ReleaseSiliconFlowSiliconFlowpublished Sep 14, 2026seen 4w

siliconflow/dsh-llm-siliconflow v0.1.0-rc.7

siliconflow/dsh-llm-siliconflow

Open original ↗

Captured source

source ↗
published Sep 14, 2026seen 4wcaptured 4whttp 200method plain

0.1.0-rc.7 — tool-call parsing hardening

Repository: siliconflow/dsh-llm-siliconflow

Tag: v0.1.0-rc.7

Published: 2026-09-14T05:48:41Z

Prerelease: yes

Release notes: Tool-call parsing hardening plus wizard guidance, same day as rc.5/rc.6.

Changed

  • Translation of wire tool_calls: call.id and function.name are now only accepted as non-empty strings (previously any present value was copied through — malformed provider payloads could inject garbage ids).
  • README: documented the agent-default-model write rule — the wizard writes only provider + model, and a leftover reasoningEffort (e.g. switching from deepseek-official) is rejected with UNSUPPORTED_REASONING_EFFORT before any network I/O. Manual settings.yaml edits must remove that field.

Evidence (byte-level tarball diff 0.1.0-rc.6 → 0.1.0-rc.7)

  • chunk: if (call.id !== void 0) → if (typeof call.id === "string"); function.name now additionally requires length > 0
  • README: one hunk, the reasoningEffort paragraph quoted above
  • everything else identical (entry files differ only by the chunk-hash rename types-Bim1q3qs.js → types-CmhCs_oH.js)

Provenance note: tag created retroactively from GitHub Releases; the attached tarball is the actual npm pack bits.