Chat

Chat is the default mode of Babel: babel "<task>" starts a lightweight, multi-turn conversational coding loop with live tool access. It deliberately loads no orchestrator and no pipeline stages — the path of least resistance for daily coding.

What a chat turn looks like

Inside the TUI you will see Babel’s tool activity as it happens: file reads, searches, edits, and verification commands, each subject to your approval settings. You approve or deny at the moment a tool needs it (just-in-time approval).

Commands inside a session

/help is the canonical list. Daily commands:

/model         list or switch models
/mode          chat / plan / deep
/diff          inspect the latest changes
/resume        continue a persisted conversation
/permissions   show or set the approval profile
/checkpoint    mark a restore point
/restore       return to a checkpoint
/cost          show session cost
/cancel        cancel the current activity

Chat from scripts (headless)

The same engine runs non-interactively for CI and scripting:

node .\babel-cli\dist\index.js chat --headless "Summarize the failing test"
node .\babel-cli\dist\index.js run "Summarize the failing test" --mode chat --headless

Under headless/non-TTY runs, missing authoritative verification hard-blocks after gate strikes — weak verifier greens are not soft-allowed.

What chat does not do

It does not load the governed pipeline (that is Deep).
It does not produce a reviewable plan artifact first (that is Plan).
A model’s “I’m done” is not accepted as success: completion is gated by the harness honesty policy. See Verification.

Boundaries

Chat runs under the same execution-profile isolation rules as other modes. See Execution profiles before any mutating task.