skillListingBudgetFraction is calculated against a fixed ~200K baseline, not the model's actual context window

Description

The skillListingBudgetFraction setting is documented in the schema as:

"Fraction of the context window (in characters) reserved for the skill listing sent to Claude (default: 0.01 = 1%)"

In practice the calculation appears to use a fixed ~200K-token reference rather than the model's actual context window. On a 1M-context model, this means the effective budget is ~5× smaller than expected.

Repro

• Model: Opus 4.7 with 1M context (claude-opus-4-7[1m])
• skillListingBudgetFraction: default (0.01)
• ~54 skills installed (~3k tokens of descriptions)
• /doctor reports: Skill listing will be truncated — 21 descriptions dropped (1.7%/1% of context)

Math

If the budget used the actual 1M window, 3k tokens would be ~0.3% of context and well under the 1% cap — no truncation should occur. The reported "1.7% of context" only makes sense if "context" = ~200K tokens (3k / 0.017 ≈ 176K).

Expected

Either:

1. The budget scales with the model's actual context window (so 1M users get 5× the room they do today), or
2. The schema description is clarified to say "fixed 200K reference" rather than "the context window", so users know to raise the fraction explicitly on extended-context models.

Version

Claude Code 2.1.138 (native, win32-x64, commit d6d494651eb4)