[zealous1]

Why

The server instructions tell an agent to call arcmira_describe before writing a program. In Claude Code's default view, arcmira_describe waited behind tool search, and so did arcmira_feedback. Programs are the main way agents use Arcmira, so the reference and the way to report a problem should be in view from the first turn.

What changes

• anthropic/alwaysLoad moves from arcmira_youtube_search and arcmira_youtube_transcript to arcmira_describe and arcmira_feedback. arcmira_execute_read keeps it.
• The server instructions route to programs: "Call arcmira_describe once, then write one program per question for arcmira_execute_read, or arcmira_execute_write for monitor changes. One video needs no reference: return await arcmira.transcript(url, { start, end }); read a long one in start and end windows, one call each." The name sentence drops the plain search tool.
• The category and call_id descriptions in arcmira_feedback are shorter.
• Manifests and the CHANGELOG move to 0.10.14.

What Claude Code loads upfront from /mcp falls from 9,325 to 6,754 characters, server instructions included.

Evidence

A blind tool-choice test read the exact text this branch builds. Each run listed the loaded tools in full and the deferred tools by name, with two rival YouTube servers. The 34 requests included 8 controls, with 3 Haiku and 3 Sonnet runs per variant.

Variant	Misses in 6 runs
New upfront set, 0.10.13 routing	4, all Haiku: two complete-transcript requests got no tool, and two video-discovery controls went to Arcmira search. Agents also loaded the plain tools through tool search
New upfront set, program routing	1, Sonnet: a complete 3-hour transcript went to another server's transcript tool, citing the 12,000-character output cap
This branch: program routing plus the long-video clause	0. Every Arcmira request went to an Arcmira tool, and every control avoided Arcmira

Tests

npm test: 145 of 145 pass. The four updated assertions fail on the 0.10.13 source.