Regression: default-context Sol models rejected before first prompt after tool definitions grow from 15K to ~260K tokens #5151

Describe the bug

Beginning around 2026-10-14, gpt-5.6-sol and gpt-6.1-sol at the default context tier fail before accepting a prompt:

compaction_static_context_blocked: Static system messages and tool definitions exceed the model's usable context budget. Reduce static context or switch to a larger-context model.

Selecting long_context immediately makes the same model, session, MCP configuration, and prompt work. Usage telemetry then reports maxPromptTokens: 922000 and approximately 258K-260K tool-definition tokens.

This did not occur previously with the same broad tool configuration. A successful October 9 session on CLI 1.0.92 used gpt-5.6-sol with the default tier and reported 615 tools / 15,457 tool-definition tokens. The current configuration reports 676 tools / approximately 260,000 tool-definition tokens: roughly 10% more tools but a 17x increase in tool tokens.

The local standalone MCP configuration was unchanged during this period. A historical default-tier session that worked before the rollout also failed with the same warning when resumed on October 14.

Affected versions

• GitHub Copilot CLI 1.0.93, 1.0.97, and 1.0.98
• macOS 26.7.1, Apple Silicon
• Models: gpt-5.6-sol and gpt-6.1-sol
• Agency versions observed during the rollout window: 2026.10.13.x and 2026.10.14.3

Because 1.0.93 also began failing after the rollout window, this may involve remotely supplied tool catalogs, Agency integration, or context-budget metadata rather than only a 1.0.97 client change.

Steps to reproduce

1. Start Copilot CLI with a broad enterprise MCP/tool configuration.
2. Select gpt-6.1-sol with the default context tier.
3. Submit a short prompt.
4. Observe compaction_static_context_blocked, followed by No response was returned.
5. Switch only the context tier to long_context.
6. Resubmit the prompt; it succeeds.
7. Repeat with gpt-5.6-sol; default context fails and long_context succeeds.

Expected behavior

The default tier should continue working with a tool configuration that worked previously. At minimum, the CLI should:

• avoid offering a model/tier combination that cannot fit its own static context;
• lazy-load or defer tool schemas instead of eagerly injecting hundreds of definitions;
• identify the largest static-context contributors;
• clearly report the effective default-tier prompt budget.

Additional evidence

• October 9, CLI 1.0.92: 615 tools, 15,457 tool tokens, default-tier gpt-5.6-sol worked.
• October 14-15: approximately 258K-260K tool tokens; default-tier Sol models fail before a model request.
• long_context reports a 922K prompt allowance and succeeds without changing the tool setup.
• Related issues: HydraFusion: after a 400 on the routed model, the session is re-routed to a small-context model that cannot load the static prompt; tool set changes mid-session #5042 and Enterprise MCP allowlist tool schemas exceed runtime token limit → persistent compaction loop #3542.

Please investigate whether tool schemas or descriptions changed around October 13-14, whether more definitions are now eagerly materialized, or whether static-context token accounting/default-tier reserves changed. I can provide redacted event excerpts or confidential debug logs privately.