[BUG] Skills prompt injection silently truncates most skills with no prioritization

[KalebCole]

Problem

The SDK truncates the skills list injected into the system prompt using a fixed character budget (Gyt() in minified source). When a user has more skills than fit in the budget, the remaining skills are silently dropped with only an HTML comment noting the truncation:

<!-- Showing 25 of 201 skills due to token limits -->

This means:

• 176 of 201 installed skills are invisible to the model
• The skill tool fails at runtime even when the user explicitly names a truncated skill by name
• The user receives no notification that skills were dropped

Root Cause

Q3r in @github/copilot/sdk/index.js (~line 1362):

function Q3r(t) {
  let budget = Gyt();          // fixed character budget
  let rendered = [], charCount = 0;
  for (let skill of t) {
    let xml = w3r(skill);
    if (charCount + xml.length + 1 > budget) break;  // hard stop
    charCount += xml.length + 1;
    rendered.push(xml);
  }
  // truncation comment appended if skills were dropped
}

Issues with this approach:

1. No prioritization — skills are iterated in arrival order (appears alphabetical). No ranking by usage frequency, recency, or relevance to the current conversation.
2. Character-based budget, not token-based — xml.length is a rough proxy.
3. Not model-aware — same budget whether the model has 200K or 1M context. On a 1M context model, all 201 skills would consume ~1.6% of available context.
4. Silent failure — the skill tool still appears available but invocations fail for truncated skills with a confusing "Skill not found" error.

Token Math

	Chars	Est. Tokens
Current 25 skills (visible)	~13,000	~3,300
All 201 skills	~65,000	~16,400
Hidden skills	~52,000	~13,000

• % of 200K context: 8.2%
• % of 1M context: 1.6%

Expected Behavior

• All installed skills should be discoverable by the model, or at minimum prioritized intelligently
• The skill tool should be able to invoke any installed skill regardless of prompt truncation
• Users should be notified if skills are being dropped

Suggested Improvements

• Decouple skill routing from prompt injection — the skill tool should resolve skill names against the full installed list, not just what's in the prompt
• Model-aware budgets — scale the budget based on the model's context window
• Prioritization — rank by usage frequency, recency, or semantic relevance
• Lazy loading — inject only skill names/triggers compactly, load full definitions on match
• User notification — surface which skills are hidden, e.g. in copilot skills list

Environment

• Copilot CLI version: 0.0.412
• OS: Windows 11
• Model: Claude Opus 4.6 (1M context)
• Skills installed: 201

[yzeng58]

Adding my experience — I have 41 skills installed, only 31 are shown. The message says "due to token limits" but that is misleading. I spent a whole day removing MCP servers (freed ~75K chars from the system prompt) trying to make room for more skills. The number did not change at all. The skill budget is clearly a hardcoded constant that has nothing to do with actual token usage.

This is frustrating because:

1. Skills are silently dropped with zero notification. I had no idea some of my skills were invisible until I tried to invoke one and it failed. There should at least be a warning listing which skills got cut.

2. The "token limits" message is wrong. It is not a token limit — it is a fixed character budget. Removing other things from the prompt does not help. Do not tell users it is a token issue when it is not.

3. The budget makes no sense. Skill descriptions are ~200-700 chars each (body is loaded on-demand), yet they hit this cap easily. Meanwhile MCP tool schemas dump 500-4000 chars per tool into the prompt permanently with no cap at all. Why are skills penalized but MCP tools are not?

Please fix this. At minimum: (a) make the budget scale with the model context window, (b) always include all skill names even if descriptions are trimmed, and (c) stop silently dropping skills.