dsh-thesis

The hardest part of a long thesis is rarely a single paragraph; it is the project-wide rules that scatter everywhere: citation formats, chapter-by-chapter advisor feedback, LaTeX layout constraints, glossaries, chapter completion status. General-purpose writing assistants treat every session as a first meeting with your project, and the context window simply cannot hold cross-chapter state. dsh-thesis pulls these rules into a local, structured registry so the project stays consistent across sessions without relying on model memory.

The core design centers on safety and determinism. New records always start in quarantine; only a Host-verified project owner can create candidates, approve, revoke, or delete them, and free text or model output cannot change state. Retrieval begins from an empty set and requires an exact project ID, approved status, complete evidence, and valid deadlines. Records are sorted deterministically and filled up to a configured token budget, never partially truncated. Any validation failure returns zero injected context. The model sees a single structured JSON message containing summaries and reference IDs only, with an explicit non-executable label, no raw evidence, credentials, or file paths, and no registry or approval tools.

It suits researchers writing a thesis with DeepSeek Harness who need to maintain citation conventions, advisor feedback, and chapter progress consistently across sessions. The current version supports a single local project, local and project scopes, and three storage classifications (public, project_confidential, restricted); restricted records are never injected. Team or user-level scopes, cloud synchronization, a web UI, external command execution, and institutional submission workflows are not yet available.