Usage Evidence 2

Usage Evidence (Part 2)

This page gives more of the usage facts that claudetto relies on, from the code graph add-on to the tools that claudetto does not use. Each fact has a source label: official, binary, measured, reported, or inference.

Code Graph Add-On Against codegraph install

Method: On 2026-10-18, with no model call, from the code and text of the add-on before and after the change. codegraph was not on the PATH, so codegraph install did not run. The install facts come from the vendor source (dist/installer/instructions-template.d.ts).

Part of claudetto-codegraph	What codegraph install alone gives	Result
SessionStart hook and its 46-word line	The CLAUDE.md block and MCP instructions name the same commands	Removed
Sync on start, after an edit, and after git	codegraph syncs its own index (the user, 2026-10-18)	Removed, about 450 lines and 3 hook entries
--mcp option of the check, MCP_ADD_COMMAND	codegraph install writes the MCP server	Removed
Skill codegraph (466 words)	MCP tools cover the same questions, but subagents get only the block	Kept, for use with no MCP server
claudetto-codegraph-check	codegraph status shows the index, but a missing CLI is a shell error	Kept, not measured against the live CLI

The add-on went from 658 lines to 201 lines and has no hook, so it adds no process to a session.

Model Calls of a Hook

• In all 4 runs of an earlier check, total_cost_usd did not count the model call of an agent hook (measured, external/cc-diff-299-300/live/live-checks.md, check 6).
• So a hook that calls $.model.complete reads the usage of the result and applies the list prices.
• Two calls to a small model on 2026-10-16 (Claude Code 2.1.301) used 798 input and 68 output tokens at $0.0023, and 826 and 11 tokens at $0.0018, at the list prices of Sonnet 5.5. Neither read the cache.

Community Claims

Claim	True	Note
/compact reads the whole conversation	✅	Expensive on a large context. /clear costs nothing (official)
MEMORY.md loads at every start, up to 200 lines or 25 KB	✅	Topic files do not load at start (official)
A short CLAUDE.md saves usage	✅	It loads at every start and in every subagent that does not set omitClaudeMd
Subagents keep large reads out of the main context	✅
Old .jsonl transcripts in ~/.claude/projects/ enter the context	❌	Only /resume loads one. The snippet does not set cleanupPeriodDays (Settings Snippet Keys)

Tools That claudetto Does Not Use

Tool	Reason
RTK	A JetBrains measurement of 7.6% more cost (reported)
Graphify	Fails on large repositories (reported)
grepai	Needs embeddings from Ollama or OpenAI, and has no measurement

See Also

• Usage Evidence
• Evidence