Feature: defer MCP tool schemas (tool search) instead of sending all 420 of them on every request

[Audrey-cn]

What happens today

DSH Desktop 0.2.1-rc.2 sends every MCP tool definition on every request. It is not filtered, summarised, or deferred — the tool array in request/header is byte-for-byte the union of all connected servers' tools/list results.

Measured on a real session with rea + cua-driver + github MCP servers attached:

source	tools	bytes	≈tokens
mcp:rea	207	774,808	≈204k
mcp:cua-driver	60	127,451	≈34k
built-in	110	84,652	≈22k
mcp:github	43	47,413	≈12k
total	420	1,034,745	≈272k

One single tool (rea.compare_web_captures) is 166 KB ≈ 44k tokens by itself.

Why it matters

1. Hard failure on smaller windows. On any model with < ~300k context the payload alone exceeds the window. The downstream host then tries to compact the conversation (it has no way to drop tools), which cannot help, and it loops: a fresh part every 15–20 s. We hit 164→175 parts before switching models; diagnostics repeated handoff note ... failed past the hard line 162 times in 12 minutes. Reported from the host side in Unbounded part-switching on deterministic model failure (missing/invalid credential) ends in V8 OOM Fei2411/DS-Bot#485 (unbounded part-switching), but the root cause is here.
2. Cost and accuracy on large windows. Even at 1M context it is 27% of every request, and tool-selection accuracy is known to degrade past ~30–50 tools. Claude Code's docs put 50 tools at 10–20k tokens and call out the accuracy cliff.

Precedent

• Claude Code: tool search is on by default — definitions are withheld, the model gets a summary and searches, ≤5 tools loaded per search, up to 10k tools in the catalog. ENABLE_TOOL_SEARCH=auto[.N] activates once deferrable definitions reach 10% of the window.
• Codex: MCP tools are deferred behind tool_search as well.

Notably, the vendored provider SDK already inside DSH carries the plumbing (defer_loading, tool_search_output, the supportsToolSearch compat flag — the latter currently only used for OpenAI Responses' additive tool updates). The agent loop never withholds schemas, and there is no config switch to opt in.

Ask

1. Deferred tool loading with an automatic threshold (Claude Code's auto semantics would already cover this case: 272k / 1M = 27% ≫ 10%).
2. A per-session or per-Bot MCP scope — e.g. attach rea only to a "reverse-engineering" preset instead of every session. Today MCP servers are profile-global, so a heavy server taxes every Bot.
3. A visible tool budget guard — surface "tools ≈X tokens / window Y" and refuse to start a turn when tools alone exceed a ratio of the window, instead of silently entering compaction loops.

Option 2 alone would have prevented our incident; option 1 is the general fix.