## **Background** Duo Chat can already hand a job to an ambient Developer session through natural language and approval ([**Run Foundational Flows in Chat**](https://gitlab.com/groups/gitlab-org/-/work_items/20484)). The slash menu work for `/new` and `/compact` is putting commands in the prompt ([**Discoverable chat actions**](https://gitlab.com/groups/gitlab-org/-/work_items/23059)). Duo CLI is exploring `/goal` as a named job with a success check and keep-going across turns ([**spike**](https://gitlab.com/gitlab-org/editor-extensions/gitlab-lsp/-/work_items/2613), [**MR**](https://gitlab.com/gitlab-org/editor-extensions/gitlab-lsp/-/merge_requests/3639)). This epic covers adding `/goal` to the Duo Chat slash command menu so that users can name a job and Duo can own continuing the job until it's done. ## **User problem to solve** As a Duo Chat user, I need Duo to keep working toward a job with a finish line I can check, so I am not the continue button and I am not treating one “I think I’m done” turn from the model as proof the work finished. Today a long ask dies at the end of a turn unless I continue with a new request. Soft prompts like “don’t ask until you’re done” do not start the next turn. They only ride inside one message. Natural language can hand work to a Developer session, but that path is easy to miss and does not teach a habit from the input. What is missing is a discoverable command from the prompt that names the work, makes the finish line clear, and keeps Duo working until that line clears or I stop it. ## **Proposal** Surface `/goal` in the Duo Chat web slash menu so users can find it from the input, name the job and how they will know it is done, see that the goal is active, clear it, and get real keep-going rather than a labeled one-shot prompt. Reuse the chat to Developer handoff and the CLI `/goal` shape to Duo Chat. Guide users toward a concrete change and a checkable end, and away from one-shot questions. Design for discovery and result states to follow. This epic covers web Chat only and does not ship IDE `/goal` or other flow slash commands. Done when a user can discover `/goal` from the prompt, set and clear a goal with a visible active state, get keep-going without babysitting for the path we lock, and we can measure that the command ran in Duo Chat in the GitLab UI. ## **Observable outcomes** * From the chat input, the user can find `/goal` without memorizing a hidden string or leaving the prompt to hunt controls. * After `/goal` with a task (and a success check if we ship that shape), a standing goal is active and the user can see that state. * Duo keeps working toward that goal without the user entering a continue prompt. * `/goal` with no args shows status or clear usage. `/goal clear` ends an active goal. * After a weak goal (vague ask, one-shot question), the user gets guidance toward a better shape rather than a silent failed long run. * Invocations of `/goal` in Duo Chat in the GL UI are measurable by command name and surface. ## **In scope** * Duo Chat web UI (agentic) * Discoverability of `/goal` from the chat input * Setting, showing, and clearing a named goal with a checkable finish line * Keep-going behavior tied to that goal (not menu-only) * Copy and examples that teach what belongs in a goal * Telemetry that `/goal` ran in Duo Chat in the web UI ## **Out of scope** * IDE and CLI product delivery (CLI spike is a reference) * `/new` and `/compact` ([**23059**](https://gitlab.com/groups/gitlab-org/-/work_items/23059)) * Agent addressability / mid-thread agent switch * Full CLI judge parity on web before we lock how keep-going works here ## **Why now** We are putting slash actions in the prompt and teaching chat to hand work to Developer-style flows. Users will reach for a named keep-going command next. If we only ship the menu label, we teach the wrong habit. Shipping `/goal` with a real continue path while that habit forms reuses the slash surface and the developer work already in flight, and avoids a second control story later. ## **What we need to figure out** ### UX • [**Figma**](https://www.figma.com/design/ygLBopdTv6h6XD5RP24vtQ/Duo-Chat---2027Q2-?node-id=1552-15584&t=QCoFB9hsYD0BLufc-4) <table> <tr> <th>Theme</th> <th>Direction</th> </tr> <tr> <td>How the active goal shows in the panel</td> <td> Show using existing session components, avoid creating goal-specific components. Add goal icon (for now this is `work-item-objective`— I think this is safe to repurpose as it is now only used as an option users have for custom work item types) {width="223" height="600"} </td> </tr> <tr> <td>Clear / pause / cancel behavior and result states</td> <td> Cancel handled within the session, no new behavior (note: pause session is [upcoming](https://gitlab.com/gitlab-org/gitlab/-/work_items/593343), not yet available) </td> </tr> <tr> <td>Menu one-liner and good / bad examples (concrete change + checkable end vs one-shot question)</td> <td> Duo's system prompt should understand the goal concept and ideally prompt the user when a poorly defined goal is provided, however it is acceptable to just immediately short circuit a goal with ambiguous success criteria (e.g. `/goal which work item should I work on next` should result in a response "this work item is best" and then goal completion) {width="223" height="600"} {width="223" height="600"} </td> </tr> <tr> <td>Whether continue turns are hidden from the transcript if we keep going in chat</td> <td>Suggest providing context back to the orchestrating agent and letting the agent incorporate into its responses, so that if nothing is happening it can provide a heartbeat update, and if the chat has continued it can add that note as an aside without completely disrupting.</td> </tr> </table> ### **Front end engineer** * How `/goal` maps onto the slash menu surface from [**23059**](https://gitlab.com/groups/gitlab-org/-/work_items/23059) and existing chat result / session patterns once UX locks presentation ### **Back end engineer** * The CLI `/goal` contract (task, optional condition, clear) should be done in 19.3, is there additional work for us to deliver for it to work in Duo Chat?