Skip to content

02-chat-mode.md

Latest commit

 

History

History
105 lines (66 loc) · 3.79 KB

02-chat-mode.md

File metadata and controls

105 lines (66 loc) · 3.79 KB
Raw
type tutorial
topic getting-started
audience
human

02 — Chat mode

reyn chat is the lowest-friction way to see Reyn in action — talk to it, watch it route your request to a skill, and read the answer back. No authoring required.

This tutorial uses only the auto-created default agent. Multi-agent setup is a later topic; pointers at the end.

Start a session

reyn chat

The first time you run this, Reyn auto-creates a default agent under .reyn/agents/default/. Subsequent runs reuse it.

You'll get a > prompt.

Type a turn

> summarize the README of this project

What happens:

  1. The chat router classifies the intent.
  2. It picks the best-matching skill — for a "summarize README" request, that's typically read_local_files followed by direct_llm, or read_local_files alone if the model already summarises inline.
  3. The skill runs and the answer prints below the prompt.
  4. The session stays open. Type the next turn.

Try a few:

> what is this project about?
> what's in src/reyn/?
> say hi in three languages

(To see the catalogue of skills the router picks from, exit and run reyn skills from the shell — chat conversations don't enumerate them.)

Each turn is logged under .reyn/agents/default/.

Exit

Ctrl+D or /quit ends the session. Re-running reyn chat resumes the same agent — its memory and history persist.

Web UI (optional)

While the TUI is running, you can also use a browser-based interface. In a second terminal:

reyn web

Then open http://localhost:8080. The Web UI connects to the same agent and shows the same session — richer rendering, better for longer conversations or screen sharing.

Stopping reyn web (Ctrl+C) doesn't affect the TUI session, and vice versa. See How-to: Chat and Web UI for options.

Slash commands

Lines starting with / are intercepted as control commands, not routed to the LLM:

  • /list — show any pending user prompts (ask_user / permission asks).
  • /answer <id> <text> — answer a pending prompt from /list.

These two are the only ones you need for default-mode use. More slash commands (/agents, /attach, /rewind, …) become useful once you have multiple agents or want to time-travel a conversation — see reference/cli/chat when you get there.

Memory is automatic

The router reads memory on every turn (no setup required) and writes durable facts back when it detects them. Two layers:

  • Shared — .reyn/memory/ — facts visible to every agent.
  • Agent — .reyn/agents/default/memory/ — facts scoped to this agent.

You can inspect what's been remembered:

reyn memory list
reyn memory show <slug>

See concepts/memory for the full model.

What's actually happening

The OS doesn't know about "chat". It just runs the chat router, which picks a skill (or, in multi-agent setups, a peer agent) to delegate to. This is the same composition pattern any of your own skills would use (P7 (principles doc removed)).

What you learned

  • reyn chat attaches a REPL to the auto-created default agent.
  • Each turn goes through the chat router, which picks a skill and runs it.
  • Memory is two-layered (shared + agent) and read/written automatically.
  • /list, /answer are the slash commands you need at this stage.

Where to go next

You've seen Reyn deliver value as a chat agent. From here:

  • Multi-agent (later): How-to: Build an agent team walks through reyn agent new, role-specific allowlists, and /attach. Background reading: concepts/multi-agent, concepts/topology.
  • The why: concepts/principles (principles doc removed).