Pascal programming that ends in a diff you can review

What the agent can do in a project

• Filesread_file, write_file, edit_file, grep, glob and list_dir work inside the chat's working folder, so a .pas unit, a .dpr file or a text .dfm form is one request away. A counter shows files created, edited and read.
• Shellruns your build command, test command or formatter on Windows, in WSL or over SSH on a server you own. Success is judged by the real exit code, and dangerous commands are blocked before they run.
• Webweb_search finds pages and http_get reads one documentation page, which helps when you need the exact signature of a routine or the wording of an error message. The agent reads the page; you decide what to use.
• Screenshots and imagesocr reads text from a saved screenshot, for example a compiler message you could not copy from the window, and gen_image saves real picture files into the working folder.
• Orchestrationtask starts parallel sub-agents for independent parts of a job, run_on_model sends one hard step to a bigger model, and skill reuses a saved playbook, such as a version-control routine for branches and commits.
• Approval modesAuto-run acts at once, Ask first pauses before actions that change something and shows the diff to accept, and Plan only writes a plan and runs nothing. Plans survive restarts of the app.
• Cronruns a prompt on a schedule, for example a nightly request to run your test command and report the exit code. Set the schedule once, and read the results in the chat the next morning.
• Modelsconnect a cloud AI provider through an API endpoint, a model running locally on your PC, or the built-in model. The Test button shows the response time and the context window of the connection.

Four tasks from real projects

Read an inherited unit

A Delphi developer who inherits a project often opens a unit nobody has touched for years. Ask Arraydeck to explain it. The agent calls read_file, then describes the interface section, the implementation section and the uses clause in plain words, and grep shows where its procedures are called from other units. Nothing is written. Use the answer as a map, not as proof: compare each claim with the source, open the unit in your own editor and run your build command, which reports the real exit code, before you change a line of the project. Repeat the request for the next unit when the first one is clear.

Make a string routine handle Unicode

To convert Unicode code in a string, first find the routine that builds or splits the text. The agent runs grep for the function name, reads the unit with read_file and proposes a change with edit_file. In Ask first the diff waits for you. Shell then runs your build command and your test command, and each card shows the real exit code. Strings use single quotes and := assigns a value, so read the diff against how the old routine worked, test it with text from your own data and reject the change if one case behaves differently. Keep a copy of the old routine until the tests agree.

Reduce the returns in a function

People argue about how many returns a function should have, and Arraydeck leaves that choice to you. Pick a function with several exits and ask for a version that sets Result once. The agent reads the unit, edits the function and shows the change as a diff. Pascal programming keeps the return value in Result, so the edit is small and easy to check line by line. Run your tests afterwards, because the agent cannot promise the new version behaves like the old one, and reject the diff if a test fails. A small, reviewed edit beats a big rewrite.

Check how a form uses dialogs

To see how to use FMX dialogs correctly in your own project, ask the agent to find every place where a dialog is shown. It runs grep over the .pas files, opens the matches with read_file and writes a short list into the chat, naming each unit and each event handler. Events such as a button click are procedures of the form, so the list reads like a table of contents. Switch to Ask first before you ask for changes, review each diff and test the forms yourself, because the agent does not design forms visually. Repeat for each form.