Skip to main content
TELOS stabilizes the request prefix (system / tools / conversation prefix) to win the KV cache. But each round, the agent also appends large chunks of tool output — bash / pytest / docker logs — to the tail of the conversation. TELOS absorbed the ideas of rtk-ai/rtk and added an orthogonal RTK output filtering layer: before a request enters the TELOS pipeline, it compresses the large repetitive output in tool_result.
The two lines are independent. TELOS shrinks what you pay full price for again (the prefix); RTK shrinks what you add as new tokens each round (the tool tail). They are controlled separately by the TelosMode four-state switch.

​
The four-state switch

@dataclass(frozen=True)
class TelosMode:
    telos: bool = True    # Run the TELOS pipeline (cache_control / ref-pool)
    rtk:   bool = False   # Run RTK tool result filtering
labeltelosrtkMeaning
none✗✗Pure passthrough — the proxy does not change a single byte
telos✓✗TELOS prefix caching only (proxy default)
rtk✗✓RTK tool filtering only, no cache markers applied
both✓✓Both enabled
An unknown or empty value degrades to the default telos, preserving the historical behavior from before the switch was introduced. Switch it live: 
telos mode both    # hot-reloads the running gateway and persists the choice

​
The filters

  • RtkFilter — shells out to the rtk binary (rtk filter --command <cmd> reading stdin). Any failure degrades to passthrough.
  • FallbackFilter — a dependency-free pure-Python filter: consecutive repeated lines folded into <line> (×N), head/tail truncation, pytest summary preserved. Guarantees the switch still takes effect when rtk is not installed.
  • CompositeFilter — runs rtk first; if it saves no bytes, falls back to the fallback filter.
  • build_filter() — rtk available → Composite(rtk, fallback), otherwise a pure FallbackFilter.
Thresholds: output shorter than 600 chars is not filtered; after dedup, output over 4000 chars goes through head/tail truncation.

​
How it’s wired into the proxy

  • The --mode CLI switch plus the X-Telos-Mode header (sticky to the session on the first request).
  • mode.rtk on → apply_filter runs before the TELOS pipeline.
  • mode.telos off → skip the pipeline and go straight to passthrough.
  • apply_filter(raw, flt) is a pure function: it deep-copies the request, rewrites the text of every tool_result (both the string and block-list content forms), and looks up the command hint from the previous assistant message’s tool_use.
  • usage_log gains the mode / compare_group / tool_output_reduction fields.

Architecture

Where RTK sits relative to the harness → bridge → engine pipeline.