Chat

The chat command group provides interactive, multi-turn conversations with LLMs. Sessions are persisted in memory per model for the lifetime of the process.

chat send

Start (or continue) a conversation with a model.

padwan-cli chat send "Hello, how are you?" -m gpt-4o-mini

In TUI mode, the command enters an interactive loop — type follow-up messages and press Enter to continue the conversation. Press Ctrl+C to exit chat mode. Messages you type while the model is responding are queued and processed in order.

In CLI mode, the command sends a single message, prints the response, and exits.

After each response, token usage is displayed:

in: 42 out: 128 cached: 0 | session: 170 | mcp: 1 | 2 queued

• in — input tokens for the last request
• out — output tokens for the last request
• cached — cached tokens (if supported by the provider)
• session — cumulative tokens across the conversation
• mcp — number of connected MCP servers (only shown when > 0)
• queued — messages typed while the model was responding, awaiting their turn (only shown when > 0)

Attaching files

In TUI mode you can drag and drop files onto the terminal while in chat mode. The terminal pastes the dropped file's path, which the chat detects and turns into an attachment — a chip appears above the input listing each queued file. Type your message and press Enter to send the files along with it; press Ctrl+U to clear the queued attachments.

• Text files (code, Markdown, JSON, …) are read and inlined into your message, so they work with any model.
• Images are sent to the model as image input — this requires a vision-capable model (e.g. gpt-4o, gemini-2.5-flash, pixtral-large-latest).
• Audio files (wav, mp3, flac, ogg, aac, aiff, m4a) are sent as audio input — this requires an audio-capable model, and format support varies per provider.

If the active model can't consume an attachment, its chip turns amber, a warning explains what will be skipped, and the file is left out of the request (switch models with -m).

You need to type at least a short message to send; dropping a file alone won't send. Drag-and-drop is a TUI-only feature.

Tools and MCP

Chat sessions run without tools by default. Pass --mcp with one or more streamable-HTTP MCP server URLs to expose their tools to the model:

# Single server (public, no-auth example)
padwan-cli chat send "Which datasets cover cycling?" --mcp https://mcp.data.gouv.fr/mcp

# Multiple servers (space-separated)
padwan-cli chat send "Hello" --mcp https://mcp.example.com/mcp https://mcp.data.gouv.fr/mcp

When a connection comes up you'll see an MCP connected notification, and any tool calls render as their own widgets (with elapsed time) inline with the response.

Errors

Failures during a response (bad endpoint, invalid API key, provider errors…) are shown inline as a ✖ message and the chat stays usable — fix the issue or just send the next message.

Thinking tokens

For Gemini models, the chat session enables includeThoughts so reasoning tokens stream into a separate "thought" widget above the answer. Other providers that emit reasoning chunks (e.g. via --stream-thinking on the one-shot path) behave the same way.

Session persistence

Each model gets its own conversation session. Sending messages to gpt-4o-mini and then to gemini-2.5-flash creates two independent sessions. Switching back to a model resumes where you left off.

chat clear

Clear conversation history.

# Clear history for a specific model
padwan-cli chat clear -m gpt-4o-mini

# Clear all sessions
padwan-cli chat clear