context_profiler
Skill2 starsChanged 3 months agoMITView on GitHub
Measure a project's always-on prefix (agent directives + core rules + MCP tool schemas + skill catalogue) in tokens and as a % of small local-model windows (64k/128k/256k), with a concrete reduction plan (lazy tool loading, on-demand skill search). Use to see how much of a modest machine's usable context is spent before any real work, and how to shrink it so weaker machines can run local LLMs usably.
In this repository75 other files it gives its agents.
- AGENTS.mdAGENTS.md
- clean-cacheSkill
- agent-cacheSkill
- app_testSkill
- asset-qualitySkill
- auto_distillSkill
- auto-memorySkill
- auto_routerSkill
- benchSkill
- bootstrapSkill
- botte-proxySkill
- cacheSkill
- capabilitiesSkill
- cardinalSkill
- cavemanSkill
- checkupSkill
- clarificationSkill
- clusterSkill
- code_fingerprintSkill
- completion_proofSkill
- conductorSkill
- context_budgetSkill
- context_slicerSkill
- context_windowsSkill
- control_loopSkill
- cost_estimatorSkill
- cwe_kbSkill
- dag_optimizerSkill
- dashboardSkill
- decision-ladderSkill
- demoSkill
- diff_languageSkill
- directives_auditSkill
- docgenSkill
- docs_stewardSkill
- eventsSkill
- fallow_likeSkill
- fast_contextSkill
- fixSkill
- harness_deltaSkill
- hermes-bridgeSkill
- hf_provenanceSkill
- infra_advisorSkill
- ingestSkill
- llm_backendsSkill
- llm_mcpSkill
- loaderSkill
- mcp_gatewaySkill
- meta_harnessSkill
- metricsSkill
- migration_auditSkill
- monte_cristoSkill
- mousquetairesSkill
- nlp_deterministicSkill
- nn_auditSkill
- prefix-prunerSkill
- prefix_treeSkill
- preflightSkill
- prompt_improverSkill
- reportSkill
- The first 60 shown.
The file
Files
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 . --jsonComponents 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.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.