projects/agent-native/slices/safety-mcp-audit/briefs/d2-r1.md
Pre-staged during S3-D1's run. The probe answered Open Question 1 "surfaces today" — no
runCommandfix; this dispatch is the description sentence plus final gates.
Update the migrate-reset tool description in packages/cli/src/mcp/MCP.ts to tell agents
the command is gated for AI agents and that, when blocked, the returned error text contains
the consent protocol to relay verbatim to the user (note the text arrives after execa's
"Command failed with exit code 1: <command>" prefix — word the sentence so an agent isn't
surprised by that framing). Review the other three tool descriptions for one-pass
consistency only if a trivially small wording touch keeps them coherent — no behavioral
changes anywhere.
migrate-reset description names the AI-agent gate and instructs relaying the
consent protocol from the error text; zod schema and behavior untouched.mcp-safety.vitest.ts still green (the description change must not break the pinned
contract); packages/cli pnpm run tsc green; eslint on the touched file green.migrate-reset tool listing (name + description) as an MCP client
sees it (tools/list via the SDK client), for the PR description.One commit; explicit staging; no push; heartbeats as usual.