Chatting with Aira

The Chat page is a conversational interface where you interact with Aira using natural language. Ask questions, create tasks, generate reports, or request analysis.

How the assistant works

When you send a message, it goes through Aira's multi-agent pipeline:

1. Supervisor — Classifies your intent and routes to the right specialist
2. Specialist agent — Processes your request (or the Assistant handles it directly)
3. Response — The result is saved and displayed

There's no keyword matching. The Supervisor is an LLM that understands intent. Asking "create a task for login page redesign" and "we need to redo the login UI, can you make a ticket?" both route correctly.

The ReAct loop

The Assistant agent uses a ReAct (Reason + Act) loop. When processing your request, it:

1. Reasons about what information it needs
2. Calls one or more tools to get that information or perform an action
3. Evaluates the result
4. Either calls more tools or produces a final response

You can see the assistant's thinking process as it works through complex requests.

What you can ask

CRUD operations

• "Create a task for implementing Google Sign-In"
• "Update the login feature status to in_progress"
• "Delete task TASK-123"
• "Add a new team member named Alex with React and Python skills"

Queries and status

• "What's the status of the current sprint?"
• "Show me all tasks assigned to Sarah"
• "Which features are in progress?"
• "What's my team's workload look like?"

Reports and analysis

• "Generate a standup report"
• "What are the biggest risks right now?"
• "Give me a sprint review summary"
• "Forecast when the dashboard feature will ship"

"Why" questions

• "Why did we prioritize feature X over feature Y?"
• "What insights led to the authentication feature?"
• "What evidence supports this risk?"

Heartbeat management

• "Every heartbeat, check if any features lack tasks"
• "Add a monitoring check: alert if more than 3 tasks are blocked"
• "Show me the heartbeat status"

Knowledge Ledger grounding

The assistant grounds its responses in your project's actual data using a Knowledge Ledger retrieval tool. When you ask about your project, the assistant:

1. Queries the Knowledge Ledger with your question
2. Retrieves matching atoms (claims, decisions, requirements, risks) with evidence snippets
3. Incorporates this context into its response

This means answers about your project are backed by evidence from your sources — not hallucinated. If the Knowledge Ledger doesn't have coverage for a topic, the assistant responds normally without refusing.

Available tools

The assistant has access to 32+ tools that call the Aira API:

Read tools (13+): List/get/search tasks, features, sprints, team members, insights, dashboard summary

Write tools (11): Create/update/delete tasks, features, sprints, team members, insights

AI tools (4): Generate reports, analyze risks, suggest assignments, forecast delivery

Heartbeat tools (4): Add/list/remove monitoring items, check heartbeat status

Knowledge tools (1): Retrieve context from the Knowledge Ledger

Chat sessions

Conversations are organized into sessions. The assistant remembers the last 10 messages for context, so follow-up questions work naturally:

You: Show me tasks for the current sprint Aira: [lists tasks] You: Which ones are overdue? Aira: [filters to overdue tasks from the previous result]

Start a new session to clear context and begin fresh.

Chat vs. the main pipeline

The chat assistant is different from the main agent pipeline:

• Chat — Interactive, handles one request at a time, uses the ReAct tool-calling loop, good for CRUD operations and quick queries
• Main pipeline — Handles bulk operations like analyzing all sources, generating features from insights, or planning an entire sprint. Triggered from the UI buttons (Generate Features, Plan Sprint, etc.)

Both use the same underlying data and API. The chat assistant calls the same endpoints your browser does.

Timeouts

Chat responses can take 20–60 seconds for complex queries that require multiple tool calls. The connection stays open for up to 120 seconds. For operations that take longer (like analyzing a large repository), use the dedicated UI buttons instead — those use SSE streaming.