dsh-dsml-artifact-guard

When routing through certain upstream model providers or API gateways, raw DSML tool-invocation closing tags can leak into the assistant-visible text stream. These stray artifacts pollute the chat bubble, trigger Markdown rendering glitches, and may confuse downstream agents or break clipboard exports. dsh-dsml-artifact-guard is a lightweight, host-only stream interceptor for DeepSeek Harness that strips these terminal tags in real time before they reach the user interface, leaving legitimate conversation content completely untouched.\n\nThe plugin is built around four design pillars: it strictly honors the Cordis synchronous stream-hook contract so that an accidental asynchronous wrap never crashes the dispatcher with a TypeError; it maintains a 96-byte sliding buffer to reliably catch protocol fragments split across TCP or WebSocket deltas; it enforces a 100 percent fail-open policy, guaranteeing that genuine prose about DSML syntax, tool-call deltas, and usage tokens are never removed or delayed; and it scopes processing to specific providerId and modelId pairs, passing all unrelated traffic through with zero overhead. Three runtime modes are available: sanitize (default, strips tags and logs the count removed), audit (emits diagnostic logs without altering the visible stream), and disabled (full bypass).\n\nThis plugin is aimed at DeepSeek Harness administrators and developers who encounter tag leakage when connecting to particular upstream providers. No gateway reconfiguration or upstream code changes are needed-enabling the guard in your local harness instance is enough to restore a clean chat surface and keep downstream pipelines free of protocol noise.