Problem

When the number of eligible skills exceeds the prompt token budget, OpenClaw truncates the skills list injected into the system prompt. Currently this happens silently from the users perspective:

• A note ("Skills truncated: included 68 of 90") appears in the system prompt, but the agent sees it, not the user. In practice, even the agent ignores it. With skills spread across ~/.agents/skills/ and ~/.openclaw/skills/, there is no easy way to tell which skills got cut without running openclaw skills check and cross-referencing the agents available_skills list.
• User-created skills can be dropped entirely. I had a custom skill in ~/.openclaw/skills/ that was eligible (openclaw skills check confirmed it), but it never appeared in the agents available_skills list. I had to manually inspect the filesystem to confirm the skill existed and was properly formatted.
• The agent behaved as if the skill did not exist. It confidently told me no such skill was found, because from its perspective, it was not there.
• There is no documentation about the skills list having a token limit or what happens when it is exceeded.

This is particularly bad for skills containing critical operational rules. If the agent cannot see the skill, it cannot follow the rules.

Expected behavior

• User-created skills should not be silently dropped from the prompt without the user knowing.
• Users should have a way to ensure specific skills are always loaded.

Proposed solutions (any/all)

1. Skill pinning / priority - Allow skills.entries.<name>.priority: "high" or skills.entries.<name>.pinned: true in config so critical skills survive truncation.

2. Per-topic required skills - Config like channels.telegram.groups.<id>.topics.<id>.skills: ["maitre-d"] that force-loads specific skills for a given conversation context, regardless of the global truncation budget.

3. Context-aware truncation - Instead of a static cutoff, prioritize skills that match the active channel/topic/conversation context before truncating less relevant ones.

4. Surface truncation to the user - If skills are dropped, warn the user (e.g., a one-time message or log entry listing which skills were excluded), not just the agent.

5. Document the limit - The Skills docs describe token impact per-skill but do not mention that the list is truncated at a budget ceiling, or what that ceiling is. This should be documented so users can plan accordingly.

Environment

• OpenClaw 2026.2.22-2
• 90 eligible skills, 68 included in prompt
• Skills spread across multiple valid locations (~/.agents/skills/, ~/.openclaw/skills/, <workspace>/skills/) - all confirmed eligible via openclaw skills check, but only a subset visible to the agent at runtime due to truncation