This page reflects the current implementation in the repository and the tested backend behavior. It is intentionally narrower than the long-term product vision.
- Overleaf is used as the IDE surface for the writing workflow
- The LaTeX project is treated as a codebase the agents can inspect, summarize, patch, and reason about file-by-file
- ResearchKit rail entry and right-side panel inside the Overleaf editor
- Message list and message input components
- Capture of active file, selected text, selection range, and cursor line
- Streaming UI fed by SSE responses from the backend
- Per-project conversation persistence
- Resume an existing conversation from the panel
- Start a new conversation
- Clear a conversation
- Conversation list summaries with last-message preview
- Manual project indexing from the UI
- Automatic reindex checks during chat when files are provided
- LaTeX structure extraction
- Abstract extraction
- Venue hint extraction from
\documentclass - BibTeX citation extraction
- Stored project summary and structure map in MongoDB
- OpenAI provider support
- Anthropic provider support
- Custom OpenAI-compatible provider support
- Per-project provider settings
- Encrypted saved API keys at rest
- Provider connectivity test endpoint
- Model discovery endpoint and UI
- Workspace-aware prompting with memory and file context
- Scoped conversation restore and save
- File viewing and editing through
str_replace_editor - Patch generation for frontend review
- Execution-only bash commands through the runner service
- Sub-agent delegation interface
- Implemented research sub-agent
- Literature search through ASTA MCP
- Citation verification flow
- BibTeX generation from discovered papers
- Read-only workspace inspection for context-aware research tasks
- ReAct-style tool loop with iteration limits
- Temporary overlay workspace creation
- Optional baseline workspace copy
- File overlay injection from request payload
- Command execution with timeout limits
- Before/after text snapshot diffing
- Structured changed-file response for patch conversion
- Research workflows are strongest when
RESEARCHKIT_ASTA_API_KEYis configured. - Bash-backed workflows require runner configuration and a valid workspace path.
- Memory summaries fall back to the LLM provider when no abstract is available.
- Request-level config overrides exist, but most users will interact through the saved project config UI.
-
Figure Agent
- class exists
- returns placeholder content only
-
Review Agent
- class exists
- returns placeholder content only
The backend test suite currently covers:
- API routes
- config loading and encrypted secret handling
- main-agent tool behavior
- research-agent tool loop behavior
- runner client parsing
- read/edit guardrails in
str_replace_editor
Current observed baseline: 67 passed.