[guidovizoso]

What this changes

Asked to "search for" something in a demo, a Bot drove its browser to a search engine instead of calling the search tool it held. The prompt invited it: the computer guidance said to navigate whenever asked to "look at, check or read" anything and to "never claim you cannot browse", and the one sentence saying a page is not a search read as a hedge beside them. A model given both picks the emphatic one.

Now a Bot is told one thing, chosen by what it holds:

• With a search tool, finding information goes to it, and the computer is for what only a browser can do: signing in, clicking through a site, filling in a form, or opening a page the person asked to see.
• Without one, the browser never goes to Google, Bing or DuckDuckGo to look something up. The Bot answers from what it knows, says it has not checked it, and calls show_app_access once with one of Parallel Search, Exa Search or Firecrawl, so the person gets a card to ask an administrator for one. The wording is "if none of your tools searches the web", not "you hold no search tool", because a search tool can arrive unrecognised (a Composio toolkit, a server added by URL) and the Bot should not be told something false about it.
• "Never claim you cannot browse" is gone. Opening a page the person names is still something the Bot does.
• computer_navigate's description now says it opens a particular site and is not a web search, so the tool list agrees with the guidance beside it.

How each kind of Bot gets the right wording. COMPUTER_GUIDANCE (a constant) becomes computerGuidance({ search }). Built-in Bots are worded in builtInAgentConfiguration from the tools they are built with (holdsWebSearch, from grant refs), so a narrowed run follows what it was offered. agent-bot and agent-langgraph work it out from the tool names each run offers them (researchRoleOfToolName, from mcp__server__tool), which are the only view they have of their grants. RESEARCH_APP_NAMES lists the catalogue's search apps by the title show_app_access resolves; a test fails if one stops matching the catalogue.

Where it runs

• New state that outlives a request? None. The guidance is built per run from the tools that run is offered.
• What happens on the second replica? The same wording: it is a pure function of the run's tools, which every process resolves from the same grants.
• Anything serialised? No.
• Anything fanned out to a browser? No.
• New listener, port, or schedule? No.

Boundary and audit

• Every acting call still goes through the gateway. Only prompt text and a tool description change; a Bot that does open a page still goes through the computer gateway as before.
• New refusals and new failures each write a row. None are added.
• Nothing new is trusted from the client that the server can resolve itself. Built-in Bots are worded from the server's own grant refs. agent-bot and agent-langgraph read the tool names in their run input, which the server builds from the grants; a wrong guess there changes only which sentence the Bot reads, not what it may call.

Changelog

• "A Bot searches with its search tool, never its browser" under Unreleased.

Merge notes

This removes the COMPUTER_GUIDANCE export in favour of computerGuidance({ search }). The calm-computer-start PR adds a test in shared/bot-prompt.test.ts that references COMPUTER_GUIDANCE (and a line to the guidance's list of acting rules, which here lives in ACTING_LINES). Whichever of the two merges second has to switch that test to computerGuidance, checking both { search: true } and { search: false }.

How it was checked

• Tests: shared/bot-prompt.test.ts (each variant's wording, the unchecked answer and the search-app offer), server/tests/copilot.test.ts (a built-in Bot is worded from the tools it is built with, and a deployment with no computer is told about neither), server/tests/web-research.test.ts (any provider's search tool counts, by ref or by the name a model sees; the offered names are the catalogue's titles), and agent-bot/tests/history.test.ts and agent-langgraph/tests/history.test.ts (the system message follows the run's tools).
• In a running app on a combined branch of this and five sibling PRs (local stack, 2026-10-26), with no search app on the Bot: asked it to search for something. It did not open its computer, and drew a Parallel Search card. As an admin, "Turn on and give it to General Assistant" (from the connect-card-resume PR) turned the card ready, "Parallel Search is ready now. Please carry on." was sent, and the Bot searched the web and cited its sources.
• The full server suite on the combined branch: 5,821 pass. The only failures were the assertion-shape guard, since fixed on each branch, and one Mastra test that needs a package not installed locally. The app suite: 1,675 pass.

[guidovizoso]

Landed in #847.