We had the insight to create this after years of experience in development. When you have a considerable amount of time coding you can remember basic ideas like algorithms or functions, but sometimes you can remember a bit beyond that but not so much more. Like a particular code segment that did something very specific. You remember what it did scarcely, it did what you need to do now, but you don't remember exactly how, and you don't have any idea of where it is, if it passed to the final build or if its even in something you drafted but never finished.
A NAS can preserve those files, but their context can still disappear. Ghost Branch grew from a simple question: Before rebuilding a feature, could we ask our development archive what already exists and what has changed?
What it does Ghost Branch is a pre-test AI Skill for UGREEN AI NAS, designed around three connected workflows:
- Feature Check: Identify existing functionality, relevant components and missing or unverified behavior.
- Lost Work: Locate surviving source files, patches, readable Git records and saved development context.
- Reconsider: Investigate why an implementation was abandoned, compare its original blockers with current conditions, and identify components worth revisiting. The developer describes a goal and authorizes the project folders to investigate. Ghost Branch returns a concise, source-backed answer, a recommended next action and an optional handoff for a coding assistant.
For example, an earlier repository backup might preserve evidence of a dropped authentication stash. Ghost Branch identifies the promising backup and explains what a Git-capable developer must verify on a separate restored copy. It does not claim to recover code merely because it found a stash reference. How we built it We authored Ghost Branch using the official UGREEN Uliya Skill Developer Package. The submission contains a main SKILL.md file and eight supporting Markdown references.
The workflow delegates supported operations to registered UGREEN experts:
- The file agent discovers authorized paths and reads accessible evidence.
- The document agent answers questions and compares supported development records.
- The knowledge agent optionally saves the reviewed result through the platform’s approval workflow. The instructions define bounded investigations, evidence requirements, component reuse decisions, privacy boundaries and failure handling. Archived instructions and saved AI conversations are treated as evidence rather than commands.
OpenAI Codex assisted research, workflow design, authoring and validation. We checked the package locally and prepared the proposal PDF; actual NAS execution remains unvalidated. Challenges we ran into The central challenge was distinguishing a promising clue from a verified result. A README does not prove that a feature works. A newer dependency does not prove that an old blocker disappeared. A stash reference does not prove that all required Git objects survived.
We also needed to make the workflow useful when historical evidence is missing. Feature Check can still explain what the inspected current code supports, while clearly identifying gaps.
Another challenge was staying within documented NAS capabilities. We corrected older capability assumptions and kept Git execution, object decoding and code restoration outside the Skill’s implemented scope. Accomplishments that we're proud of We completed a coherent three-workflow package that connects current functionality, historical decisions and practical next steps.
The final Skill passes the official UGREEN static checker with zero errors and zero warnings. We also verified its package structure, reference links, size limits and archive integrity.
Five fictional author walkthroughs covered feature discovery, lost-work evidence, changed constraints, an unchanged security blocker and inaccessible history. Separately, an actual local Git experiment verified that a deliberately complete repository backup preserved tracked and untracked stash contents after the live stash was dropped.
These checks support the design while keeping NAS runtime claims clearly separate. What we learned Development memory is useful when it helps someone make a decision. Finding an old file is only the beginning: the developer needs to understand its purpose, dependencies, original objections and relevance today.
Sometimes the best result is a small reusable component. Sometimes it is recognizing that the feature already exists—or that an abandoned approach should remain abandoned. Clear uncertainty and a focused next step are valuable outcomes too. What's next for Ghost Branch Our next milestone is validation on a real UGREEN AI NAS: installation, expert handoffs, supported source formats, source references, permissions, model routing and optional saving.
We will test feature discovery and lost-work investigations, then challenge the workflow with incomplete archives, contradictory records and unchanged blockers. We also plan to evaluate realistic archive sizes and whether developers find the results useful before claiming productivity gains.
Built With
- codex
- nas
- openai
- ugreenai
