This page holds the decisions D6, D8, D9, D22, and D26. Read it to learn why claudetto uses each tool, file, and feature in the form that it has. The index is Decisions.

Each decision gives the problem, the choice, the alternatives, and the evidence. Each number has a source in the code, in an evidence file, or in a binary fact with its version.

The problem is that the status line shows the session state under the prompt. It must stay small. It must read no file of Claude Code that the docs do not name.

The choice is one statusLine command. It is the documented extension point. The code is in src/plugins/claudetto/slices/statusline/. Its files import only src/kernel/ and their own folder.

The command reads the status object from stdin. It reads the active change of MiniSpec from minispec/changes/<name>/tasks.md (adapters/change.ts). It makes one git status call (GIT_ARGS in domain/git.ts). It caches the result of that call in a claudetto-git-<key>.jsonl file in a folder of the temp folder of the user (adapters/git-cache.ts). The cache holds the result for GIT_CACHE_MS.

The alternatives are the old state files of the status line, the reset credits segment, and the /context breakdown. They are removed. The code has no word credit and no word breakdown in the slice. So no segment reads them.

The bounds are in slices/statusline/limits.ts:

• MAX_ROWS is 3.
• REFRESH_SECONDS is 1.
• GIT_TIMEOUT_MS is 1500.
• GIT_CACHE_MS is 3000.
• DEFAULT_COLUMNS is 120, and COLUMNS_MARGIN is 4.
• SUBAGENT_ROWS_MAX is 20.

The bounds in src/kernel/budget.ts are these:

• CONTEXT_WINDOW is 300000 tokens. It is the upper bound of the window that the status object reports.
• USAGE_LEVELS is 75 and 90 percent (paintOf in color.ts).

renderRows in domain/rows.ts fits the segments into at most MAX_ROWS rows of COLUMNS width. NO_COLOR turns off each color, and the icons stay (colorOff in color.ts). The status line has no progress bars. That choice is the user's decision of 2026-10-12 in the dotclaude 1.0.0 change.

The basis of the slice changed in 0.1.0: it was rebuilt, and it now shows the limit windows with a pace, the prompt cache, and the change of MiniSpec. See openspec/changes/claudetto-0-1-0/features.md and the page Status Line. Also see Undocumented Reads.

The problem is that the model needs the API of a library version that the project does not show.

The choice is the docs skill of the add-on claudetto-docs. The model invokes it. The code is in src/plugins/claudetto-docs/slices/docs/skills/docs/SKILL.md. The skill runs [email protected] through bunx -y, with CTX7_TELEMETRY_DISABLED=1. Its allowed-tools rule allows only the two read forms, library and docs. The skill makes at most 3 ctx7 calls for one question.

The alternative is an unpinned ctx7@latest. The pin keeps the allowed command the same until a release changes it. The settings snippet for the add-on has the same runner and version, and it allows the two hosts context7.com and registry.npmjs.org (see Settings Snippet).

See Library Docs.

The problem is that claudetto needs checks that run the plugin in a real session.

The choice is the claude plugin eval command of Claude Code. Claude Code owns the case format, so claudetto keeps no harness of its own. Each case has a case.yaml and a scaffold.sh. src/plugins/claudetto/evals/ holds 3 cases. The evals/ folders of the slices of the core plugin hold 10 more. The add-on claudetto-you-should-know holds 1. The recipe eval in justfile runs claude plugin eval plugins/claudetto on the path of the checkout. It uses --max-cost-usd 5, --keep-temp, and a lock from tools/lock.ts, and it needs --model. The recipe eval-variant compares the checkout with a variant at a git ref. The add-on claudetto-lab has the eval-compare slice. That slice runs claude plugin eval on two plugin folders in turn. The type in src/kernel/slice.ts says that a hook that runs a model names its claude plugin eval case in eval.

The alternative is a Bun harness with its own case format, its own oracle, and its own reports. It is not in 0.1.0, because the command of Claude Code covers the cases (inference).

The binary fact is that the strings of the 2.1.296 bundle name the plugin eval subcommand. A text search of the bundle found this fact.

See Evals.

The problem is that each read adds text to the context of a subagent or of the main session. The context grows with each read. The bound is a choice. No measured figure in the repository supports its value (inference).

The parts are these:

1. A search-first read rule. The fragments src/fragments/search-route.md and src/fragments/read-from-project.md are included in the agent bodies. The agent bodies are in src/plugins/claudetto/slices/agents/agents/.
2. A context bound for each subagent, with a report-now message. slices/agents/limits.ts gives SUBAGENT_CONTEXT_MAX (100000 tokens) and HAIKU_CONTEXT_MAX (85000 tokens). The hook tells the run to report now when the bound is reached.
3. No maxTurns in an agent file. A run at its turn limit returns no report, and a test checks that no agent file has maxTurns (tests/claudetto/agents/roster.test.ts).
4. Delegation text. The description of the investigator agent says when to delegate a question that spans several files.
5. A Bash deny for a command that only prints a file. slices/guard/domain/reads.ts parses the print. deny.ts turns it into a deny.
6. A report bound. REPORT_MAX_WORDS is 400 in slices/agents/limits.ts. The report-size fragment states it.
7. bashOutputMaxChars at 16000. BASH_OUTPUT_MAX_CHARS in src/kernel/budget.ts gives the value. slices/prompt/slice.ts sets it in the settings. The settings schema of Claude Code 2.1.302 clamps the key to a range of 4000 to 128000, with a default of 30000. An output above the cap goes to a file. Claude gets a preview and the path. The comment of BASH_OUTPUT_MAX_CHARS says that 16000 brings 196 of 240 saved outputs inline. The dotclaude 1.0.0 value was 4000, its lowest value. 0.1.0 changed it, and openspec/changes/claudetto-0-1-0/design.md has the basis of the sweeps.

The alternatives are these:

• A lower BASH_MAX_OUTPUT_LENGTH was the first choice. It cut build output. No evidence file in the repository checks this reason.
• A filter in updatedInput for test output is not built for now. No evidence file in the repository checks the reason. The reason is that test output was small.

bashOutputMaxChars replaces the first choice, because it keeps the full output in a file.

See Handoff.

The problem is that Claude needs a code graph of the project, such as the callers of a function.

The choice is the add-on claudetto-codegraph. It uses the codegraph CLI and no MCP server. The code is in src/plugins/claudetto-codegraph/slices/graph/. Each hook runs only when .codegraph/codegraph.db exists (INDEX_PATH in domain/graph.ts). The hooks in hooks/hooks.json do three things:

• They sync the index at SessionStart, and they add one line about the graph.
• They sync the index after an Edit, Write, MultiEdit, or NotebookEdit.
• They sync the index after a git command that changes the tree (a mod handler in adapters/mod.ts).

The add-on runs codegraph by its name, and a missing CLI gives ENOENT in adapters/codegraph.ts. The bounds are in slices/graph/limits.ts. QUERY_TIMEOUT_MS is 2000. START_SYNC_TIMEOUT_MS is 6000. With HOOK_MARGIN_MS (1000), the total is 7 s, which is 3 s under the 10 s hook timeout (SESSION_START_HOOK_TIMEOUT_S).

The alternatives are these:

• An MCP server. It is not the default, because a skill with a CLI costs less at the start of a session. The MCP server is the choice of the user (Code Graph).
• The built-in LSP tool. The comparison with it is not in the repository. So the page cannot say which one to ship. This is open.

See Code Graph.

• Core Plugin
• Plugin Layout
• Decisions