dsh-focal

In a multi-turn DeepSeek Harness agent session, stuffing the full history into every model request wastes the context window and pushes sensitive material into a network hop that may never need to happen. FOCAL (Filtered On-device Continuous Activity Logging) addresses exactly that: it watches committed session records, routes user messages into task-isolated memory buckets via a small deterministic similarity filter, and injects only the active task's bounded context snapshot into the next request. Everything else stays in a local file and never leaves the machine with the prompt.

The filtering is selective rather than aggressive. It records tool names, call success or failure, and path-like arguments, but deliberately never captures assistant token streams, raw tool arguments, or full tool-result payloads. Common credentials and email addresses are redacted before anything is written, and absolute paths outside the workspace are reduced to their basename. Each session ends up as a single bounded JSON document under the local DSH_HOME directory, with zero network calls and no telemetry. On the model side, focal_status and focal_recall expose two read-only tools, while the human-facing focal command retains sole authority over deletion.

If you run several tasks side by side in daily development and would rather not ship your entire conversation to a remote model, or if you want a privacy-boundary-aware memory layer that stays on disk, FOCAL is currently the only host plugin in the DSH ecosystem that combines task isolation, bounded injection, and an anti-injection policy section. It is at v0.1.1 (experimental), ships as a host-only plugin with no Web UI yet, and is best suited to developers comfortable tracking the DSH 0.1.0-rc.6 interface.