GooDAnDReaDY/dsh-dsml-artifact-guard
A fail-open stream interceptor that strips leaked DSML protocol closing tags out of the assistant's visible text before they reach the user.
The README describes the problem this solves: when interacting with certain upstream model providers or API gateways, raw DSML (DeepSeek Markup Language) tool-invocation protocol tags can leak into the assistant's visible text stream, polluting the chat bubble and breaking Markdown rendering. @goodandready/dsh-dsml-artifact-guard is a lightweight, host-only runtime stream interceptor that cleans those terminal artifacts in real time before they reach the user. Its README lists four design points: synchronous stream contract preservation (interceptors must return an AsyncIterable, not a Promise), a split-chunk buffer pipeline with a 96-byte sliding buffer so tags split across deltas still match, 100% fail-open safety so legitimate text is never dropped, and targeted provider/model scoping so other traffic passes through with zero overhead.
Install
dsh plugin --profile web add @goodandready/dsh-dsml-artifact-guardCopied verbatim from the README's install block. The README's audit mode emits ctx.logger.info(...) diagnostics without modifying the user-visible stream; the default mode strips the leaked tags.
Compatibility
The README's badges state Node 20+ and the DSH plugin topic. It is a host-only runtime stream interceptor, so it processes the assistant's output stream inside the harness rather than requiring a client half.
Details
- Repo: GooDAnDReaDY/dsh-dsml-artifact-guard
- Category: Coding & Development
- Stars: 3
- Version: GitHub repo GooDAnDReaDY/dsh-dsml-artifact-guard (MIT, per the README's License section); npm package @goodandready/dsh-dsml-artifact-guard at 0.2.9 (registry check 2026-10-07; the README shield shows the same package).
FAQ
- Could it delete legitimate text?
- The README states a 100% fail-open design: it never drops legitimate user or assistant text, and discussions about DSML syntax are preserved intact.
- Does it process every model's output?
- No. The README says it scopes processing to the provider and model configurations that exhibit tag leakage, passing other traffic through with zero overhead.
Alternatives
030611/dsh-context-provenance · 23swccp/dsh-undo · 121103qwq/dsh-vision-sidecar