The problem: an evolved, partly unfamiliar application
WebTViOS is a shipping App Store application owned by the investigator. Its older Swift codebase had been changed repeatedly by AI coding agents without subsequent manual source review. The reported symptom was a performance problem after downloading new electronic programme guide (EPG) data.
The goal was not to repair WebTViOS. It was to test whether Code Universe could reconstruct the feature, identify a plausible cause, and make the supporting evidence reviewable without editing the application.
Reconstructing the feature boundary
The local scan covered 23 Swift files and produced 823 graph nodes and 1,745 relationships. Searching for EPG exposed a path spanning toolbar actions, views, the fetcher, XML parser, cache, channel filtering, and timestamp formatting.
The useful moment was seeing that the feature was not confined to EPG.swift. The map supplied a connected starting point for source inspection across the UI and data-processing code.
Following a source-supported diagnosis
The inspect-only Codex review completed with 27 retained evidence events. It traced a critical boundary: XML parsing runs in the background, but its completion moves to the main queue. Cache encoding, storage, channel filtering, and state publication then follow synchronously.
- The complete feed is JSON-encoded and stored in
UserDefaults. - The programme array is repeatedly filtered for configured channels.
- The result is published from a root-owned observable object, potentially triggering SwiftUI work.
The measured feed snapshot contained 16,514 programmes and produced approximately 7.3 MB of JSON. Git history connected the behavior to commit c564610, “Cache daily EPG feed data,” dated 11 May 2026. That change replaced caching a filtered result with caching the complete feed before filtering.
This is historical evidence consistent with the reported timing—not proof of the dominant device runtime cost, or proof that this particular commit was AI-authored.
What Code Universe contributed
The map re-established the current implementation’s boundaries after several unreviewed changes. The review connected its conclusion to the relevant source and Git history, while the replay retained the investigation sequence. Together, they made the diagnosis easier to inspect before accepting another proposed change.
Limits and disclosures
- This is a founder-owned, single-project study, not an independent benchmark.
- The project had existing uncommitted changes, which were preserved. No WebTViOS files were edited for the study.
- Some heuristic member relationships were false positives caused by repeated local-variable names. Important edges still require source verification.
- The review consumed substantial model context. Efficiency remains a product concern.
- No device before/after performance test was performed. The study does not claim a measured speed improvement. Device profiling is optional follow-up work.
The result
Code Universe produced a source-supported, historically informed diagnosis and a retained evidence trail. The demonstration is about understanding and reviewing an AI-modified application—not claiming the application was fixed.
Read the detailed investigation record, benchmark caveats, and optional profiling runbook on GitHub.