Cloud Native Weekly: HiRoute v0.6, Higress 2.2.6, AgentTeams 1.2.7 & 5 More Project Updates
This weekly report covers November 9–16, 2026 updates across eight cloud-native open-source projects: HiRoute v0.6 adds Omarchy desktop support and agent onboarding guides; Higress 2.2.6 isolates MCP access and AI cache per consumer; AgentTeams 1.2.7 unifies harness images and adds team/task tracing; LoongSuite Pilot 1.17 links sub-agent traces to parent tasks; AgentScope Java 2.0.5 hardens command approval and memory isolation; Nacos advances cross-namespace Skill sharing; RocketMQ optimizes transaction batch threads; and AI Agent Handbook launches an interactive AgentLoop game map.
HiRoute v0.6 Series — Model Routing & Agent Onboarding
HiRoute released v0.6.0, v0.6.1, and v0.6.2. v0.6.0 introduces signed Omarchy desktop packages for ARM64 and x86_64 installable via pacman, reusing existing Linux services or configuring built-in runtimes with explicit upgrade, resume-after-interrupt, and runtime removal. Codex context compression now preserves tool-call/reasoning continuity; fixed routing lets users select reasoning strength, save the route, and restart Codex to load the new model catalog.
Quota management: decision-model calls count against the API key quota; unknown-price calls execute without pricing. From v0.6.1, key spend caps also apply to regular inference, and budget commands work directly. Deleting a decision model first shows which routes reference it — useful for teams sharing keys and controlling model spend.
Stability: home and session pages auto-recover from brief read interruptions; history gaps no longer break tool conversations; Claude routes retain prompt prefixes; Windows detects MSIX-installed Codex Desktop. v0.6.2 adds copyable diagnostic paths on Windows, records which step a decision connection test fails at, and reduces large tool-definition timeouts.
The website now publishes bilingual onboarding guides for Claude, Codex Desktop, Qoder, DeepSeek Harness, Hermes, Pi, OpenCode, WorkBuddy, plus subscription login, API quota fallback, and agent-driven route optimization. Developers follow illustrated steps to set routing rules, connect subscriptions, and complete integration. Platforms: macOS 15+, Windows x64, Linux web UI/CLI, and Omarchy desktop. Ongoing work includes Codex hosted-login model discovery, Codex Responses Lite tool declarations, restoring original settings when disabling an agent, and Windows subscription re-read/agent discovery.
Repository:
https://github.com/higress-group/HiRouteHigress v2.2.6 — AI Gateway with Consumer-Isolated MCP & Cache
Higress 2.2.6 supports MCP protocol version 2025-11-25 with fallback to 2025-06-18. REST-to-MCP path parameters are now escaped per path segment; legacy behavior retained via allowUnsafePathArgs. MCP user access binds to authenticated consumers — no more injecting config via x-higress-mcpserver-config header; ongoing SSE sessions continue on the original Redis.
AI cache is isolated per authenticated consumer so multiple apps sharing a gateway hit their own caches. Cluster rate-limiting and anti-replay plugins gain fail_open (allow) or fail_closed (deny) on Redis failure. Gateway route cache now invalidates only on the actually selected gateway, speeding route distribution in clusters with many Ingresses. Claude streaming state is request-isolated; token usage in segmented streaming is more accurate; instructions passthrough is optional.
Bundled Envoy upgraded to 1.36.12. Controller xDS gRPC now listens only on 127.0.0.1; Service manifests no longer expose plaintext ports — custom clients relying on external plaintext connections must adjust. Console adds AI routing plugin display, Qwen model provider, regex path rewrite, and upstream service search. Upgrading users of header-injected MCP config or slash-containing path parameters must adapt.
REST-to-MCP work continues: custom server.instructions returned in initialize and server/discover; proxy tools retain input schema; AI data desensitization checks restore tool parameters; legacy JSON-RPC response acknowledgment compatibility improved.
Repository:
https://github.com/higress-group/higressAgentTeams v1.2.7 — Unified Harness Image & Multi-Dimension Tracing
AgentTeams 1.2.7 ships with Dashboard v1.4.1. One-click install starts the workspace first, then configures the Manager model in Dashboard; upgrades preserve data and config. OpenClaw, QwenPaw, Hermes, DeepSeek, OpenHuman, and Codex all use the same agentteams-harness image (amd64/arm64); Manager and Worker share this image, with Manager selectable as OpenClaw or QwenPaw.
Universal Tracing is managed by Controller; Dashboard adds a Tracing entry. CLI agt tracing filters by Agent, Team, Session, or Task — enabling team leads to locate a task's execution flow. Tracing is optional, requiring Collector and query backend configuration. Dashboard unifies list styles: model selection now picks Provider then Model; thinking and tool calls grouped; Chinese IME issues fixed. Old agentteams-manager and worker images are no longer defaults — watch image config on upgrade.
Parent/child task progress: child tasks keep own assignee and status; parent cannot complete while children remain open; Dashboard shows parents by default with drill-down; offline migration tool converts legacy data. Parallel work: Codex as Manager, CLI diagnostics for Manager health, runtime metrics & container logs, Harness native approval linked to Dashboard cards, and stop/redirect for direct Agent chat.
Repository:
https://github.com/agentscope-ai/AgentTeamsLoongSuite Pilot v1.17 — Sub-Agent Trace Linking & Multimodal Collection
Pilot 1.17.0 and 1.17.1 attach Qoder CLI native sub-agent traces to the parent task; multimodal summaries attach to the corresponding model or tool call. Users see task decomposition and per-step inputs in a single trace.
Collection scope adds Doubao App work mode (explicit config in listeners.doubao-work.client, default still collects Doubao workspace). Multimodal collection extends to common local video formats for Qwen Office and Qoder, captured when users actively provide video. AgentLoop audit queries and dashboards share the same schema and SQL knowledge, keeping query scope consistent.
Stability: large Qoder Work sessions resume via byte checkpoints; OpenClaw streaming shows chunks promptly; Hermes long-run and Windows file-watch CPU issues fixed. v1.17.1 distinguishes Qoder Desktop App vs CLI telemetry identities — data viewable separately. Ongoing: OpenClaw/Hermes agent-call audit logs, AgentLoop Skill cleanup, checkpoint-based raw log reclamation with disk pressure limits, and Pilot v2 cloud-embedded Rust SDK on a separate branch.
Repository:
https://github.com/alibaba/loongsuite-pilotAI Agent Handbook — AgentLoop "Self-Evolving Islands" Game Map
The Handbook launches AgentLoop's "Self-Evolving Islands" interactive map linking five islands to the Agent continuous-improvement loop: Observe (Mist Lighthouse), Audit (Border Customs), Evaluate (Echo Theater), Experiment (Twin Experiment Island), Optimize (Growth Garden). Observe and Audit are playable now; progress persists in-session, resets on refresh. Game entry separate from code exercises.
Five-part "AgentLoop Data Flywheel Practice" series documents a full end-to-end demo: overall approach, data ingestion via OTel probes, evaluation from golden metrics to rubrics, replay/offline experiments, and automatic experience-asset mining validated by ablation. Teams wanting to use production Agent data for evaluation and iteration can follow stepwise.
Special topics: AI Coding chapter includes a HiRoute developer's first-person account — critical decisions to flagship model, routine work to economy model, same acceptance criteria for both. Jev topic adds an article on replacing generative LLM-as-a-Judge with decision models, adapted from Hamel Husain & Antaripa Saha's "How To Use Decision Models for Evals" plus author practice notes warning against default thresholds or raw probabilities. Personal Agent topic covers DeepSeek SuperAgent analysis and team explorations.
Repository:
https://github.com/aliyun/ai-agent-handbookAgentScope Java v2.0.5 — Maintenance Hardening: Command Approval & Memory Isolation
AgentScope Java 2.0.5 is a 2.0.x maintenance release — no new modules, no public API removals. Security: when Shell allowlist non-empty, commands with variable expansion, pipes, or redirections require approval; Skill path traversal blocked. Memory reads isolated by namespace; A2A cancel stops remote streams. State, cancel, and tool-parameter handling more robust; K8s and E2B sandbox issues fixed.
Upgrade notes: default models changed to claude-sonnet-5-5 and gemini-3.8-flash — pin modelName for fixed behavior. File queue uses new layout — upgrade consumer instances first. Namespace memory users must migrate root memory files into namespaces. Structured output no longer forces strict mode. JDBC extension key-column definition changed — new tables adopt automatically; existing tables need manual migration.
Ongoing: recoverable Durable runtime simplifies Session/Thread/Turn interfaces, compatible with 2.0 session state; separates persisted events from optional real-time increments; on reconnect restores history first; provides Claude Managed Agents and AG-UI adapters. Per-user capability scoping — model sees/invokes only tools, tool groups, and Skills granted to that user — fits one Agent serving multiple user classes. Also advancing: A2A input-required on tool-confirm wait, OpenAI Responses document input, Lettuce Redis support, custom AG-UI MVC handler.
Repository:
https://github.com/agentscope-ai/agentscope-javaNacos — Cross-Namespace Skill Sharing via Agentic Resource Hub
In the Open Source Summer feature branch, Nacos advances Agentic Resource Hub Skill support: share, discover, read, download Skills; import to Hub or cross-namespace copy preserves provenance; console integration and feature flags added; later extending to other resource types. Teams reusing Skills across teams/environments gain unified management and traceability.
Console Copilot multi-model connection: declare multiple under nacos.copilot.connections.*; validated on load; first available chosen in order; falls back to existing settings if unconfigured. Same branch also progresses file-based Raft log storage.
Mainline and 2.x maintenance fix MCP registration, Istio MCP initial push config, MySQL config list pagination reading large fields, Oracle namespace creation, subscription-recovery blocking cache discovery; add Native Image metadata; adjust 2.x Derby connection timeout.
Repository:
https://github.com/alibaba/nacosRocketMQ — Transaction Batch Thread Idle Wait Optimization
RocketMQ improves the transaction message batch processing thread. Broker batches transaction-op messages per queue; thread skips queues with no pending ops; when all queues empty, waits configured interval, yielding CPU instead of pinning a core. Businesses using transaction messages — especially with clear peak/off-peak traffic — free Broker CPU for other tasks during low periods.
Repository:
https://github.com/apache/rocketmqSigned-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
