The live in-memory diagnostic view — IR structure, PIN/FOLD/DROP distribution, cache breakpoint slots, and per-tool stats.
The developer page is the engineer’s counterpart to the savings dashboard. It renders the live
in-memory state from SessionInspector, re-rendering on every GET and lost on process restart.
GET /__telos/developer # HTMLGET /__telos/developer.json # JSON
Anthropic allows only 4 cache_control breakpoints, cut by the priority below. Physical order is
guaranteed by tools → system → messages, so the long TTL (1h) naturally precedes the short (5m).
Each row is a snapshot of one call (most recent 25 retained):
Column
Meaning
#
call index (monotonic since session start)
lat
call latency (seconds)
raw_in / cache_read / cache_write / output
the normalized usage quadruple
tools / system / messages chars · Δ
per-segment chars + diff from last turn
plan slots
the BP slot names actually placed this turn
uses / results
tool_use blocks this turn / tool_result blocks this turn
uses and results are off by one turn: the tool_use the assistant initiates in turn N generally
waits until the turn N+1 user to have its tool_result sent back.