These decisions explain the shape of the system prompt, the cuts to built-in text, and the limits on result size.

Choice: The prompt slice answers by exact id. An unknown id passes on unchanged.

• prompt.compose replaces the text of lean_body with the tailored text, because Claude Code 2.1.302 and 2.1.303 never send that section to prompt.section.
• prompt.section answers the ids that have a transform in slices/prompt/transforms/.
• tool.describe replaces the description of Agent when the new text is shorter. ListAgents is deferred behind ToolSearch.
• prompt.attachment leaves out goal_status and tool_hosts_notice.
• command.describe hides 29 commands from the typeahead and /help. A hidden command still runs when typed in full.
• The context_management section follows the compaction state: handoff, built-in, or off. With built-in, Claude Code text stays as written.

Reason: Some built-in text does not help claudetto and costs tokens on every request. Nothing cuts session_context or prompt.context, because a cut needs an eval with no loss first.

Evidence: slices/prompt/lab/ tests the handlers. tools/builtins/verdicts/ holds the verdict for each part. See Prompt.

Choice: The prompt slice builds on the short prompt (CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=1). It replaces the text of lean_body with one text for every model, through prompt.compose. LEAN_BODY joins the sections intro, doing_tasks, system, blocks, verification, actions, tools, and counts.

• The text is the same on each turn, so the prompt cache stays valid.
• A subagent gets its prompt from Claude Code, so agent.spawn adds a <claudetto_notes> block once. It holds the verification, block, and evidence lines.
• An output style reaches the model as a message, so Concise holds only role, tone, and format.
• Rules are plain imperatives. A reason appears only where it changes behavior.

Reason: The same rules are needed for every model, and one text serves them all.

Evidence: tests/claudetto/prompt/rules.test.ts pins the least-code rule, the pushback tiers, and the missing-tool rule. slices/prompt/lab/section.test.ts checks that prompt.compose replaces the lean_body section and keeps the other ids.

Choice: claudetto-capture is a local logging proxy in src/plugins/claudetto-lab/slices/capture/adapters/cli.ts. Claude Code reaches it through ANTHROPIC_BASE_URL. It records each request body and forwards it to --upstream URL, then to ANTHROPIC_BASE_URL, then to https://api.anthropic.com. It refuses an upstream that is its own address. The report splits the first request into system blocks, tool schemas, and messages.

Reason: A size figure for the start of a session must come from a real request, not an estimate. The agent sandbox can block a local port, so the user runs just capture in a terminal.

Evidence: Prompt Size.

Choice:

1. guard denies a Bash command that only prints a file, and names the Read call with the same range. A Read of an unchanged range returns a short stub, and a Bash print does not.
2. The tools section tells the model to read only the part it needs, with offset and limit.
3. bashOutputMaxChars is 16000. Claude Code 2.1.303 to 2.1.305 allow 4000 to 128000, with a default of 30000. A longer output goes to a file, and Claude gets a preview and the path.

Reason: Each read and command output adds text to the context. The value 16000 is a choice. It brings 196 of 240 saved outputs inline.

Evidence: tests/claudetto/guard/print.test.ts and slices/guard/lab/print.test.ts.

Choice: A rule that a program can check goes to a hook, a lint rule in just check, or an eval case.

• sembr rewrites line breaks, and guard denies a file print, with no model token.
• just eval runs claude plugin eval with the cases in the evals/ folders.
• just validate fails on a case that checks only the words of a reply (tools/case-checks.ts).

Reason: A rule written only as prose is not checked, and some fail in practice. No gate checks repeat reads, recaps, or multi-sentence chat lines yet.

Evidence: tests/tools/case-checks.test.ts.

• Prompt
• Styles
• Settings Snippet
• Sembr
• Sandbox
• Decisions