Cortask

Interactive Chat

The cortask chat command drops you into a persistent REPL where you can talk to the agent, invoke tools, and see streaming responses directly in your terminal.

Starting a session

cortask chat                       # uses current directory as workspace
cortask chat -w /path/to/project   # explicit workspace path

If the path has not been registered as a workspace yet, the CLI registers it automatically and prints a confirmation line.

On startup the REPL displays the active workspace path, the current provider and model, and a short help hint:

Workspace: /path/to/project
Provider:  anthropic (claude-sonnet-4-5-20250929)
Type a message, /help for commands, /quit to exit.

The provider and model are determined by config.yaml. The default model is claude-sonnet-4-5-20250929 unless a different model is pinned via cortask providers default.

How interaction works

Each REPL session creates a unique session ID (cli_<timestamp>). Messages within the session accumulate in the session store, so the agent retains context across turns until you exit.

Type a message at the you> prompt and press Enter. The agent responds under the cortask> prompt. If the agent calls tools during its turn, you see inline annotations:

you> List the TypeScript files in src/

cortask> [tool: list_files] -> src/index.ts, src/config.ts, ...
Here are the TypeScript files in your src directory...

Tool calls show the tool name in brackets, followed by a truncated preview of the result (up to 200 characters).

Streaming output

Responses stream token-by-token to the terminal. The CLI subscribes to the agent runner's runStream generator and writes each text_delta event to stdout as it arrives. There is no buffering -- you see text the moment the provider emits it.

Tool call events interrupt the text stream with a [tool: name] marker, followed by the result preview on the same line once the tool finishes.

Slash commands

Inside the REPL two slash commands are available:

CommandDescription
/helpPrint the list of available commands
/quit or /exitEnd the session and exit

Any other / input prints an "unknown command" warning.

Skills and tools

The chat REPL loads the full tool set: all built-in tools, the cron scheduling tool, the artifact tool, and every eligible skill from the bundled, user, and configured skill directories. Skills are resolved using the same logic as the gateway -- SKILL.md files are loaded, eligibility is checked, and tool definitions are built before the first prompt.

One-shot alternative

If you want a single-prompt, non-interactive run:

cortask run "Summarize the README in this project" -w /path/to/project

This executes the prompt, prints the response, and exits. It uses the same agent runner and tool set but does not load skills.