Inspiration
CodeGnosis began as an ambitious prototype assembled by four AI collaborators working without one shared domain model. The vision was powerful: let a human and an AI understand the same codebase in their own native languages. But the architecture drifted out of sync. Control centers multiplied, calculations contradicted one another, and a test project containing a real dependency loop could still receive a perfect health score.
For OpenAI Build Week, we recovered the original vision and rebuilt it around one truthful structural model.
What it does
CodeGnosis scans a project locally and read-only, resolves real dependency relationships, and creates one canonical evidence- bearing model.
Humans receive that model as an explorable galaxy with guided investigations, dependency paths, entry points, loops, unresolved references, Gravity, Blast, Drift candidates, uncertainty, and line-level evidence.
AI receives the same reality as complete versioned JSON containing files, relationships, diagnostics, evidence, exclusions, completeness, hashes, and bounded source context.
CodeGnosis also provides a printable human report, configurable privacy exclusions remembered per project, and a Vault containing plain-language definitions, the public lexicon of Another, and the Law of the Living ethical base.
It reports evidence and limitations instead of inventing a universal health score.
## How we built it
The Build Week rebuild was authored in one primary Codex session using GPT-5.6.
Codex helped recover the product’s intended architecture from the failed prototype, dynamically prove its semantic failures, define the domain contract before implementation, build the Python analyzer and React interface, create adversarial fixtures, construct privacy barriers and the local launcher, and verify that every human and AI surface projects from the same canonical model.
Veris, operating through Claude Code CLI, served as the independent architecture challenger and Key-2 auditor. She tested the contracts adversarially from disk, surfaced truth and privacy defects, and verified the finished system without contributing implementation code.
Challenges we ran into
The hardest problem was not drawing a graph. It was preventing different surfaces from quietly describing different realities.
We also had to distinguish deep dependency parsing from simple file recognition, preserve uncertainty around dynamic and runtime-only relationships, protect private customer files before analysis, prevent symlink and junction escapes, and make complex structural language understandable to someone encountering the product for the first time.
Real browser use exposed problems that isolated tests missed, including misleading upload language, local-context privacy assumptions, launcher port races, obstructed tooltips, confusing terminology, and tutorial layout imbalance.
Accomplishments that we're proud of
CodeGnosis now has one deterministic domain model shared by the galaxy, guided investigations, inspector, AI export, printable report, and Vault.
The analyzer produces evidence-backed dependency relationships across major language families, identifies real cycles and supported entry points, records exclusions and completeness, and packages bounded source context without writing into the analyzed project.
What we learned
A sophisticated vocabulary cannot compensate for an untruthful model. not merely the builder’s machine. Tests must challenge semantic truth, not only whether code executes.
We also learned that AI collaboration becomes much stronger when implementation and independent audit remain separate, evidence is inspected from disk, and uncertainty is treated as useful information rather than something to conceal.
What's next for CodeGnosis
Next we will add Structural Delta: honest comparison between two analyses, showing what changed without pretending that change automatically means improvement or decline.
We also plan source-aware search, evidence-backed unfinished-intention detection, stronger temporal diagnostics, expanded parser depth, and continued refinement of the adaptive galaxy viewer.
The destination remains the same: instantaneous mutual intelligibility between a human, their codebase, and their AI.
Log in or sign up for Devpost to join the conversation.