Greeting! I have made 1 bid decision and ported the whole hag to python scipt using external libraries instead of building every agent myself. But for this we need gemini-cli configured with a paid account(which I dont have!). Sad, but I am happy rest all of hag works without any issues

Got 208/208 tests to pass. Also I have implemented the providor for LLM as Gemini(Primary) -> Openrouter -> NVIDIA. So that maximum success rates can be achived.
Also for anyone to test out HAG, just use README.md from github to install HAG, then setup up your API keys in the .env (don’t worry it’s all local), and you are ready to go.

Heyyy, I have updated my README.md and also fixed my failing tests by now. I guess I am ready to ship Hack-A-Gent for Ship#1

I have realised that using NVIDIA NIMS is the issue, my app is almost working. , but the providors is failing at this point. I will try using gemini and openrouter to chek if it will work or not. I guess I will not be able to. And I also need to update my README.md for some reason it looks like AI written, I will write it in a funny way now.

Fixed a pretty annoying bug today. HAG can now actually reach Nemotron during auto-repair after NVIDIA 503s. One step closer to a full successful run.

Major progress today! I completed the Phase 1 stabilization of Hack-A-Gent by resolving all remaining test and build issues (1721/1721 tests passing, clean TypeScript build). I also performed a deep forensic investigation into the AI generation pipeline and discovered a critical issue where the router was rejecting recoverable LLM responses before they reached the JSON extraction layer. With the core architecture now stable, the focus shifts to eliminating unnecessary template fallbacks and improving the quality of AI-generated hackathon projects.

Today’s work focused on stabilizing Hack-A-Gent for a production-ready release. I fixed the Improvement Pass so it now completes the full critique -> patch -> build -> judge cycle with measurable score improvements, strengthened template fallback generation with complete project scaffolding and improved styling, added automatic generation of deployment assets like LICENSE and vercel.json, and implemented comprehensive child-process lifecycle management to ensure the CLI exits cleanly after execution. I also streamlined the interview flow by removing unnecessary prompts, improved validation, and verified the project with a clean TypeScript build and extensive unit and feature test coverage.

This is session I added a better parser so I get almost 99% accuracy and also made many improvements in the code generation. Almost done and drained out from all this

Spent this session improving Hack-A-Gent’s reliability and generation pipeline. Fixed a major architectural issue where generated apps were being overwritten by fallback templates, allowing AI-generated pages to be preserved correctly. Also resolved several runtime and build issues, improved validation, and verified the full pipeline by generating, building, and running a real hackathon project successfully. The next focus is improving project quality so the generated apps are more original, context-aware, and competitive in real hackathons.

Today was mostly about making Hack-A-Gent more reliable instead of adding new features. I finally tracked down a long-standing parser issue that turned out to be an architectural problem: the CLI and the analysis pipeline were using two completely different parsers. I merged everything into a single parser, updated the production build to include it, and verified that all parser regression tests and the full test suite still pass. After that, I confirmed the CLI now extracts real judging criteria like Impact and Innovation instead of mistakenly treating navigation links as judging criteria. With parsing in a much better place, I shifted focus toward hardening the overall pipeline by planning phased reliability work around generation, auto-repair, validation, and recovery so the goal is no longer just generating code, but consistently producing projects that typecheck, build, run, and are ready for submission.

This week I focused on making Hack-A-Gent stable and release-ready instead of adding new features. I cleaned up dead code, removed unused files, improved CLI logging and error messages, and verified the project with a full build and test pass. The CLI now parses real hackathon pages, plans projects, and gracefully falls back when AI providers are unavailable. I also ran end-to-end tests on real hackathon URLs to identify the final issues before release. The remaining work is improving template fallback and validation so every generated project is immediately usable, bringing Hack-A-Gent closer to its first stable release.

Might be on final stages, making sure the agent doesn’t blows up in users hands. Taking time but needs to be done, Hope I can ship it by this week

Spent today improving Hack-A-Gent’s long-term direction instead of just fixing bugs. I mapped out a roadmap to turn it into a real Devpost hackathon agent, including a hackathon qualification engine, real Devpost benchmark suite, autonomous repair loop, browser validation, and organizational learning. Also audited the current architecture to identify what already exists, what needs refactoring, and what should be built next before writing more code.

I built this project using my AI (Actual Intelligence) and a really well-written Stardance guide. It was my first time working with the NASA APOD API, and I learned a lot about fetch(), Vite, and environment variables. I also ran into a GitHub Pages issue that took some debugging before I finally got everything working. The CSS was surprisingly fun too, especially experimenting with clip-path and Google Fonts. Overall, this was a simple project that taught me a lot more than I expected.

The website generation is pretty irritation, partialy fixed it. Needs more work

Hack-A-Gent

I built Hack-A-Gent, a CLI-based AI agent that helps developers turn a Devpost hackathon challenge into a working project in minutes. Simply paste a Devpost competition URL into the terminal, and Hack-A-Gent analyzes the challenge, extracts the judging criteria and submission requirements, plans an appropriate architecture, and generates a production-ready project scaffold.

You bring your own LLM API key (Anthropic, OpenAI, Google Gemini, NVIDIA NIMs, Ollama, and more), and Hack-A-Gent orchestrates the rest—from understanding the hackathon prompt to generating code, documentation, and a submission-ready starting point.

The goal isn’t to replace developers, but to eliminate the repetitive setup work so builders can spend more time refining their ideas, polishing their projects, and creating a great demo.

Added proper guest mode to StartupOS! Users can now complete the founder interview, generate a blueprint, and explore the workspace without creating an account. Also fixed a bunch of authentication edge cases, improved session handling, and ran a full Playwright audit to make the app much more stable.

Thanks for the reviews, those helped really well. For some reasons my app broke during the QA Testing and I would apologize for that. (I have fixed it now hope so!). Also I have reworked the CSS(to make it less AI slop). Also this is 100% AI free. I hope you like it