---
title: "Chat and actions"
summary: "Chat in the TUI and web, driga ask, what Chat can do, the Confirm, Auto and Full auto trust modes, action cards and what is retained."
description: "Ask Driga about your sessions through your installed Claude Code CLI, and approve the actions it proposes with trust modes and action cards."
read_when:
  - "You want to ask questions about your sessions in plain language."
  - "You want to know what Chat may run without asking."
---

Chat answers questions about your sessions using your installed `claude` CLI
and its existing login. Driga has no API key and never calls a model in the
background: Chat and [summaries](#summaries) run only when you ask. Codex and
Copilot are listed as coming.

## Open it

- **Terminal UI:** press `c` anywhere for Chat with the current selection.
  Enter sends; Shift+Enter (or Alt+Enter) adds a line; Esc closes the panel and
  keeps the draft. Alt+0, Alt+1 and Alt+2 switch tabs while typing.
- **Web:** press `c` on any page, or open the Chat page, to dock Chat on the
  right. Context chips describe the current page, selection and machine; remove
  a chip to leave it out.
- **One shot:** `driga ask "which agents are waiting on me?"`. `driga ask` is
  read-only and has no action cards.

Session references in answers select the session. Tool calls appear in a
collapsed **Tools used** list, and replies are marked **answered by Claude
Code**.

## What Chat can do

Chat reads through Driga's read-only tools. With approval it can also
acknowledge inbox items, jump to a session terminal, kill an agent or orphan,
clean up a candidate, refresh project summaries and answer Claude permission
prompts. It also has Claude's built-in shell and file tools. Actions are
local-only and keep the CLI's process, cleanup and jump safety checks. Chat
never types into a live agent terminal.

## Trust modes

Each conversation starts in **Confirm**. Change mode with the web header
buttons or `m` in the TUI's navigate mode.

| Mode | Runs without asking | Asks |
| --- | --- | --- |
| Confirm | Driga read tools; reads inside the working directory | Reads elsewhere and every other call |
| Auto | Confirm's list, plus inbox acknowledgments, jumps and summaries; Edit/Write only inside a git project working directory that is not your home, the Driga data directory or an ancestor of either, and never on hidden paths or agent instruction files | Everything else |
| Full auto | Everything | Nothing |

Answering a permission prompt always asks in Auto.

**Full auto** needs confirmation, is shown in red with its remaining time, and
lasts at most 30 minutes. It ends when you switch it off, close the
conversation or restart the broker, and is never persisted. Only you, through a
client, can change modes or decide cards; model text, files and tool output
cannot.

## Action cards

A card shows the tool, its exact input, the working directory, status, who
decided and the result.

- **Web:** click **Allow** or **Deny**.
- **TUI:** in navigate mode `y` allows and `n` denies the oldest pending card;
  while typing use Alt+y and Alt+n. The status line says **action waiting**.

Stop denies pending cards before stopping the turn. Unanswered cards expire
after ten minutes.

## Where conversations live

The collector owns the Chat broker. Without the service, the client you open
runs a private broker and labels it **conversations end with it**. With the
service, conversation IDs and capture-filtered messages persist for 30 days;
interrupted work is marked interrupted after a restart and never replayed to
the agent.

Your question is scrubbed and sent to Claude. Context and tool results pass the
capture policy first. Under the default `capture_text = "titles"` Driga keeps
no chat message bodies; `snippets` keeps scrubbed 200-character snippets.
Claude's own session history is managed by Claude and is not controlled by
Driga's capture policy.

## Summaries

```sh
driga projects summarize <project>...   # or --all
```

Summaries need `summaries.enabled = true`. They send the project name,
branches, dirty/ahead/behind counts, the last five commit subjects, recent
session titles and, only with `capture_text = "snippets"`, handoff text. They
never send transcript lines, file contents or paths, and run with a $0.05
budget cap.