zedarvates context_profiler — how much window is gone before you start?
On a modest machine the usable window is shared between the model weights' RAM and the KV-cache, and every always-on token is paid twice (RAM + each turn). This measures the prefix an agent carries before its first message and frames it against real local windows, so you can shrink it.
python -m skills.context_profiler.cli . # prefix tokens + % of 64k/128k/256k
python -m skills.context_profiler.cli . --json
Components measured:
- directives — CLAUDE.md / AGENTS.md instructions ([[metrics]] always-on).
- core_agent — the shared
core-agent.mdrules, if present. - tool_schemas — the MCP tool definitions injected into the agent (the hidden cost: on this repo ~3.8k tok for 38 tools).
- skill_catalog — the skills' descriptions IF the whole catalogue is injected.
It then reports the % of 64k/128k/256k windows and a reduction plan with honest token savings:
- lazy tool loading — expose ~5 core tools + a
find_tool(query)that loads a schema on demand (the pattern this very harness uses via ToolSearch). - on-demand skill search — don't inject the catalogue; use [[skill_finder]] / [[context_budget]] to load only the relevant skills per task.
On this repo: prefix 7.9k tok (12% of a 64k window) → **1.4k tok (2%)** once
lazy tools + on-demand skills are applied. Exposed via [[llm_mcp]] as
context_profile. Pure measurement, 0 cloud tokens.