Conversations

A conversation is one task with the model: its timeline, its settings and the tools it may use. A preset gives a new conversation its settings and opening messages, and the Conversation settings pane changes them as you go.

Start a conversation

New task at the top of the sidebar, or ⌘N (CtrlN on Windows), opens a conversation in the selected project from the built-in mework preset. + on a project row uses the project's default preset, set with ⋯ → Default conversation preset on that row; without one, the settings you last used in the project; without those, mework.

A new conversation takes the preset's settings and, when the preset has a template, its opening messages: with mework, the first card is an engineering-practice system prompt.

Drafts. A conversation is a draft until you send its first message, and there is only one draft. While it is unsent, New task and + return to it with its text and settings and move it to the project you asked for; a preset applies only to a fresh draft.

Branch and fork. Branch from this message and Fork conversation (Editing the context) start a new conversation from part of this one's history, with the same settings; a fork also keeps the tool lock.

Conversation settings

Open them with ⋮ More options → Conversation settings in the conversation header, or ⌘⇧, (CtrlShift, on Windows). Each change is kept as you make it and reaches the model from the next turn; the security level and the Plan switch apply at once, even mid-turn.

Besides Tools and Advanced tools (below), the pane has Skills, MCP and Hooks (Skills, MCP servers, Hooks), Agent roles (Agent roles) and Conversation presets (Presets).

Skills, MCP and Hooks list what Mework finds in ~/.mework on this computer and in the .mework folder of each of the conversation's workspaces, which for a WSL or SSH workspace is read on that machine; the conversation can select from all of them. Each page has a Global section first, then one section per workspace, headed by the workspace's absolute path with its machine (such as "SSH: devbox") and, when the conversation has more than one workspace, its number. The folder button on a section's heading, Open the global config folder or Open the config folder of {path}, opens that level's folder, creating it if needed; for a WSL or SSH workspace it opens that machine's folder in the Files pane. A workspace whose machine could not be read says so in its own section. The selected count at the top of the page covers every section. In a preset window, there is a section for every workspace that has entries. A selected entry gone from disk shows Dangling, and every run fails, naming it, until you uncheck it.

Note Delete on a Skills, MCP or Hooks row removes the item from disk — the skill's folder, the server's entry in mcp.json, the hook's entry in hooks.json — not just from this conversation.

Choosing tools

The Tools page lists the tools you turn on one by one; Shell shows only the shells a machine in this conversation has. Click a row to switch it; + and − on a group heading switch the whole group. Tools without a row come with another setting, such as web access, memory or Plan; Built-in tools lists which.

The Reviewed badge marks tools whose calls ask for your approval at the Manual level whatever they touch: the shells, write, edit, workflow, agent_spawn, fork and the preview tools that act on the page or read its console, network or screenshots. Tools that never ask carry no badge. See Security levels.

Advanced tools

The tool lock

Providers cache the start of a conversation, declared tools included. Changing what the last request declared throws that cache away, and the next request pays for the whole conversation again, so Mework marks such changes.

Orange: the cache is still warm, for 30 minutes after the last request by default, or the model's Cache lifetime (minutes) in its model properties. Orange rows are:

• a tool, web access or a skill the last request carried, if you turn it off;
• MCP servers, hooks, either memory switch, Load skills on demand, Tool discovery and Tool descriptions, in either direction;
• a web backend Mework runs itself, if its tool went out in the last request.

Turning a tool on is never orange. An orange row still changes, but the first time in a conversation Mework asks This change throws the cache away first. Change it back before the next request and the cache holds.

Gray: the model cannot change tools mid-conversation. On a model without Mid-conversation tools, the tool setup is fixed at the first request; only skills still change, and a Dangling row can always be unchecked.

The lock belongs to the model that sent the last request: select another model and the marks go away; select it again and Mework undoes the changes the lock covers. Agent roles and the security level are never locked.

Presets

A preset is a saved copy of the Conversation settings pane, with an optional template. It holds every setting on the pane, the security level included, but not reasoning effort, Plan mode, attached workspaces or the tool lock.

The built-in mework preset turns on every tool that has a row, with this machine's usual shell (zsh on macOS, PowerShell on Windows), plus web search, both memory switches, Load skills on demand and Tool discovery, at the Manual security level with four agent roles. You cannot edit it, but you can open it, change anything and click Save as new preset.

Apply. In Conversation settings → Conversation presets, click the arrow at the start of a row. The settings are copied in, and later edits to either side do not carry over. The preset's template replaces the timeline; Mework asks first (Replace) unless the timeline is empty or has as many messages as the template applied last.

Save. Click Save as preset, under the pane's page list, and enter a Preset name. Tick Use the current context as the conversation template to keep the timeline as it is, real tool results included, as the preset's template.

Edit. Click a row's settings icon (Open preset). The preset window has the same pages plus Conversation template. Settings changes wait for Save preset; template edits are saved as you make them. Deleting a preset deletes its template too.

Templates

A conversation template is a set of opening messages that a preset or an agent role carries. A preset's template opens every new conversation started from that preset; a role's opens every subagent agent_spawn starts with that role, with the task in place of {input}.

Edit a template on the Conversation template page of a preset or role window. It is a timeline you edit like a conversation: right-click and choose System prompt, User input, Reasoning field, Tool call or Model reply. Tool calls and their results are written by hand and never run. A system prompt at the top is the conversation's system prompt; one placed lower down takes effect from that point.

While the model works

Queue. Sending while the model is busy queues the message. Queued messages are listed above the composer, where you can delete them; when a turn ends on its own, they go out one at a time as new turns.

Steer. Steer into current turn, on a queued message, hands it to the running turn after its current batch of tool calls. If the turn ends first, the message stays queued.

Stop. While a turn runs, the send button becomes Stop generating. Stop ends the turn, the commands it is waiting on and its approval cards, and holds the queue: queued messages wait, unsent, until you send a message again. Subagents, workflows, background commands, dev servers and terminals keep running; stop those from the Tasks pane (⋮ More options → Tasks).

Questions. With ask_user on, the model can ask you questions in a card above the composer. The card never times out; closing it ends the turn.