dsh-web-search-glm

If you are already running Zhipu GLM as the conversational backend in dsh, forcing the search step onto an entirely different provider is friction you do not need. dsh-web-search-glm closes that gap: it executes the native web_search server tool through GLM's Anthropic-compatible Messages endpoint and returns real, structured search sources rather than prose summaries or scraper fallbacks.

The outgoing request is byte-for-byte aligned with the official DeepSeek provider — same /messages path, same server-tool declaration bounded by max_uses, same AbortSignal wiring. Everything that differs lives in the response layer. GLM replies with paired server_tool_use and tool_result blocks whose content is a stringified Python-repr list; a small hand-written tokenizer parses that into {url, title, snippet} entries, deduplicates by URL, and merges multiple searches that appear in a single response. When no usable structured results are present, the provider fails cleanly with WEB_PROVIDER_ERROR and does not guess.

It is built for workflows that already orbit the ZAI / GLM ecosystem — one API key, one endpoint, the same Coding Plan billing. Both glm-5.3 and glm-5.3-flash are verified live, the overseas api.z.ai deployment is a one-line baseURL change, and every knob lives in the web-search-glm section of settings.yaml with zero modifications to your conversational settings required.