igrock88/dsh-llm-gigachat
Connects Sber GigaChat to DSH through a built-in OAuth2 proxy, so the provider lands on Settings → Models like any other.
Adds a working Sber GigaChat provider to the standard Models page: the plugin runs a local OAuth2 proxy at 127.0.0.1:8787/v1, creates the provider card with the openai-completions protocol and a pre-filled model list, and keeps the authorization key out of settings.yaml — it lives in the managed store ~/.dsh/.credentials.yaml as SBER_API_KEY, with the card holding only a reference. It is careful not to fight an existing setup: if an sber route already exists from the older standalone-proxy approach it is not overwritten, except that a baseURL still pointing at another loopback port is redirected at the local proxy. Diagnostics are exposed through a /stats endpoint that counts served requests, rate limits, server errors, degenerate retries, translated tool calls and queue depth, and the README's troubleshooting table maps symptoms — tools unavailable, no boot — to the specific row or setting that causes them.
Install
dsh plugin --profile web add dsh-llm-gigachatThe npm install line. After it, the README says dsh must boot right away and sends you to a Rollback chapter if it does not — the plugin never patches other entries, so undoing it is one removal plus, at most, deleting an llm-pi-ai.providers.sber route that this plugin created. Verification without starting anything: dsh --profile web --dump-config should show a row id: llm-gigachat and leave the llm-pi-ai config untouched.
Compatibility
Sber uses OAuth 2.0 client credentials, a two-step key-for-token exchange that DSH's built-in OpenAI-compatible gateways do not implement, which is why configuration alone cannot connect it and code is required. The plugin ships that code: key→token exchange, 30-minute token caching, request serialization (personal plans allow about one concurrent request and otherwise answer HTTP 429), retry of degenerate "<" answers, and translation of tool calls into the legacy functions/function_call format GigaChat 3 understands. Model choice is a hard constraint, not a preference — the README states only GigaChat-3-Ultra and GigaChat-3-Pro reliably handle agentic tool use, while lighter models reject complex schemas or hallucinate tool arguments, so they run without tools by default via stripToolsFor. TLS verification is off by default because the Russian NCC root certificates are not in Node's default trust store on Windows; the README gives the hardening path rather than leaving it implicit.
Details
- Repo: igrock88/dsh-llm-gigachat
- Category: Models & Providers
- Stars: 0
- Version: 0.1.1 (npm)
Alternatives
ffyuuu/dsh-llm-longcat · cyjyyd/dsh-llm-xai-oauth · NOirBRight/dsh-llm-ollama