CosmoOrchestrator: AI Agents That Turn a Chat Request into a Pull Request

Context

Small engineering teams get a steady stream of small feature and fix requests over chat. Each one needs clarifying, a spec, code, tests, and a pull request before anyone can review it.

Problem

Coding agents that act on a one-line prompt either guess the requirements or change the repo with nobody signing off, and their claims about what they changed and whether tests passed cannot be trusted.

What I did

• Built a Nuxt + Nitro platform with a Lead agent that takes requests from web chat, Telegram, Lark, or WhatsApp, asks up to 4 clarifying questions, and writes a structured PRD.
• Added a human approval gate: nothing touches the repo until the operator replies ACC (or TOLAK to reject) in chat or approves on the task page.
• After approval, the pipeline researches the repo (web search, README, pgvector memory search), forks when needed, branches, and hands the spec to a Coder agent chosen by the repo’s language, which edits code inside an E2B sandbox; files to commit come from git status, not the model’s claims.
• QA detects Gradle, Maven, npm, or pytest, runs the suite, and loops failures back to the Coder up to 3 times; a mission that still fails stops without a PR, otherwise a pull request is opened with Octokit.
• Made it recoverable: LLM errors halt before committing, failed tasks retry twice, and missions interrupted by a restart resume on boot.
• Added multi-workspace isolation (per-workspace GitHub tokens, encrypted provider keys, chat binding with /bind codes), per-agent LLM provider and model, MCP tools, a research agent that fills an Idea Inbox, GitHub security and Postgres slow-query agents, and a kanban, live log, and 3D agent dashboard.
• Deployed it on Kubernetes through GitHub Actions with cert-manager TLS and an HPA, backed by 240 Vitest cases and Playwright smoke tests.