan container cleanup with configurable TTL and async-safe threading
• Docker compose local stack with stack_control.sh tooling (build-sandbox, patch-sandbox, cleanup, setup)
• e2b.Dockerfile: gh CLI v2.89+ installed, adapter deps included
• Storage proxy router for MinIO-backed local file serving
• Frontend: local sandbox support in workspace state and UI

A2A Inner Loop Framework

• 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
• AdapterServer (FastAPI/uvicorn) running inside sandboxes on :18100
• Backend registry: simulate, copilot, claude_code, codex

Chat A2A Turn Loop

• A2ATurnLoopService for chat mode with event translation layer
• Routes chat requests through A2A adapter (Copilot) in local mode
• A2A client singleton with URL tracking and auto-refresh on sandbox change
• Council service with parallel LLM execution and synthesis via A2A
• Fix council text doubling: separate delta collection from full_content
• Fix async coroutine bug in file_processor, vectorstore, and LLM providers
  (await get_storage().read() instead of anyio.to_thread.run_sync)

Copilot Backend

• CopilotBackend using GitHub Copilot SDK (github-copilot-sdk>=0.1.25)
• 15 native sandbox tools bridged to Copilot CLI
• Fresh sessions per run for reliable tool availability
• CLI path resolution: bundled SDK binary primary, gh fallback

Session Lifecycle

• delete_after column and schedule-delete endpoint for timed session cleanup
• Orphan sandbox cleanup with async-safe threading
• Frontend session delete UI (sidebar + project list)
• Alembic migration for session delete_after column

Bug Fixes

• A2A reasoning events visible in frontend (delta_status tracking)
• SSE stream kept open across SDK continuation turns
• ToolInvocation TypedDict argument extraction + sandbox FOWNER
• Council text doubling (dual accumulation of deltas + full message)
• Image coroutine bug (get_storage().read() not awaited in 5 locations)
• A2A client stale URL when sandbox recycled

Config & Infra

• AgentSettings: A2A billing strategy, multipliers, inner loop mode config
• SandboxSettings: SANDBOX_PROVIDER=docker, host config
• StorageSettings: STORAGE_SERVE_BASE_URL for proxied URLs
• CreditsSettings: CREDITS_BILLING_ENABLED toggle for self-hosted
• Health endpoint reports A2A inner loop mode and backend
• Alembic migrations: summary_authority column, session delete_after
• Untrack docker/.stack.env.local (contains local secrets)

Testing

• Comprehensive E2E test suite: 32 tests across 11 categories
  (INF, CHAT, IMG, WEB, CODE, SESS, AGEN, XFEAT, HIST, CNCL, A2A)
• Content doubling detection and server error scanning in E2E
• E2E test-cycle prompt for autonomous fix/rebuild/retest workflow
• Unit tests: council billing, A2A turn loop, orphan cleanup
• Repository integration tests
• Latest run: 31/32 PASS, 1 SKIP (OpenAI quota), 0 FAIL

Docs

• A2A billing model design doc with strategy comparison
• Chat A2A inner loop integration assessment
• Inner loop parity assessment and tool bridge gap analysis
• Implementation guide updates

[mdear]

This rebased PR is based on #172

[mdear]

Coming soon : A2A replacement inner loop backends for Chat Mode

[khoangothe]

Hi @mdear, could you split this pr to resolve DockerSandbox first? I think it could be splitted into 3 prs, and we can work with you and resolve each individually! Thank you so much for your constant support

[mdear]

Yes, I can do that. Stand by, I'll split into three progressive PRs, each that can stand on their own, each building on the previous PRs, with appropriate unit and e2e tests accompanying each.

I'm happy to say the local model with CoPIlot inner loop appears to be stable and working well. I haven't done extensive testing yet but the proof of concept is now holding up.

First PR : Docker local and core unit test rewrite (most unit tests were written against the outdated develop branch). I held off writing many front end unit tests due to lack of technical experience, as my expertise is primarily back-end, and these features are heavily backend and sandbox weighted.

Second PR : Agentic A2A inner loop replacement

Third PR : Chat A2A inner loop replacement, council billing foundation, A2A council billing enhancements

[mdear]

This PR has been split into three progressive PRs for easier review. All content from this PR is covered across:

1. feat: local Docker sandbox infrastructure (1/3) #198 — Local Docker sandbox infrastructure (1/3): Docker sandbox runtime, local deploy stack, session lifecycle, frontend, test overhaul (389 files)
2. feat: A2A agent inner loop framework (2/3) #199 — A2A agent inner loop framework (2/3): A2A inner loop strategy, backend registry, billing strategies, adapter server (74 incremental files)
3. feat: chat A2A inner loop, council routing, compaction authority (3/3) #200 — Chat A2A inner loop, council routing, compaction authority (3/3): Chat A2A turn loop, council A2A routing, cross-authority compaction (16 incremental files)

Merge order: #198 → #199 → #200

Closing this PR in favour of the split.

[mdear]

Churn Metrics

Diff	Files	Insertions	Deletions	Net
main → 1_of_3 (#198)	389	+27,707	−69,207	−41,500
1_of_3 → 2_of_3 (#199)	74	+26,596	−37	+26,559
2_of_3 → 3_of_3 (#200)	16	+6,142	−214	+5,928
Total	479	+60,445	−69,458	−9,013