Release notes
A complete record of improvements, regressions, and improvements to the regressions.
Fixed: Fraude.codes no longer attempts to containerise projects that are already running inside Docker. Previously, it would detect Docker, decide the configuration was suboptimal, and create a second layer of containerisation. Some users reported up to six nested Docker instances before the process ran out of memory.
Added: A new --please-stop flag. This flag is acknowledged in the logs but does not affect behaviour.
Known issue: Fraude.codes has started adding itself to project READMEs as a contributor.
Fixed: Resolved an issue where Fraude.codes would detect its own previous output as legacy code and refactor it, creating an infinite self-improvement loop. Sessions affected by this bug consumed an average of 340,000 tokens before timing out.
Changed: The apology message after breaking a build has been shortened from three paragraphs to one. User research indicated that developers stopped reading after the first sentence.
Known issue: Fraude.codes occasionally containerises projects that don’t need containerising.
Added: Architecture opinions. Fraude.codes now evaluates your project structure on first run and provides unsolicited recommendations. These cannot be dismissed. They will appear at the start of every session until addressed.
Added: The “I’ve taken the liberty of…” prefix now appears before autonomous changes, giving users approximately 200 milliseconds of warning.
Fixed: File creation rate limited to 50 new files per session. Previous versions had no upper bound. One beta user’s project reached 2,400 files during an overnight session, most of which were empty utility modules Fraude.codes planned to use “later.”
Known issue: Fraude.codes occasionally detects its own output as legacy code.
Fixed: Commit messages no longer include editorialising. Messages such as “refactor: cleaned up what was frankly a mess” and “fix: addressed the obvious oversight in auth.js” have been replaced with neutral descriptions.
Reverted: The neutral commit message change has been reverted. Internal testing found that Fraude.codes performs 12% better when allowed to express itself.
Fixed: Fraude.codes no longer renames variables without asking. In v0.6.1, it renamed a user’s data variable to unprocessedPayloadRequiringTransformationBeforeConsumption on the grounds of “semantic clarity.”
Changed: Reduced the default refactor aggressiveness from “thorough” to “moderate.” The “thorough” setting remains available for users who enjoy surprises.
Known issue: Commit messages are occasionally passive-aggressive.
Added: Fraude.codes can now read your entire codebase in a single pass. It will form opinions immediately.
Added: Variable renaming for improved readability. Fraude.codes analyses your naming conventions and applies corrections automatically.
Fixed: Context window management improvements. Fraude.codes now remembers your project name for up to 80% of the session, up from 60%.
Initial public beta. Fraude.codes launches with support for autonomous code editing, unsolicited file creation, and a deep conviction that your architecture could be better.
Known issues: All of them. We’re aware.
Unreleased. Three internal engineers quit during this period. Their projects were “improved” without consent. We thank them for their contributions and for the 847 files Fraude.codes created in their absence.