DevAgent Studio

##Project Highlights - ** Multi-agent collaboration rather than a single Prompt**: Orchestrate Planner, project analysis, code review, RAG, supervision and reporting nodes through LangGraph. - **Harness Runtime governance **: Unify task context, event timeline, artifacts, policies, manual review, persistence and recovery execution. - ** Truly executable visual Workflow**: The front-end drag canvas will be compiled into LangGraph workflow execution, not just a display diagram. - **LLM observability and governability **: You can view the Prompt version, call Trace, token/cost, fallback, A/B Test, and configure the model by Agent. - ** Safe and governable Skill plug-in system **: Skills have contract verification, privilege grading, strict approval, version snapshots, dependency detection, test cases and execution logs. - ** Plug-in market is compatible with external sources **: Supports built-in packages, local paths, URLs, GitHub style sources, and external `SKILL.md` automatic conversion. - **Docker code-based Skill Sandbox **: Third-party code can run in temporary containers of blocked, read-only, time-limited resources. - ** Controlled long-term memory **: Conversations are first extracted into candidate memories through LLM/rules, and then deposited into RAG through quality scoring, conflict detection, manual confirmation, life cycle strategies, and scope authority boundaries. - ** Oblivion decay rather than infinite accumulation **: The stable rule is retained for a long time; ordinary preferences enter a 90-day review, mark them as 'expired' after expiration and exit from retrieval, but the audit record is retained and can be refreshed, updated, or deleted by the user to avoid outdated preferences and outdated decisions continue to interfere with the Agent. - **RAG governance and evaluability **: The knowledge base supports incremental indexing, document versioning, ACL permission filtering, BM25/vector hybrid retrieval, optional LLM Rerank and Chunk Gold Set. - ** Colorful/clear dual-theme UI**: The top can be switched between the "colorful" theme with yellow, green and blue powder colored environmental light beams and the "clear" theme of SVG refraction water glass; the clear mode supports edge refraction, floating shadows, click the Ripple and Mode buttons to switch continuously in the liquid state, and save user choices locally in the browser. ###First use process 1. When the application's private workspace is completely empty for the first time, the system will automatically generate the sample project `/workspace/DevAgent-Studio`. Only when analyzing your own project, you need to first put the project directory into the `workspace` folder of the application data, and then fill in `/workspace/<project name>` on the page. See the attached figure. 2. Enter `LLM` and verify that "LLM configured·model name" is displayed at the top. If "LLM not configured" appears, please first check the API Key, Base URL and model ID in the installation configuration; if not configured, only deterministic analysis and fallback can be used, and a complete model report cannot be obtained. 3. Enter Run and select Collab. Fill in `/workspace/DevAgent-Studio` for the project path, and the goal can be filled in: `Please analyze the project structure, core modules and data flow, list the three highest priority improvement suggestions, and provide a verification method for each suggestion.` Fill in '80' for the number of documents and turn off "manual review". 4. Click "Run Multi-Agent Collaboration". The execution event will display Planner, Project Analyzer, Code Reviewer, RAG Processor, Supervisor, and Reporter in order; wait for the current state on the right side to change to `completed`. The number of documents and slices in project analysis and RAG results should be greater than 0, otherwise the project path and workspace content should be checked first. 5. Enter 'Report' and open "Final Report","Architecture Mentor Perspective","Mermaid Chart" and "Governance Suggestions" in turn to confirm that all four types of content can be displayed. After refreshing the page, the current report column will still be restored through the `reportTab` parameter in the URL. 6. Enter 'Question' and enter: 'Break the highest priority recommendation into a three-step plan that can be completed in one week, and add acceptance criteria for each step.' The page should return answers combined with the task and report you just completed. At this point, the first core processes of project scanning, multi-agent collaboration, RAG, reporting and context questioning have been completed. ###Daily usage - ** Select execution mode **: Select the mode by task in Run.` Agent`is suitable for quickly completing project analysis,`Collab`is suitable for complete planning, analysis, review, RAG and reporting collaboration,`Workflow`executes the current canvas, and`Planner`generates the process based on goals;`Tool`and`Knowledge`are used for tool invocation and knowledge retrieval respectively. - ** Tracking and review tasks **: Check node status, Agent output and tool call in "Execution Events" during runtime; go to 'History' to reselect tasks and view reports. When it is necessary to continue to refine conclusions, go to 'questioning' and asking questions around the current task without having to rerun the entire analysis. - ** Read and share results **: Switch to "Final Report/Architecture Mentor Perspective/ Mermaid Chart/Governance Suggestions" in `Report`. The 'reportTab' parameter will be written to the column, and the same perspective can be restored after copying the current link or refreshing the page. - ** Orchestrate reusable processes **: Drag and drop nodes and connect the execution order in "Orchestration", first click "Verify Workflow", and then save or run. The Skill node is approved according to `skill_code + agent_code`, and the MCP node is approved according to `server_id + tool_name + agent_code`; the approval identity must be consistent with the actual execution identity. - ** Maintain project knowledge **: Use the RAG Processor in Collab to incrementally process projects, query `project-memory 'in reports or knowledge-related entrances, save knowledge notes, and check the retrieval quality through ACL, Gold Set and evaluation results. - ** Governance model calls **: Check the number of calls, Token/Cost, fallback, Agent/Model/Prompt summary and single Trace in LLM; use Prompt A/B Test to compare the two prompt versions, and then select a version based on Trace, time consumption and output quality. - ** Use MCP and Skills**: After saving the server in MCP, click Discover to use the specific agent_code approval tool, then make a test call and view the log. View permissions, risks, input and output schemas, version snapshots and execution logs in Skills; third-party code skills should be reviewed first and then approved. - ** Management plug-in **: Preview the source, TRUST, permissions and manifest in Market, and then install the built-in resource pack, ZIP, plugin.json or separate SKILL.md.