Inspiration
Returning to a paused project should begin with reliable context. A completion message may sound reassuring while the last test still fails. A large cache may be reproducible, while a tiny local patch is the only copy of unfinished work. Rethaw puts those distinctions at the center of a restart.
What it does
Rethaw reviews one manually sanitized coding-project export staged on a UGREEN AI NAS. It produces three linked outputs:
- A restart capsule covering the goal, approved decisions, last evidenced state, unresolved blocker and next test.
- A tool-neutral restart prompt that carries evidence into a fresh development session.
- A conservative storage review separating what to preserve, conditional derivative candidates and unknowns.
Every conclusion stays tied to its source. Missing evidence stays visible. Rethaw does not delete files or promise exact session restoration.
How it is built
The native Skill consists of SKILL.md and six linked Markdown references. Its orchestration uses registered file-agent, document-agent and knowledge-agent bindings verified against the official UGREEN creation kit 0.7.3 and capability catalog 2.0.0.
File experts discover and read the authorized records; document experts extract and reconcile evidence; optional report saving requires a complete preview, separate confirmation, safe new-file creation and read-back verification. If that safety guarantee is unavailable, the report remains an unsaved preview. Development scripts, fixtures and artwork stay outside the runtime ZIP.
The demonstration case
Cartmill is a fictional checkout project. A historical completion claim conflicts with a dated currency-rounding failure. A unique uncommitted patch has no passing rerun. Rethaw preserves that blocker, the patch, the database and the approved decision digest.
The supplied inventory contains 12 GiB of dependency cache and 8 GiB of build output. With matching synthetic rebuild records, these are 20 GiB of conditional logical candidates. They are not freed space or permission to delete. Removing rebuild evidence withdraws that classification; missing secondary backup evidence remains a retirement prerequisite.
Challenges and what I learned
Evidence must agree on project, item, revision and time before it can support a conclusion. A lockfile alone is not rebuild proof, a prior rebuild does not guarantee future package access, and an absent-file check does not guarantee an overwrite-safe save. These limits shape the workflow itself.
Current state
The native Markdown workflow is implemented and passes the unmodified official static checker. Local model-only exercises use synthetic inputs and explicitly simulated tool states; they do not demonstrate native NAS execution. The package includes truthful execution and assistance disclosures. No real cleanup, NAS installation or verified native save is claimed.
What is next
If selected for Round 2, test the same Skill on the assigned NAS: first native scoped discovery and extraction, then the restart and storage cases, missing evidence, permissions and confirmed safe saving with read-back. Record a demo only after those operations actually run.
Log in or sign up for Devpost to join the conversation.