Decisions

Decisions

This page is the index of the design decisions of claudetto. Read it to find the page that holds the detail of a decision. Each row gives the number, the title, and the page with the detail. Design holds the principles.

The ids D1 to D47 come from dotclaude 1.0.0 and stay as they were. The ids from D48 are new in claudetto 0.1.0. Engine Facts holds the facts of the engine that a change must not break. A missing id is a removed decision. The removed ids are D7, D11, D19, D20, D21, D32, D33, D35, D36, D40, and D41. They described the removed design of dotclaude 1.0.0, such as the spec store, the append-only event log, and /dotclaude:handoff. A kept decision whose basis changed in 0.1.0 says so in its section and points to openspec/changes/claudetto-0-1-0/design.md.

The restart has its own list of decisions in openspec/changes/claudetto-0-1-0/design.md. Its numbers D1, D2, D8, and the others are not the numbers on this page. This page always writes "(D3 there)" when it cites that file.

Decision	Title	Page
D1	Rule of Least Power	Settings
D2	No guard route, settings.json only	Settings
D3	Tool-description and section cuts by id	Prompt
D4	Measured budget	Prompt
D5	Who invokes what	Settings
D6	Status line	Tools
D8	Context7 skill	Tools
D9	Evals through claude plugin eval	Tools
D10	Schema extraction	Settings
D12	Agent frontmatter search	Router
D13	No router, the main loop keeps the session choice	Router
D14	Off switches	Settings
D15	claudetto owns the system prompt	Prompt
D16	Request capture and the session start	Prompt
D17	Plugin and marketplace layout	Layout
D18	Separation of concerns, size bounds, and bound tiers	Layout
D22	Context bound for subagents and the main session	Tools
D23	Development frictions	Layout
D24	Add-on plugins stay separate	Layout
D25	Agent roster	Router
D26	Code graph through the CodeGraph CLI	Tools
D27	Measured facts over model reports	Evidence
D28	Compliance by checks, not by prose	Prompt
D29	Semantic line breaks	Evidence
D30	Hold work that is right under an unsupported claim	Evidence
D31	Guard hooks fail open	Hook Events
D34	Adherence by checks in code, with no new judge (now the asks slice, see D43)	Architecture
D37	Ports and adapters inside vertical slices	Architecture
D38	A generated composition root	Architecture
D39	One hook module in place of classic command hooks	Architecture
D42	Zero token cost where a hook, a script, or a CLI can do the work	Architecture
D43	asks only on a weaker check or discarded work	Architecture
D44	Jev is a slice of the core plugin, with a small model as the fallback	Architecture
D45	Add-ons import nothing from the core plugin or from each other	Architecture
D46	Bounds in three tiers	Architecture
D47	Arrow functions, plain data, and classes only for state	Architecture
D48	User notices through $.ui.log, not a transcript row	Engine Facts
D49	No catch-all in a hook that calls next first	Engine Facts
D50	A marker in $.env for each Bash call of Claude	Engine Facts
D51	The sandbox is the hard stop, and setup writes it	Engine Facts
D52	Each adapter keeps its own $ calls	Engine Facts
D53	A plugin verdict only makes a call stricter	Engine Facts
D54	The attribution reminder in place of includeGitInstructions	Engine Facts
D55	No session-wide env through CLAUDE_ENV_FILE	Engine Facts
D56	Toolchain facts come from the environment, not from model memory	Engine Facts
D57	A plugin bin/ command is named by its path, not by its bare name	Engine Facts
D58	Stop sees the reply after the user and gives the background tasks	Engine Facts
D59	A local background command has no time limit	Engine Facts
D60	The built-in read dedup covers only the same range	Engine Facts
D61	A tool.call hook sees the outcome of the permission dialog	Engine Facts
D62	Start context, served model, and fetch timeout come from three documented points	Session Facts
D63	The cost of the judge comes from its own usage	Session Facts

Related Pages

• Design
• Plugin Layout