Release date: October 26, 2026
Overview
OAgent V2.2.0 is designed for complex database O&M scenarios, with a focus on enhancing inspection, log analysis, and long-running task collaboration capabilities. This version introduces member token usage rankings, native task continuity, OCP log analysis, and a closed-loop inspection plan feature; it also optimizes tool discovery, context usage, multi-cluster inventory, SQL tuning knowledge loading, and the report evidence, streaming response, and standalone knowledge base experience when OCP/OAS is not available.
New features
Member token usage rankings
- Added a member token consumption ranking on the dashboard, which calculates the total token consumption for members within the selected time range.
- Supports viewing details such as input, output, cache read, and cache write, and can expand to view all members.
- When session details are not fully attributed to a member or statistics are incomplete, the page clearly indicates confirmed subtotals, avoiding mistaking incomplete data for a complete cycle's statistics.
Native task continuity
- Adds structured records for task context, work objectives, constraints, success criteria, current progress, and next steps for complex tasks.
- Supports resuming unfinished work within the same task, reducing the need to repeatedly describe backgrounds and re-collect information in lengthy conversations.
- Task context records are bound to tool evidence, preserving verified facts and their sources, facilitating task recovery and subsequent report delivery.
- Supports checking task status, saving checkpoints, viewing task metrics, and assessing task context quality.
OCP log analysis
- Introduces OCP log analysis capability, supporting log retrieval by specified log type, host, and query conditions.
- Supports log sample location, paginated re-reading, and evidence citation. Analysis conclusions are based solely on evidence from the same source returned by the current query.
- Limits the number of queries, response size, display quota, and evidence save scope to prevent large logs from overwhelming the context.
- Log evidence is retained during compression, restoration, and follow-up inquiries. When evidence is inaccessible, the budget is exhausted, or the query fails, clear limitations and next-step suggestions are returned.
Inspection plan closed loop and real-time progress
- Automatically checks OCP self-healing plans associated with alerts during inspections, and includes the plan status in the inspection summary and report.
- Supports executing eligible self-healing plans after user confirmation. The real-time status is automatically checked after execution.
- Execution is reported as successful only when the plan status is COMPLETED and the execution result is successful. The true status is retained if it is In Progress, Failed, or Unknown.
- Inspection metrics support batch collection, and the tool execution progress is displayed in real-time on the page, reducing wait time during long operations.
Experience optimization
Inspection, diagnosis, and reporting
- Optimized the single-cluster inspection process, reducing duplicate rewriting and repeated querying to improve inspection speed.
- Reports include a health conclusion, evidence association, and handling suggestions based on observed facts. Content that cannot be verified is clearly marked as unknown.
- Retains metric accuracy, SQL values, execution plans, and alert facts. The conclusions in the report are consistent with the actual returned evidence.
- Supports batch inventory across clusters and tenants, simplifying the payload returned by OCP/OAS tools and reducing context truncation and thought loops in multi-cluster scenarios.
- Verifies facts and evidence scope before report delivery. If report generation fails, retains the collected facts and provides hints for further troubleshooting.
Stress test reports and SQL results
- Optimized the generation and delivery of performance testing reports. Metrics such as CPU, load, and throughput are collected for clusters, tenants, and complete time windows specified by users. Inspection data is no longer used to replace performance testing data.
- Distinguished the boundary between "observed load" and evidence of whether the performance test was actually initiated. When data is missing or execution is incomplete, it is clearly marked as unassessed, rather than treating partial results as a complete success.
- Streamlined fields before inspection results of slow SQL and top SQL enter the model while retaining the top 5; by default, monitoring curves retain window statistics and curve data is passed only when users explicitly request charts.
Tool discovery and context
- The tool catalog supports on-demand discovery. The model no longer loads all tool definitions in each new session, reducing the context footprint in the initial round.
- Tool search supports Chinese and English queries and returns matching results based on tool name, description, and capability information.
- Tool definitions, parameters, and return results are loaded on demand. When a tool call fails, clearer parameter and retry prompts are provided.
- Streamlined the context for monitoring, logging, and reporting, retaining only the facts, evidence, and delivery information truly needed for the current task.
- Purged irrelevant skills and tools, reducing unrelated capabilities from entering the context of new sessions and lowering the likelihood of accidental tool selection and waiting times.
SQL tuning and OceanBase knowledge
- Stably integrated SQL tuning knowledge into OceanBase Agent, which is automatically enabled after an upgrade.
- SQL tuning suggestions are generated based on collected SQL, DDL, execution plans, and index dependencies, with applicable scope and validation basis retained.
- Optimized the transmission of SQL analysis pagination, risk classification, numerical metrics, and chart data to avoid losing key facts when simplifying the returned content.
- Enhanced the analysis and optimization capabilities for slow SQL, top SQL, and execution plans, retaining key SQL numbers, risk classification, and applicable scope.
Knowledge base status
- Displays the vectorization status of the knowledge base, including statuses such as No Vector Model Configured, Processing, Vector Search Available, and Failed.
- Supports re-embedding after vectorization fails, facilitating the recovery of the knowledge base's retrieval capability.
Dialogue and service availability
- When OCP/OAS is not connected, OAgent automatically downgrades to pure knowledge base and general analysis mode: you can still use the built-in knowledge base and analyze SQL, DDL, execution plans, logs, or ASH provided by users. In this mode, OCP/OAS tools are not called, nor is installation or connection to these products required.
- If the connected OCP/OAS service is temporarily unavailable, steps that depend on this service are skipped. The system continues to use currently available capabilities to answer questions and resumes streaming responses.
- The model configuration process automatically handles interface compatibility parameters for models such as DeepSeek. No manual modification is needed in the background after configuration is complete.
- Retains successfully saved configurations on the frontend after a restart or failure recovery. Provides clear feedback on persistent results if saving fails.
- Optimized the process for task interruption, service restart, and dialogue resumption. After recovery, the correct running state and task ownership are retained.
- Model selection takes effect per session. After switching or saving the default model, new dialogues use the correct model, and existing sessions are not accidentally switched.
- Optimized the dialogue input area, tool execution status, and report delivery feedback. After a task is aborted, the tool bar ends correctly and no longer continuously displays a loading status.
- Unified the layout, font size, spacing, color scheme, and component styles across main pages, and improved the display states for loading, empty data, error prompts, and long content.
- Optimized the presentation of thought processes, tool calls, and intermediate states to reduce interference from complex process information when reading final results.
- Hidden navigation entries for pages without accessible content to reduce instances of entering pages with no content or no permission.
Bug fixes
Domain |
Fixed issues |
|---|---|
| Inspection and Reporting |
|
| OCP Log Analysis |
|
| Plan Execution |
|
| Dialogue and Tasks |
|
| Tools and Context |
|
| User Interface and Deployment |
|