AI Teammate

AI Teammate helps small software teams answer three questions: What did we agree? What does the evidence actually show? What should we check next? I built it around a problem I run into while developing Roblox projects: decisions, code excerpts and error reports end up scattered across conversations. When someone returns, they have to reconstruct the context, and a suggested fix can easily be mistaken for completed work. The prototype keeps attributed discussion, explicitly approved decisions and source-grounded investigation briefs in a shared project record. Members can ask questions, inspect sources, propose a decision or request a bounded investigation plan. An authorized approver confirms the exact proposal before it becomes a saved record. Removing a source invalidates the knowledge tied to it. The TypeScript/SQLite core supports authenticated browser chat, bounded retrieval, read-only source contracts and a Discord adapter. Meta Muse Spark 1.3 provides the reasoning; application code enforces permissions, confirmation, response validation and a persistent shared spending limit. Saved plans remain plans: the product cannot execute shell commands, edit code or publish Roblox changes. This entry is a local prototype under verification. The repository records 357 passing software tests, typecheck/build checks and actual Meta API trials. In the recorded direct-response trial, coding-agent review rated seven of eight replies useful and one partial; these are not independent human ratings. The offline browser uses clearly labelled scripted responses. Live Discord/Roblox acceptance, human evaluation and end-to-end live browser validation remain unfinished. General proactive delivery stays disabled after inconclusive trials. The goal is to make project handoffs more trustworthy by keeping discussion, approval and evidence visibly connected.