Skills produce no output under Claude Opus 4.8 in the Desktop App when many MCP servers are enabled (works with GPT-5.x)

### Short summary

Invoking any skill under Claude Opus 4.8 shows "Loaded skill …" and then produces no output at all — no text, no tool calls — for that turn and all following turns. The same skill and config work under GPT-5.5/5.6, and work under Opus 4.8 once most MCP servers are disabled.

### Affected version or release

1.0.25

### Installation context

Desktop App on Windows. 15 MCP servers enabled (~580 tools total; heaviest: ecs ~135, azure-devops ~90, geneva ~85, azure ~65, icm ~45). Global custom instructions ~152 KB (~26K tokens). Affects all skills/agents whenever the selected model is Claude Opus 4.8.

### What happened?

_No response_

### Steps to reproduce

_No response_

### Expected behavior

### Observed behavior
Selecting Claude Opus 4.8 and invoking a skill (e.g. /s360-sla-status) shows `Loaded skill …` and then nothing — no assistant text, no tool calls. Follow-up messages ("continue", plain text) also return no output. The turn appears to fail silently with no error.

### Steps to reproduce
1. Enable ~15 MCP servers (~580 tools total).
2. Select **Claude Opus 4.8**.
3. Invoke a skill (e.g. `/s360-sla-status`).
4. Result: `Loaded skill …`, then no response; subsequent messages also produce nothing.
5. Switch to **GPT-5.5 / GPT-5.6** → works. **Or** disable all but the 2 MCP servers the skill needs, under Opus 4.8 → the skill runs end-to-end successfully.

### Expected
Opus 4.8 executes the skill — or, if a context/tool limit is exceeded, surfaces an **explicit error** instead of empty output.

### Actual
Silent no-op: zero output across multiple turns, no error shown.

### Likely root cause
The Desktop App appears to load **all** enabled MCP tool schemas into the model request up front. ~580 tool schemas (~120–175K tokens) plus ~26K tokens of instructions likely exceed Opus 4.8's context budget (or a per-request tool-count limit), causing a silent request failure. GPT-5.x tolerates the larger payload.

**Key signal:** the **GitHub Copilot CLI does not reproduce this** — it defers MCP tools (surfaces them on demand via tool search), keeping requests small, and Opus 4.8 works there with all 15 servers enabled.

### Suggested fix
1. Add **on-demand tool loading / tool deferral** in the Desktop App (parity with the CLI).
2. When a request would exceed context/tool limits, **fail loudly** with an error rather than producing empty output.
3. Optional: show a per-model tool/context-budget warning in the MCP settings UI.

### Workaround
Enable only the MCP servers a task needs when using Opus 4.8.

### Additional context

_No response_

Start by reproducing the failure in the Desktop App with Claude Opus 4.8, about 15 MCP servers, and the reported instruction size; compare behavior with GPT-5.x and the GitHub Copilot CLI. Trace MCP tool loading and the MCP settings UI, then verify that skills produce output or that an explicit context/tool-limit error appears instead of a silent no-op.

Scritto dal modello di indicizzazione a partire dal testo della issue.