Collapse `configure`'s 36-property schema — it costs ~1.9k context tokens every session

[jphein]

Problem

MCP tool schemas are fixed per-session context overhead — they live in the model's system prompt for the entire session whether or not the tool is ever called. Measured via Claude Code's /context:

• configure = 1.9k tokens — the single largest tool in the server
• Whole speech-to-cli server = ~4.5k tokens
• So configure alone is ~42% of the server's footprint

The schema enumerates 36 properties, each with a full description (key, region, tts_region, player, recorder, mic_source, all the chime_*, subtitle_color_*, barge_in_*, etc.).

Why it's avoidable

handle_request already treats configure as a generic key→value setter ("Pass any setting as a key-value pair to update it"). The 36-property schema is effectively documentation stored in the most expensive place possible — re-sent every session.

Proposal

1. Replace the 36 enumerated properties with a single free-form object:

   {"settings": {"type": "object", "description": "Key→value settings to update. Call with no args to list all available settings + current values."}}

2. Make configure() with no args return the full catalog (key, type, current value, one-line doc) — discovery moves to runtime, on demand.
3. Move the per-key reference into README.md / docs/.

Impact

Reclaims ~1.7k tokens of always-on context (~90% of this tool). No behavior change — the handler already accepts arbitrary keys.

Filed from a context-footprint audit of the MCP tool surface.