[zealous1]

Why

In Claude Code's default view, the two plain read tools almost never got picked: 0 to 2 of 24 per run in a blind tool-choice test on 0.10.11. Only arcmira_execute_read loaded upfront, and the server instructions sent one-video reads to it. With every description visible, the plain tools took 82% to 85% of picks.

What changes

• arcmira_youtube_search and arcmira_youtube_transcript carry anthropic/alwaysLoad, beside arcmira_execute_read.
• Server instructions:
  • Routing sentence. It now reads "One video, read whole: arcmira_youtube_transcript. A quote or topic across videos: arcmira_youtube_search."
  • Name sentence. It now reads "arcmira_youtube_search takes a show, speaker or subject by name. In a program, resolve a name with arcmira.resolve when it becomes a filter; filters there take ids only."
  • Length. 1,980 of 2,000 characters.
• arcmira_youtube_search description. It adds "who sponsors a channel, or which creators recommend a product", for /mcp/apps, which has no server instructions.
• arcmira.price(video), merged in Rename the code-mode price check from quote to price #30, ships here. arcmira.quote still works as the old name.

Evidence

• Routing pilot. In the deferred view with the routing sentence changed, 3 runs each on Haiku and Sonnet, pasted-link reads went to arcmira_youtube_transcript 54 of 54 times, up from 3. Both whole-transcript misses went away.
• Pilot of both sentences. Same view, 3 runs per model. Cross-video asks went to arcmira_youtube_search 30 of 45 times for Sonnet, up from 2, and 32 of 45 for Haiku. Every Arcmira request and every control still scored right: 14 of 14, 12 of 12 and 8 of 8 in all six runs.

Checks

• pnpm test: 145 of 145 pass. Each new assertion fails on 0.10.12's source.
• tsc --noEmit, manifest:check, skill:check and the sandbox source check pass.