Skip to content

v1.3.124

Choose a tag to compare

View all tags
@topcheer topcheer released this 04 Jul 06:55
· 3673 commits to main since this release
v1.3.124
4e4280c

v1.3.124

Released: 2026-07-04

Improvements

Context Window Usage Metrics in Reflection

Based on 2025-2026 context engineering research (Anthropic, Manus, LangChain),
context utilization metrics are essential for understanding agent performance
degradation in long sessions.

The reflection system now tracks and reports:

  • Peak tokens: highest token count observed during the run
  • Context utilization: peak tokens as percentage of context window
  • Compaction events: number of compaction triggers (precompact, reactive, autopilot guard)

These metrics appear in the "Context usage" section of run reflections, giving
future sessions visibility into how context pressure affected previous runs.

Bug Fixes

  • EndpointMetrics unbounded growth: AppendMetricForEndpoint appended a metric
    event on every LLM API call without any cap. In long-running sessions (hours/days),
    per-endpoint metric slices grew unboundedly, causing increasing memory consumption.
    Capped to 200 most-recent entries per endpoint key.
  • OOM kills in CI tests: Added GOMEMLIMIT and reduced test parallelism (-p 1)
    to prevent OOM kills during make test on memory-constrained CI runners.
Assets 51
Loading