Claude Code's skillListingBudgetFraction: The Undocumented Setting Silently Killing Half Your Skills

I run 27 skills in Claude Code. Custom skills I have built for WOWHOW, ecosystem plugins from wshobson, the graphify integration, the Supabase agent, the Cloudflare MCP bridge — 27 total. For about three weeks, I had a subtle problem: certain skills would trigger reliably on Monday and then go silent by Thursday of the same week. Not broken. Not misconfigured. Just… absent. The skill keyword would appear in my message, the trigger condition was clearly met, and the model would proceed as if the skill did not exist.

I blamed my trigger definitions. I rewrote them. I tightened the keywords. I rebuilt two skills from scratch. The problem persisted. It was not the trigger definitions. It was a setting I had never heard of, buried in Claude Code’s internal config, that was quietly dropping skills from the context on every single turn — and the model never saw them because they were never listed.

The setting is called skillListingBudgetFraction. Here is everything I found.

The Symptom in Detail

The failure mode has a specific fingerprint. If you are experiencing it, you will recognize these characteristics:

• Skills fail to trigger intermittently, not consistently. The same skill works in a fresh session and fails in a long session.
• There is no error. No warning in the Claude Code output. The model does not say “I could not load skill X.” It simply behaves as if the skill were not installed.
• The failure rate increases as conversations get longer. Short sessions are fine. Sessions past the 20-30 turn mark start losing skills.
• If you open a brand-new Claude Code session and try the same skill, it works immediately.
• The skills that fail are not always the same ones. Different sessions drop different skills from the visible set.

That last point is the tell. If the same skill always failed, you would debug that skill. But when which skills fail varies session to session, the problem is not in any individual skill — it is in the selection mechanism that decides which skills to expose to the model on a given turn.

What skillListingBudgetFraction Actually Does

Claude Code has a skill system that works roughly as follows: installed skills have metadata — a name, a description, trigger conditions, and usage instructions. Before each model call, Claude Code assembles a skill listing: a block of text that enumerates the available skills and their descriptions. This listing is injected into the context so the model knows what tools it has available. When the model decides a skill is relevant, it signals Claude Code to load and execute that skill.

The skill listing is not free. It consumes tokens. Twenty-seven skills, each with a name, trigger description, and brief usage summary, consumes somewhere between 3,000 and 6,000 tokens depending on how verbose each skill’s metadata is. On a model with a 200,000-token context window, that is 1.5% to 3% of the window — not catastrophic in isolation.

But Claude Code does not have an unlimited token budget for skill listings. It has a fraction. skillListingBudgetFraction is the percentage of the remaining context budget that Claude Code is willing to spend on skill listings per turn. “Remaining” is the operative word: it is not a fraction of the maximum context window. It is a fraction of what is left after the system prompt, conversation history, and other injected context have consumed their share.

In a long conversation, the remaining context budget shrinks. As it shrinks, the absolute token count available for skill listings shrinks proportionally. When the available tokens fall below what it would cost to list all installed skills, Claude Code truncates the listing. Skills at the bottom of the listing order do not get included. The model never sees them. It cannot invoke what it cannot see.

This is not a bug. It is a deliberate design decision. The alternative — always including all skills regardless of context pressure — would push out conversation history or system prompt content, which are arguably more important. The problem is that the default fraction is set conservatively, and it is not documented anywhere in the Claude Code user-facing materials as of version 2.1.129.