[httpsphl]

What changes for the owner

When a bot asks to run a command, the card in the chat used to show the raw command (a long cd ... && python - <<'EOF' ... line) with "Full input" folded under it. Someone who does not read commands allowed it without knowing what it was.

Now the card leads with what the bot says the command is for, in plain words, and keeps the command itself behind an arrow:

• Scout wants to run a command
• Opens every link in the report draft and lists the ones that no longer work
• Scout wrote this. What actually runs is the command below.
• ▸ See the command (the whole command, with its line breaks)

When the bot gave no explanation, the card says so ("Scout did not say what this command is for. If you are not sure, deny it and ask.") and the command is already open.

The tool line above the card follows the same idea: "Run a command" and the bot's words; a click opens the command and its output. The conversation list shows the same text.

Where the explanation comes from

Claude Code's Bash and PowerShell tools take an optional description next to command, written by the model (documented in the hooks reference). It reaches our permission_prompt tool unchanged. The daemon passes it along as explanation on the tool and approval chat items; summary is now always the command itself.

The generated rules (.claude/rules/botloft.md) now ask every bot to describe each command for a non-technical owner, in the language the owner writes in. Existing bots get the rule on their next start.

Security

The explanation is written by the bot, so it never stands in for the command:

• it is shown as plain text (no markdown, no links), with a line saying the bot wrote it and that what runs is the command;
• the command is one click away and whole: a command's input is now kept up to 32 KB (it was 4 KB, and real commands were being cut), and a longer one says it is cut;
• Claude Code gets back the original input, as before.

What was checked with the real Claude Code (2.1.284)

Recorded in spec section 19.

• Docs: description is optional in both shell tools.
• A probe permission tool that logs its input and denies, on Haiku 4.5, Sonnet 5.5 and the plan default (Opus 5.5): all 32 requests carried description, the same as the tool_use block.
• Language without the rule: short and technical; with the owner writing in Portuguese, Haiku and Opus described in Portuguese, Sonnet in English.
• With the rule: all three wrote plain Portuguese with the reason ("Instala o pacote requests, que ajuda o Python a acessar sites").
• Through the dev daemon and the app preview (?live), a Sonnet bot in Manual mode: the card showed the explanation, "Ver o comando" opened the whole heredoc script, Allow ran that command, and the debug log had none of the text.
• Not guaranteed: in real use, two Sonnet bots made 28 commands with no description at all, while the other bots always sent one. I could not reproduce that in a test, so whether the new rule fixes it is not proven; the card covers the case.

Manual test

1. cargo run -p botloftd -- serve with a dev BOTLOFT_HOME, and pnpm dev with ?live=<port> (or pnpm tauri dev).
2. Create a crew, keep its chief in Manual mode, and ask in your own language for something that needs a command, for example "crie uma pasta chamada relatorios e grave nela um arquivo com os números de 1 a 5 usando um script Python".
3. The card shows the explanation in your language, then "See the command". Open it: the full command is there. Allow: the command runs.
4. After updating, restart an older bot running on Sonnet and ask for a command: check that it now explains. If it still does not, the card should say the bot did not say what the command is for, with the command open.
5. pnpm dev alone (fake data): the Revisão bot in the Research crew shows both cards.

Not in this PR

• Other tools keep their current look (the file or URL, with the input under "Details"). Claude Code's Monitor tool also takes a command and a description, but it was not checked and is shown like any unknown tool.