electric-sql/electric @electric-ax/agents-runtime@0.6.3

on GitHub

Patch Changes

• 886c423: Context compaction for the agents runtime. Modelled on OpenAI Codex's
  summarization but adapted to the event-sourced timeline (a context_inserted
  checkpoint placed at a stored watermark, so reconstruction folds older messages
  into a summary):

  • A context-window usage gauge (cache-inclusive context_input_tokens +
    context_window persisted per step) and <token_budget> notices injected at
    25 / 50 / 75% usage.
  • Oversized tool-output truncation, and a synchronous mid-turn compaction floor
    at the 90% hard ceiling (runs before every model step via the adapter's
    transformContext hook).
  • Non-blocking background (turn-end) compaction that starts at 85%: a detached
    summarize whose checkpoint is applied at the next turn's start, or immediately
    if it finishes while idle. Each generation uses a watermark-unique checkpoint
    id so a new run can't supersede a prior completed one. Summarize calls are
    bounded by a hard timeout.
  • UI: a "Compacting…" indicator (blocking vs. background) and a collapsible
    "Context compacted" entry in the conversation timeline.

  Thresholds are env-tunable (ELECTRIC_AGENTS_COMPACT_CEILING,
  ELECTRIC_AGENTS_COMPACT_BG_CEILING).

• 7139c53: Context-usage indicator: a circular ring gauge plus a hover breakdown popover
  showing how the prompt decomposes across system prompt, tools, messages, and
  free space (à la Claude Code's /context). The runtime now persists an
  approximate context_breakdown of the stable request parts (system + tools) on
  each step alongside the cache-inclusive total, and exports new
  computeContextBreakdown / parseContextBreakdown helpers from the client
  entry; the UI derives the "messages" bucket as the real remainder so the
  segments always sum to the gauge.