dsh-cognition

Every new DSH session starts from a blank slate: the agent has no recollection of files changed last week, which approaches were already validated, or which edits crossed the boundary of the current task. dsh-cognition (DSNLE) injects project-level memory into the agent — past edits on similar tasks surface as precedents, out-of-scope changes are hard-blocked, and co-change patterns, churn trends, and rollback history accumulate as cross-session knowledge. All of this is enforced through DSH-native mechanisms (per-tool pre-execute gates, kernel-level version guards, the skills knowledge layer, session event sourcing) rather than prompt engineering.

The core workflow is an auditable ritual chain: each task passes through constraint locking, content-hash observation, retrieval recall, and a reliability contract gate. Out-of-order code mutations are rejected outright. An optional embedding-based semantic channel (MiniLM-L6, 384-dim) degrades gracefully to BM25 plus symbol and precedent retrieval when disabled, leaving the main pipeline unaffected. Publicly documented evidence includes 47 real-repo fixes verified by upstream test suites, a gold-free 20-task continuous run (20/20), and 245 deterministic regression assertions.

Best suited for developers who run DSH agents long-term and need cross-session consistency and traceability, especially when maintaining codebases of moderate-to-high complexity where the agent must remember project context instead of re-exploring from scratch every time.