Chat and models

How a conversation works in ws, how a model gets chosen, and what happens to long conversations. For the commands to get ws running, see Start here.

Projects and conversations

A project holds conversations. The sidebar’s “New conversation” starts a chat in your most recently active project, or creates one called “General” if you have none; a project can also be a code project, which adds a sandbox. Every turn is an agent run: one or more model calls, plus any tool calls the model asks for. Runs are saved as they go, so they survive a restart of the server; a run whose process stopped is picked up and finished by the worker.

Choosing a model

The model picker has three groups:

• Routing: aliases defined in your policy files (the shipped ones are best, cheap, local and code), plus the built-in auto, which is the default. A routing name lets ws pick the endpoint for each request from your policies, filtered by what the request needs (tools, images) and by which endpoints are healthy.
• Local: endpoints on your own hardware.
• Hosted: endpoints from providers such as Anthropic, OpenAI and OpenRouter.

An endpoint that is down is shown disabled with “(down)”; an endpoint an admin has disabled is not listed.

The brain icon next to the picker asks for extended thinking. It has an effect only on Anthropic-protocol endpoints that support it; local and OpenRouter endpoints ignore it today. It is not remembered and is off again when you reload the page. Which endpoints exist, and what each can do, comes from config/endpoints.yaml and from the Admin page; see Admin and Configuration.

Attachments

“Attach” accepts images, PDF, plain text, Markdown, JSON and CSV files. What the model sees depends on the endpoint. The router only filters on images (it picks a vision-capable endpoint) and on tools. A PDF is read only by Anthropic-protocol endpoints; an OpenAI-compatible endpoint gets a short placeholder saying the file was omitted, and text and JSON files are included as text there. A file whose type the browser leaves blank can be omitted the same way.

Tools in chat

A chat turn can call these built-in tools: create_artifact and update_artifact (see Artifacts), web_fetch (public URLs only; private, loopback and tailnet addresses are refused), read_blob (to page through a long tool result that was stored outside the conversation) and ask_user. Tools from MCP servers listed in config/mcp.yaml appear as mcp__<server>__<tool>. Code projects are offered every tool the worker has, including these and the sandbox tools. Each tool has a policy that decides whether it runs on its own or asks you first; see Approvals.

Long conversations

ws never edits what is stored. When a request would be larger than the compaction budget (WS_COMPACTION_BUDGET_TOKENS, 60,000 tokens by default), it sends a stored summary of the older turns plus the recent turns instead, and drops the oldest turns for that one request if it is still too long. A background job writes a fresh summary with a cheap model. Your full history stays as it was.

Cost and limits

Every model call is written to a usage ledger with its tokens, cost and the routing decision. The owner can set budgets that block a call or switch it to local endpoints once a limit is reached; see Admin.

Checked against the code at master 3364287.