Skip to content

v0.19.8

Choose a tag to compare

@github-actions github-actions released this 20 Nov 10:36
· 2 commits to master since this release
41d5000

0.19.8 (2026-11-20)

Features

  • emergent: a trace on every forge result, kept off forge_tool's output (2469b2f)
  • emergent: an onForgeTrace hook with a time bound and content-free log lines (1da8ee4)
  • emergent: build, freeze and deliver a forge trace (028c4bc)
  • emergent: forge_tool takes an optional justification and traces its own refusals (6805ae2)
  • emergent: keep the judge's dropped detail for the trace (14b350a)
  • emergent: the forge trace's types, stage codes and JSON-safe caps (f247977)
  • emergent: the wrapper's capture carries its call's trace (e3c0a72)
  • tools: the answering model on a tool call's context, and no error text in the quiet failure line (56e57d8)
  • tools: trace the orchestrator's refusals of its own forge_tool, and carry the justification to its approval request (deeb127)

Bug Fixes

  • emergent: forge trace review follow-ups (121c319)
  • emergent: keep the space after the colon in the effects_need_ceiling refusal (9de1b84)
  • emergent: read a BigInt array length as unconvertible, name the answered model on OpenAI abort and cut-off chunks, and state the fail_closed wait (ebbe676)
  • emergent: replace a malformed nested array in place, bound a failed round's hook wait once, and keep requested model ids (3d2fed2)
  • emergent: trace a refused forge call the round never reached, label a model id filled in from the request, and count array slots before reading them (61aa3a7)
  • emergent: walk each trace field once within a bound, and the remaining review follow-ups (6c13467)
  • providers: OpenRouter effort yields to reasoning_effort and follows customModelParams.model (c20d9d0)
  • providers: send effort to OpenRouter as reasoning.effort on OpenAI reasoning slugs (060d455)