fix: research node uses actual repo tools instead of hallucinating answers"

[simonabler]

Goal

The research_node currently sends the user'"'"'s question to the planner model with zero file access or code context. It reliably produces hallucinated answers for questions like "what does auth_middleware do?" or "where is X used?". Fix it to use the code intelligence, search, and file-reading tools that already exist and are wired into the coding workflow.

Current behavior

def research_node(state: GraphState) -> dict:
    prompt = "Answer the user'"'"'s request in analysis mode only.\n\n" + state.user_request
    answer = _invoke_agent("planner", [HumanMessage(content=prompt)])
    # ^ No tools. No repo context. Planner just guesses.

Target behavior

Research queries get the same context and tools as a coder, minus any write/git tools.

What changes

research_node rewrite

1. Ensure context is loaded (context_loader_node result available in state)
2. Build prompt including:
   • _format_intelligence_summary_for_prompt(state) (call graph, smells, dep graph — already implemented)
   • _format_context_summary(state.repo_facts) (tech stack, structure)
3. Invoke agent with read-only tools: read_file, list_directory, search_in_repo
4. Stream response to UI as normal

Read-only tool set

RESEARCH_TOOLS = [read_file, list_directory, search_in_repo]
# Explicitly excludes: write_file, git_command, run_shell, run_terminal

Code intelligence trigger

If state.call_graph is empty (intelligence not yet run), trigger code_intelligence_node first, then research. This ensures call graph is available for structural questions.

Example queries that must work correctly after this fix

• "What does planner_plan_node do?" → reads the function, explains it
• "Where is GraphState used?" → searches repo, lists files
• "What are the main entry points of this codebase?" → uses context listing + call graph
• "Find all places where we call the OpenAI API" → searches for API calls

Acceptance criteria

• research_node uses read_file, list_directory, search_in_repo tools
• Code intelligence summary injected into research prompt
• If call graph is empty, code_intelligence_node runs first
• Response streams to Web UI in real-time (same as coder)
• No hallucinated file paths or function names in answers
• Tests: research with mocked tools, verify tool calls are made

Dependencies

None — all required tools and intelligence already exist.