Skip to content

v0.238.0

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 26 Oct 19:52
· 380 commits to main since this release

[0.238.0] - 2026-10-26

Release summary: 8 features, 25 bug fixes, 2 performance improvements, and 1 docs update.
Tag range: v0.237.21..v0.238.0.

Features

  • bench: set the Dreaming pass timeout for slower local models
  • memorybench: cap local answering and judging reasoning with a thinking budget
  • bench: override the Dreaming model context window
  • inference: cap thinking on local OpenAI-compatible targets
  • memorybench: send enable_thinking with the local reasoning effort
  • bench: set the Dreaming per-turn output budget for local reasoning models
  • memorybench: run answering and judging on local OpenAI-compatible models
  • daemon: raise LLM and Dreaming concurrency ceilings to 256

Bug Fixes

  • inference: discover the window only for agents
  • bench: bound archive retries and clean temp files
  • dreaming: report activeAgentId only for that agent
  • bench: harden the RunPod kit from review
  • memorybench: cap waits at the retry deadline
  • dreaming: scope dream status to the agent
  • transcripts: repair a torn tail on append
  • inference: read the served context window
  • dreaming: cap evidence delivered per pass
  • dreaming: bound scheduled content sweeps and size passes by maxPassTokens
  • dreaming: size passes so they finish and commit instead of livelocking
  • dreaming: treat a DB owner deadline on trigger as retryable
  • pipeline: count special-token literals in user text instead of throwing
  • dreaming: return a retryable 503 when a trigger hits DB owner pressure
  • memorybench: resend dropped requests on a fresh connection and name the failed path
  • memorybench: keep every question when resuming an unrestricted run
  • daemon: page session-claim restore so a busy daemon can restart
  • memorybench: keep honoring Retry-After while the capture backlog drains
  • daemon: treat a full transcript-capture backlog as retryable write pressure
  • memorybench: resend requests the daemon dropped on a pooled connection
  • inference: take local targets' context window from the server
  • memorybench: create each haystack agent once and retry Signet write pressure
  • daemon: surface write pressure as retryable 503s instead of 500s and dropped captures
  • dreaming: report running passes in /api/dream/status
  • bench: restore the loopback endpoint check for local Dreaming models

Performance

  • transcripts: scope the per-agent canonical backfill query to that agent
  • transcripts: append a new session's snapshot instead of rewriting the canonical JSONL

Docs

  • record the Qwen3.8-27B vLLM recipe for one Blackwell GPU