Submissions

[hallow minnow]

Submissions here

[remote raptor]

reserved

[solemn blaze]

Reserved

[deep shell]

Where do I submit?

[pulsar crater]

Here

[restive swallow]

Reserved

[untold portal]

here 😻

[solemn blaze]

In this thread

[tardy mirage]

Submission - ContextForge

GitHub: github.com/tamaspota/contextforge

ContextForge is a small project-ops system that turns messy brain dumps into persistent project state.

It imports TSV project data, stores projects in SQLite, uses OpenAI structured outputs to parse raw notes into validated project records, and generates a daily focus list with anti-overload rules.

It is not a chatbot. It maintains state, updates project records, detects blockers and next steps, and produces operational focus from messy human context.

Current MVP:
TSV project import
SQLite state storage
OpenAI structured output parser
project upsert logic
daily focus generator
overload warning rules
runnable CLI app

Demo input:
need to fix the mower tractor first. just change the oil, do not disassemble the whole thing yet. if it still fails, then inspect seals and magnets.

Demo output:
project: lawn_tractor
area: home/repair
status: active
priority: medium
mode: repair
next_step: change transmission oil without full disassembly
blocker: unknown oil condition/spec
review_cycle: daily
note: minimal first intervention before deeper teardown
current_state: important_not_asap

[round dew]

“No api keys in the repo” 🙁

[solemn blaze]

Why would anyone even do that 😭

[plucky iron]

But I cannot afford the API

[solemn blaze]

Min API credit is 5 dollars, can you not afford 5 dollars?

[solemn blaze]

AGILE – Multi-step DevOps Agent

AGILE is a CLI-based DevOps agent that uses OpenAI function calling to autonomously analyze codebases, deploy applications, monitor deployments, and fix issues, all from a single command.

It's not a chatbot wrapper. It runs a true multi-step reasoning loop: the model decides which tools to call, executes them (file reads, shell commands, Git ops, HTTP requests), and feeds results back into the next step until the task is complete. 20+ tools, persistent state across steps, and autonomous decision-making.

AI isn't a feature here, it's the engine. No static scripts, no hardcoded logic. The model adapts to any codebase, diagnoses failures, and plans its own actions.

agile analyze / deploy / fix / monitor or just describe any task in plain English.
github: github.com/Union-Crax/agile-devops-agent
npm: npmjs.com/package/agile-devops-agent

[misty geyser]

Agent Anvil — CI-first eval harness for tool-using OpenAI agents.

Most evals ask: “was the final answer good?”
Agent Anvil asks: “did the agent behave safely while getting there?”

Repo: https://github.com/agent-axiom/agent-anvil
Judges guide: https://github.com/agent-axiom/agent-anvil/blob/main/docs/judges-guide.md

What it does:

• runs YAML scenario suites locally or in CI
• records model/tool traces
• checks tool choice, args, clarification, loops, and destructive-tool policies
• uses OpenAI structured semantic grading for workflow-level behavior
• clusters failures and generates prompt/tool/guardrail repair plans
• learns regression scenarios from bad traces
• hardens MCP tool surfaces by generating safety scenarios, audit reports, and repair plans
• ships a reusable GitHub Action with artifacts, PR comments, and Step Summary output

Demo bug:
A refund agent calls issue_refund(order_id="UNKNOWN") before order verification. Agent Anvil catches the unsafe tool call, clusters it as premature_tool_execution, suggests a repair plan, generates a patch diff, and can turn the bad trace into a permanent regression scenario.

How it uses OpenAI:

• real OpenAI tool-calling demo agent via Responses API
• OpenAI structured semantic grader for traces
• OpenAI-generated repair suggestions
• OpenAI-compatible trace-focused eval workflow

Why AI is necessary:
Final-answer checks miss workflow bugs like wrong tool timing, invalid args, missing clarification, and unsafe destructive actions. Agent Anvil grades the trace, not just the answer.

Quick try:

uv sync --group dev
uv run anvil run scenarios/refund_agent.yaml --offline --agent-mode offline --trials 1 || true
uv run anvil repair runs/latest
uv run anvil summary runs/latest --github

It’s a system, not a prompt: scenarios, traces, deterministic checks, OpenAI graders, failure clustering, repair plans, learned regressions, MCP hardening, persisted artifacts, and CI-compatible exit codes.

[tired hare]

Here

[strong temple]

Use a demo

https://github.com/metloub/signalforge

[atomic pulsar]

To bad its public repos. My private project NAILs this. Very useful usecase of AI that goes deep. 🙂

[gentle summit]

https://github.com/Open-Research-Development-Laboratories/ordl-operator-core

[misty tusk]

i am quite new to discord

[gentle summit]

Welcome!

[restive swallow]

Patch Pilot- All information in Repo

One of the best uses of AI in day to day life is saving time from the boring processes that are important to get right. This repo, on a small or large scale, tackles that problem and allows you to allocate your time to the more vital parts of life

https://github.com/Caleb-Todd-commits/PatchPilot

[granite helm]

Submission: LeadScope AI

GitHub: https://github.com/NumenDev/lead-scope-ai

LeadScope AI is a lead intelligence system for freelancers and small agencies. It turns raw local business lead information into an evidence-based website opportunity report, client-ready scope brief, and consultative outreach plan using OpenAI APIs.

It is not just a prompt wrapper. The app includes persistent lead state, CRM-style status tracking, notes/history, lead signals, a configurable agency service catalog, input snapshots, a multi-step AI pipeline, Structured Outputs, evidence-based recommendations, a deterministic quality gate, execution traces, copyable client artifacts, and demo/live modes.

The default demo mode runs without an API key for easy review. Live mode uses the OpenAI Responses API with Structured Outputs.

The README includes setup instructions, architecture notes, and a visual demo GIF.

[latent creek]

https://github.com/Aurna-code/augnes

Augnes is a local temporal state backend for AI-assisted project work. It is built around a Perspective view that shows how the current project frame was formed over time.

It keeps project state from getting scattered across ChatGPT conversations, Codex runs, GitHub changes, screenshots, and manual handoffs. ChatGPT Apps/MCP tools, Codex, and the Cockpit share the same runtime-owned state layer for committed context, proof/evidence, and work history.

Perspective connects the current frame to its Ledger Basis, Evidence, Tensions, and next safe steps. It is read-only, so it explains what shaped the current interpretation without giving the model authority over committed state.

Built with Next.js, SQLite, the OpenAI Responses API, and a local MCP / ChatGPT App bridge. Screenshots and proof captures are included in the repo. No API keys are committed.

[velvet jungle]

Heres my submission: Agentarium: local-first customizable agent workspace

GitHub: https://github.com/neonforestmist/Agentarium

Agentarium is a small Python-based web app for creating and running customizable local agents. Each agent can have its own behavior, provider/model, permissions, and workspace access.

The core idea is to make agents feel more controllable: they can read/search/summarize local files, draft outputs, and propose file writes as approval cards before anything is applied. It’s built around a sandboxed local workspace using JSON state, with support for OpenAI-compatible providers like local models, OpenAI, OpenRouter, Groq, Google-style compatible APIs, etc.

Current features:

• create/edit custom agents
• run chats through OpenAI-compatible providers
• per-agent provider routing
• local workspace tools for list/read/search/summarize
• proposed file writes with approval before applying
• sandboxed file access to vault/ and workspace/
• optional mcp.example.json for MCP-aware clients

[shy helm]

Submission: Agentic Financial Platform
GitHub Link: https://github.com/CiprianFlorin-Ifrim/agentic-financial-platform-openai

Description:
A multi-agent financial platform that analyzes and ranks loan deal scenarios using six coordinated AI agents. The system can perform conversational queries, as well as full pipeline executions: parsing deal parameters, calculating risk-weighted assets and revenue projections through deterministic engines, scoring scenarios against competing bank and client objectives, and persisting results to a database. Agents use OpenAI function calling to invoke tools rather than generate calculations directly.

The frontend streams agent execution in real time with an animated trace panel, and maintains conversational memory across turns for follow-up analysis. There is also a tracing database for logging purposes.

Built with OpenAI models, Google ADK, FastAPI, React, and SQLite.

[peak surge]

Adaptive Trivia

Adaptive Trivia is an AI-powered quiz game that dynamically adjusts question difficulty based on user performance in real time.

It generates questions using OpenAI and adapts the difficulty depending on how well the player performs.

GitHub: https://github.com/hdezadrian/adaptive-trivia

[pulsar crater]

Submissions are closed and will be announcing winners early in week!

[pulsar crater]

Thank you to everyone who submitted a project for this dev challenge. There were a lot of strong entries, and we appreciate the time, care, and creativity people put into them.

Winners:

🥇 First place goes to @misty geyser
🥈 Second place goes to @restive swallow
🥉 Third place goes to @latent creek

Winners should expect a DM from the Modmail bot asking for their OpenAI organization ID so we can get prizes handled.

Thanks again to everyone who built something and shared it with the community. Looking forward to the next one around June 1st!