[CaffeinatedCoder]

Closes #167

Problem

AGENTS.md loads into every agent session in this repository: 22,230 characters, about 6k tokens, most of the roughly 10k tokens of memory files a session starts with. Five sections repeated what another file holds, or detailed what an agent needs only at one moment.

Change

Section	Before	After	Where the rest lives
Finding code	3,693 chars	2,310	design/code-graph.md (new, linked from design/plan.md): the graph's measured gaps from 2026-10-13, and refreshing it
Commands	3,691	2,825	the guard's own refusal message; decisions 92 and 105
Issues	1,808	1,032	CONTRIBUTING's "Issues and labels" and "Pull requests"
Workflow	3,223	2,576	step 7's mechanics: test/oracles/README.md
Releasing	1,210	608	CONTRIBUTING's "Releasing", step 1

Each section keeps what an agent must know before it acts:

• the graph leads and the source decides, and an empty result proves nothing;
• what the verdict prints and what the guard refuses;
• sub-issues, branch names and closing an issue by hand after a release-branch merge;
• a promise gets a claim or a waiver;
• the arm64 TYPEKIT_MSSQL_IMAGE, and never deleting a seed or reading an ERROR as caught.

"Rules that are easy to break" is unchanged, and so is every heading another document cites: decision 93 cites "Finding code".

AGENTS.md goes from 22,230 to 17,956 characters, about 1.1k tokens less per session start. That's about half the 2–2.5k #167 estimated; the rest of those sections is what an agent needs up front.

Tests

• CodoMetis.TypeKit.Conventions.Tests: 2,002 passed.
• docs-query check: 0 problems in 854 links.
• Each dropped sentence was checked against the file it now points to.

🤖 Generated with Claude Code