Project Assistant: Ask Questions About Your Project

Overview

The Project Assistant is a chat panel built into Shift-Left Studio. You ask a question in your own words, about your own project, and the assistant reads your project's data before it answers: requirements, API specifications, endpoints, tests and their assertions, run history, run analyses, workflows, mocks, data sets and more.

It is built for the questions that screens do not answer directly. For example: "Why did the checkout tests fail last night?", "Which requirements have nothing verifying them?", or "Does this test actually check anything beyond the status code?" Every answer shows what the assistant looked at, and every item it mentions can be opened with one click.

The assistant can also propose changes, such as running tests, fixing a test, or creating an endpoint. It never makes a change on its own. You review each proposal and accept it. See Project Assistant: Proposals You Approve.

Key concepts

Before you begin

• An AI provider must be configured. An administrator sets this up under AI Settings. Without it, the launcher does not appear, and the conversation sections inside drawers explain that AI is not configured.
• Open a project. The assistant answers questions about one project at a time. With no project open, the launcher does not appear.
• Permission to view test cases. The assistant reads the same data you could open yourself, and nothing more.
• Daily AI usage limit. Each question uses your installation's AI provider and counts toward the same daily AI limit as test generation and run analysis.

Note: The assistant is available in every edition. Some of the things it can do for you, such as running packs or performance tests, still depend on your edition, exactly as the matching buttons do.

Step 1 — Open the assistant

You can open the assistant in two ways.

From the launcher (project-wide or current screen)

1. Open a project.
2. Click the purple sparkle button at the bottom right of the screen. Hover over it first to see what it will ask about, for example "Ask about this project" or "Ask about POST /orders".
3. The Assistant panel opens. The line under the title names the subject and the project.

The launcher follows the screen you are on: an endpoint, a workflow, a run report or a mock. Elsewhere it opens about the whole project.

Tip: The small link button above the launcher copies a one-line reference to what is on screen. Paste it into a ticket, a chat message, or a question to the assistant from another screen.

From inside a drawer or panel (anchored conversations)

Some screens carry their own conversation instead of opening the panel:

The launcher hides while one of these drawers is open, so you never have two assistant panels for the same thing.

Step 2 — Ask a question

1. Type your question in the box at the bottom (placeholder: "Ask about a failure, a requirement, a test…").
2. Press Enter to send. Press Shift+Enter for a new line.
3. While the assistant works, a progress line shows what it is reading, for example "Reading your requirements, tests and runs…".

On an empty conversation, the panel offers starter requests you can click, such as Show what failed in the most recent run, List the requirements that nothing verifies yet, and Summarise this project and its test coverage. A failure cause drawer offers Explain why these tests failed, Tell me whether this is a product bug or a test problem, Recommend what to do about this cause and List the tests in this group.

Refer to a specific item with @. Type @ followed by at least two characters of a name, for example @TC08 or @Login. A list of matching tests, endpoints, features, workflows, packs and requirements appears, this project first and then other projects you can open. Pick one with the arrow keys and Enter or Tab. The assistant receives the exact item, so two tests called "Login" in different projects are never confused.

Step 3 — Read the answer

An answer can include several parts:

• Citations open in the project tree. Clicking a feature, endpoint or test opens the tree down to that item and selects it. A chip without an arrow cannot be opened and is plain text.
• Counts are totals. Asked for a list, the assistant gives the total and says when it shows only part, for example "10 of 31".
• Only what it read is cited. Every ID in a citation or offer came back from a lookup in that answer; the assistant cannot point you at an invented item.
• Disagreements are labelled. If the assistant reads a failure cause differently from the run analysis, it says so and offers to reclassify it. It never changes the verdict itself.

Step 4 — Attach a specification or document (optional)

Attach a file so the assistant can work with it, for example "Import this spec and generate tests".

1. Click the paperclip button beside the message box (Attach a specification or a document), or drag and drop or paste a file into the panel.
2. Each file appears as a chip showing its name, size and status: Uploading…, then what it was recognized as: a specification (with the kind detected, when it can tell) or Document.
3. Type your question and send it.

Other file types are refused with a pointer to Setup → API Specs or Requirements → Documents. A failed upload never blocks your question; the chip shows why.

Step 5 — Continue, resume or switch conversations

• One conversation per item. Opening the assistant from the same place again resumes the same conversation. Conversations anchored to a cause or finding are shared with anyone looking at that cause. A project-wide conversation belongs to the person who started it.
• Earlier conversations. When the current conversation is empty, the panel lists Earlier conversations with their message counts. Click one to resume it.
• Close the panel with the X button or the Esc key. Your conversation is kept.

What you can ask about

The assistant reads the whole project. The table below lists example questions by goal.

Product questions are answered from the assistant's knowledge of Shift-Left Studio; questions about your data are always answered from lookups.

Change requests are answered with a way forward. If you ask for a change that cannot be offered yet, the assistant tells you why and the one step that would make it possible. For example: "This test has never been run. Run it once; the run is analysed automatically and the missing checks can then be offered here."

Questions across projects

The assistant starts in the project you have open. To ask about another project, name it, for example "Compare the login tests in Orders API and Payments API", or use an @ mention from that project.

It crosses only into projects you can already open. If you name one you cannot access, it tells you so rather than answering as if that project were empty. Workflows that span several projects are read in full.

When something goes wrong

• The AI provider fails. Your question is kept, and the failure appears under it, for example "No answer came back." Ask again later.
• The assistant cannot be reached. The panel says so and offers Try again.
• A lookup fails, or the lookup budget runs out. The assistant answers with what it has and says what it could not check. Ask a narrower follow-up to continue.

Privacy and data handling

• Your permissions apply. The assistant reads only projects and items you can open.
• Raw request and response bodies are not stored in conversations. When the assistant reads a run, the bodies are redacted (credential-shaped values removed) and size-limited while the answer is written, then discarded. The conversation keeps the answer, not a copy of your run data.
• Your AI provider sees the question and the redacted lookup results, as it does for test generation.
• Never paste secrets. Your messages are stored as part of the conversation. The assistant refuses to store passwords, keys and tokens you type, but it cannot un-send a message. Use the Authentication screen for credentials.

Troubleshooting

Best practices

• Ask from where the thing is. A question inside a failure cause's drawer starts with that cause's evidence already read.
• Name items precisely, or use @ mentions.
• Tell the assistant what the data cannot show, for example "that order ID does not exist on staging".
• Expand Looked at N things when an answer matters.

FAQ

Does the assistant change anything when I ask a question? No. Reading never changes data. Changes happen only when you accept a proposal card.

Can other people see my conversations? Project-wide conversations belong to you. Conversations inside a failure cause or validity finding drawer are shared with teammates who open the same item and can view the project.

Does it support every protocol? It reads REST, SOAP, GraphQL, JSON-RPC 2.0 / MCP and WebSocket-RPC projects. JSON-RPC and WebSocket-RPC need the Trial or Enterprise edition.

Can I make it remember something about my project? Yes. See Teaching the Assistant: Feedback and Learned Notes.

Related articles

• Project Assistant: Proposals You Approve
• Project Assistant: Create Projects, Endpoints, Tests and Workflows
• Teaching the Assistant: Feedback and Learned Notes
• Run analysis: failure causes and bulk repair
• Test run

For administrators (self-hosted installations)