Skip to main content

Context And RTK

Inferoa uses context settings and RTK to reduce token waste while preserving the evidence required for accurate work.

Context Settings​

context:
  compression_threshold: 0.75
  context_window: 256000
  # Optional automatic compact trigger overrides:
  # output_reserve_tokens: 16000
  # compact_buffer_tokens: 12000
  auto_compact_failure_limit: 3
  compact_recent_file_limit: 5
  compact_recent_file_token_limit: 5000
  compact_recent_total_token_limit: 25000
  protected_recent_loops: 3
  engine:
    provider: auto
    startup: welcome
    require_ready_before_chat: true
    watch: true

By default, automatic compaction is triggered from usable headroom: effective_window = context_window - output_reserve, then trigger = effective_window - compact_buffer. If output_reserve_tokens or compact_buffer_tokens are not configured, Inferoa derives bounded values from the context window. Setting compression_threshold to a value other than the default keeps the older ratio-based trigger as an explicit override.

Automatic compaction pauses after auto_compact_failure_limit consecutive model-summary failures. Manual /compact remains available. Compaction retries large summary requests as prefix-preserving, standalone, then trimmed standalone payloads before falling back to deterministic memory. Recent read file/resource evidence, the active plan, and invoked skills are attached to the post-compact epoch with the configured caps. protected_recent_loops keeps recent model and tool work visible. The context engine can use automatic detection, CodeGraph, built-in behavior, or be turned off.

RTK Settings​

rtk:
  enabled: true
  delivery: managed
  version: 0.42.3
  auto_download: true

Environment overrides:

INFEROA_RTK=false inferoa
INFEROA_RTK_PATH=/path/to/rtk inferoa
INFEROA_RTK_AUTO_DOWNLOAD=false inferoa

Operational Views​

/context
/compact [instructions]
/tokenmaxxing
/tokenmaxxing trend
/tokenmaxxing signals

/context reports context usage and compression state. /compact forces a conversation summary immediately and accepts optional summary instructions. /tokenmaxxing reports RTK savings, prefix-cache safety, compact boundaries, and token pressure alongside endpoint usage evidence. /tokenmaxxing trend shows pageable metric charts; /tokenmaxxing signals shows raw lifecycle and endpoint evidence.