Skip to content

v0.1.46 — DSH 0.2.0-rc.2 compatibility

Choose a tag to compare

View all tags
@Anionex Anionex released this 01 Oct 04:49
v0.1.46
b825d03
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Compatibility

  • Verified DSH 0.2.0-rc.2 using an isolated, hoisted tarball Profile, actual visual-tool calls, and Web Settings. Add an exact peer/compatibility declaration and a required CI Profile lane for this release. The fixture explicitly selects its OpenAI-compatible protocol after the host's default provider protocol changed.
  • 已通过 DSH 0.2.0-rc.2 隔离压缩包 Profile、真实视觉工具调用及 Web 设置验收；增加确切兼容声明及 CI 验收矩阵。官方默认协议变更后，测试 fixture 显式选择原有 OpenAI 协议。核实时官方尚无正式稳定版。

Added

  • Added an explicit OpenAI Responses protocol for custom vision providers. An optional, constrained reasoningEffort is forwarded only for Responses requests, exposed conditionally in bilingual Settings, and included in both live-Session and durable evidence cache identities. Existing OpenAI configurations remain on Chat Completions and omit effort by default.

  • Added hardened provider.headers for non-secret deployment metadata and provider.sessionHeaders for gateways such as OpenCode Zen that require per-session routing. Session values are process-keyed HMAC identifiers that never expose raw Session ids or workspace paths; configured headers are bounded, conflict-checked, error-redacted, cache-isolated, restricted to the provider origin/base path, and shared by both requests in a connection-and-model health operation. The manifest-verified vendored snapshot remains unchanged (#144).

Fixed

  • Restored startup and live Settings on DSH 0.1.7, whose Config projection replaced the old settings.register/get/watch API. The plugin now uses the host's matching Schema form, unwraps live configuration values for runtime use, and follows Settings document updates. Earlier DSH versions continue using their registration API.
  • Vision API subprocesses now try IPv4 addresses before IPv6 when both are returned by DNS. This avoids exhausting the connection timeout on unusable fake IPv6 addresses in TUN proxies while preserving IPv6-only endpoints.
  • Stopped the image-input variant route from crashing on request messages that carry no source. The variant read message.source.kind while assembling the evidence pass, but the official createUserMessage() helper writes no source at runtime (its published type still declares one as required), so a plugin that built a programmatic ctx.llm.stream() call on a vision-toolkit-* route — the usual summarize, title, or classify pattern (@modusensus/dsh-mneme ≤ 0.8.1 did exactly this) — failed with TypeError: Cannot read properties of undefined (reading 'kind'), which the host surfaced only as a generic UNKNOWN failure chunk. A user message without provenance is now treated as a user turn, and assistant history without provenance is skipped instead of throwing; messages that do carry a source behave exactly as before.

Documentation

  • Documented /responses endpoint composition, common and provider-specific effort values, possible token/latency/cost effects, and the need to review provider data-retention policy even though requests send store: false.

Verified host: DSH 0.2.0-rc.2. No official stable DSH version was available at verification time.

Assets 2
Loading