## User story As a developer using Duo Chat, I want to upload files directly from my computer into the conversation, so that I can show Duo what I mean instead of describing screenshots and documents in text. ## Problem Duo Chat only accepts text today. When Sasha hits a UI bug, gets a design mockup on their desktop, or has a spec as a local PDF, the workaround is painful: upload somewhere else, copy a URL, paste a description, or give up on visual context entirely. Issue and MR editors already support paste-to-upload. #### Example use cases to support 1. User wants to fix a bug which is rendered in the product without having to describe the UI, instead show the agent. 1. **Successful outcome:** User uploads images of the UI, agent parses the image to understand context, combines with existing context of the repo, to determine what fix needs to be made to address the problem as shown. 2. User wants to implement a particular change based on an existing design 1. **Successful outcome:** User uploads design as an image, agent parses the image to understand the request, combines with existing context of the repo and/or associated work items, and implements the change to match the uploaded image 3. User wants to build out work items based on content in an image, e.g. from a slide or diagram created for team communication 1. **Successful outcome:** User uploads an image with information related to upcoming work, agent parses the image, defines the work items, and creates work items structured in a way relevant to how it's organized in the image (e.g. visual hierarchy becomes work item hierarchy) 1. _This will be further enhanced beyond this epic with non-image uploads, e.g. upload a spreadsheet of work item details and have it automatically produce work items combining project context with spreadsheet info._ ## Scope **Scope lock:** Duo Chat web (Agentic) · local file upload · **Done when:** user attaches a file from their computer and Duo uses it in that conversation. [**🎨 Designs**](https://www.figma.com/design/VqpH0V6joIOokfeXS71Yqz/Duo-Agent-Platform-Side-Panel?node-id=19171-399854&m=dev&focus-id=19171-400499) <details> <summary> Prior design </summary> [Design file](https://claude.ai/design/p/12d30f75-7622-4221-93aa-d38700d111bf?file=File+Upload+Support+in+Duo+Chat+-+Epic.html&via=share) </details> **In scope** - Paste, drag-and-drop, and file-picker upload into the Agentic Chat input (web) - Inline preview before send - v1 file types chosen in discovery (expected: common image formats first) - Upload storage and passing attachment references through the chat/workflow pipeline (Layer 1) - Admin disable and documented size/type limits **Out of scope (this epic)** - Issue attachment fetching / vision over GitLab-hosted issue images (prior epic draft — future iteration or sibling epic) - Office/PDF document ingestion — https://gitlab.com/gitlab-org/gitlab/-/work_items/595072+ - IDE/CLI upload surfaces - Repository binary/image file reading - Design Management files - Automatic ingestion without user attaching a file ## Plan of record (updated 2026-09-09, from the 9 Sep sync — @tbulva DRI) | Phase | What | Owner | Target | |---|---|---|---| | **1 — Ship v1 behind flag** | Merge https://gitlab.com/gitlab-org/gitlab/-/merge_requests/250665 as-is: images (PNG/JPEG/WebP, ≤5 files, ≤2 MiB each) via `+` menu, paste, drag-and-drop; sent as `attachments` additional-context entries; gateway side https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/merge_requests/6731 (merged 8 Sep). Behind `dap_web_chat_file_attachments`. | FE @igor.drozdov / @tbulva · BE approval @Kaveh_Nejad_Gitlab | 19.4 (or first days of 19.5) | | **2 — Design parity + BE hardening** | FE: previews, composer layout, error/limit states to @nickleonard's designs — https://gitlab.com/gitlab-org/gitlab/-/work_items/628537+. BE: keep non-base64 attachment data in `ui_chat_log`; prevent retries on turns that carry attachments. Team-member rollout via https://gitlab.com/gitlab-org/gitlab/-/work_items/627543+. | FE @tbulva / @nickleonard · BE @Kaveh_Nejad_Gitlab | 19.5 | | **3 — Flow-registry integration** | Test against the flow-registry backend once it lands (John S. is accounting for attachments in the registry requirements); scope gaps then. | TBD | No date — depends on flow registry | ### Decisions - **Prompts with attachments are not queueable in v1.** A draft with images can't be sent while a turn is running; the images are held for the next sendable turn. IndexedDB persistence deferred. (Agreed 9 Sep; @nickleonard confirmed on !250665; revisit if we move to upload-then-reference.) - **Sequencing:** the PromptComposer refactor (https://gitlab.com/gitlab-org/gitlab/-/merge_requests/253612, merged 8 Sep) landed first; !250665 is rebased onto it. - **Known v1 limitation:** an image is visible to the model only until the first interrupt of the turn it was sent on — durable fix tracked in https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/work_items/2826. - Related work elsewhere: LLM vision over issue-hosted images — https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/work_items/2820 (not this epic). ## Platform dependency - AI Gateway multimodal input: https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/work_items/2148 - Architecture reference (Slack adapter Layer 1/2 pattern): https://gitlab.com/gitlab-org/gitlab/-/work_items/602535+ - Prior team discovery (attachments nice-to-have, Aug 2025): https://gitlab.com/gitlab-org/gitlab/-/work_items/559987+ ## Measurement ### North star (outcome) **Visual-context task completion:** share of chat turns where the user attached a local file and received a completed assistant response they could act on, without falling back to text-only workarounds. **Goal**: completion rate increases as upload ships. No regression in overall chat completion. ### Friction levers (inputs) - **Upload adoption:** % of agentic web sessions with ≥1 local attachment - **Time to first useful response** on attachment turns vs text-only baseline (dogfooding metric) ### Guardrails - **Upload failure rate:** failed paste/pick/drop / attempted uploads - **Conversation abandon after failed upload** - **Security block rate:** scans or policy rejects (should be visible, not silent) ### Baseline TBD — discovery dogfood. No production telemetry for chat attachments today. ## Child work - https://gitlab.com/gitlab-org/gitlab/-/work_items/591047+ — image attachments as context (Phase 1) - https://gitlab.com/gitlab-org/gitlab/-/work_items/628537+ — attachments UI to design parity (Phase 2, FE) - https://gitlab.com/gitlab-org/gitlab/-/work_items/627543+ — `dap_web_chat_file_attachments` rollout (Phase 2) - Phase 2 BE hardening issues (`ui_chat_log` without base64; no retry on attachment turns) — @Kaveh_Nejad_Gitlab to open