Deep Discovery is a Codex plugin that implements a structured, 100-question self-interrogation process. It does not generate answers or models responses using AI—it provides a fixed sequence of prompts, each informed by the user’s prior input, designed to force methodical exploration of a topic before finalizing decisions. The framework assumes that thoroughness matters more than speed: whether reviewing an API design, auditing a trading strategy, or preparing to ship a Codex skill, it asks the same core question in 100 distinct, layered forms—“What could go wrong here?”—then branches into domain-specific variants. It was built to counter shallow reasoning in technical and strategic planning, especially in contexts where early oversights cascade into costly rework.

What it does

Deep Discovery does not run independently. It operates only inside Codex, a local-first, plugin-driven AI assistant framework. Once installed, it surfaces as a skill—deep-discovery—that users invoke by name or intent. Its behavior is defined in skills/deep-discovery/SKILL.md, and its domain-specific logic lives in skills/deep-discovery/references/, where separate question sets target software architecture, code review, Codex plugin development, business strategy, trading systems, and general analysis. The plugin includes a marketplace manifest (marketplace.json) and a Codex plugin manifest (plugin.json)—both validated using standard Codex tooling. It does not include a web interface, CLI binary, or standalone executable. It has no external dependencies beyond Codex itself and Python for validation.

  • Executes a strict, sequential chain of 100 questions, each conditioned on the user’s previous response
  • Routes to domain-specific reference patterns (e.g., architecture, trading, codex-plugin) based on the user’s stated goal
  • Integrates as a third-party plugin via Codex’s marketplace system—no custom runtime or server required
  • Includes built-in validation scripts compatible with Codex’s local skill-creator tooling
  • Is licensed under MIT and contains no telemetry, cloud calls, or proprietary components

Getting it running

Deep Discovery is installed exclusively through Codex’s CLI and UI. There is no pip install, npm install, or Docker image. First, register the plugin repository with Codex’s marketplace system:

codex plugin marketplace add forsonny/deep-discovery

On Windows, this command may fail with a Git staging-path error. In that case, clone the repository locally and register it from the filesystem:

git clone https://github.com/forsonny/deep-discovery.git
cd deep-discovery
codex plugin marketplace add .

After registration, restart Codex. Open the Plugins view, switch the filter to “Third-party plugins”, and install deep-discovery from the Deep Discovery Plugins listing. That final step—clicking install in the UI—cannot be done from the command line. To update the plugin later, run:

codex plugin marketplace upgrade deep-discovery-marketplace

Validation is optional but documented. From the repository root, you can verify JSON manifests and run Codex’s local skill validator:

python -m json.tool .codex-plugin\plugin.json
python -m json.tool .agents\plugins\marketplace.json
$codexHome = if ($env:CODEX_HOME) { $env:CODEX_HOME } else { Join-Path $HOME ".codex" }
$validator = Join-Path $codexHome "skills\.system\skill-creator\scripts\quick_validate.py"
python $validator skills\deep-discovery

The validator script is part of Codex’s internal tooling, not this repository—so if it’s missing, users must rely on Codex’s current skill validation workflow instead.

Who this is for

Deep Discovery targets people already using Codex who want rigor in early-stage decision-making. It is not a general-purpose question-answering tool. It is most relevant for software architects evaluating system boundaries, developers preparing to merge a complex feature, Codex plugin authors validating their own skills before publishing, or product strategists stress-testing assumptions about market fit. It assumes the user is comfortable paraphrasing, reflecting, and typing answers manually—not expecting AI-generated summaries or auto-filled fields. The 100-question sequence is intentionally heavy. If a user wants quick feedback or automated analysis, this is not the right tool. It’s built for depth, not throughput.

How it compares

Deep Discovery has no direct analog in widely used open-source tooling. It is not a LLM fine-tuning framework, a prompt engineering library, or a documentation generator. Unlike tools like llama.cpp or Ollama, it does not host or run models. Unlike promptfoo or LangChain, it does not test or chain LLM outputs. It shares conceptual ground with structured thinking methods—like the Five Whys, Red Teaming, or FMEA—but embeds those ideas into Codex’s plugin architecture as a repeatable, domain-aware script. Compared to lightweight alternatives like custom prompt templates in Obsidian or Notion, Deep Discovery offers tighter integration, versioned reference patterns, and enforced sequencing—but only inside Codex. It is heavier than alternatives like manual checklists or ChatGPT custom instructions, and less flexible than building your own Codex skill from scratch.

The project has 53 GitHub stars and no listed primary language in its repository metadata. Its documentation assumes familiarity with Codex’s plugin model and CLI workflow. It does not provide tutorials for Codex itself—only for its own integration. There is no hosted version, no API, and no mobile or web frontend. It is a narrow, opinionated tool: if you’re not already running Codex and committed to structured, human-led interrogation, it will not fit your workflow.

Source code is available at https://github.com/forsonny/deep-discovery under the MIT license.

View on GitHub →