session.skills_loaded event: when is it intended to fire?

[K8Rayner]

Environment

• Copilot CLI: 1.0.36-0
• Python SDK: github_copilot_sdk==0.3.0
• OS: Windows 11

Question

The SDK's SessionEventType enum defines SESSION_SKILLS_LOADED = "session.skills_loaded" with payload SessionSkillsLoadedData(skills: list[SkillsLoadedSkill]), matching the shape returned by session.skills.list(). The schema has no docstring on the firing condition.

In our environment the event never fires. We verified this across many sessions and by grep -r '"type":\s*"session.skills_loaded"' ~/.copilot/session-state/**/events.jsonl (zero hits).

When is this event intended to fire?

• On session init, immediately after skills.list() is internally populated?
• After every successful session.skills.enable(...) / disable(...) / reload()?
• Only when the CLI's discovery pipeline re-scans skill_directories?
• Some combination of the above?

Why it matters

We're a host integration that needs to keep a UI "active skills" indicator in sync with the SDK's view of the world. We've worked around the absence of this event by combining:

1. SKILL_INVOKED (fires only on explicit /skill invocation, well-documented)
2. Inferring skill activation from EXTERNAL_TOOL_REQUESTED via our own tool→domain map

…but a confirmed authoritative event would let us replace the inference path with a pull from skills.list() triggered by session.skills_loaded.

If the event is intended to fire and isn't, we're happy to file a follow-up bug with this discussion linked. If it's by design (e.g. only fires in a CLI mode we haven't enabled), a docstring or docs note clarifying that would be very helpful.

Thanks!

[K8Rayner]

Additional finding: SKILL.md body size also silently prevents activation

We found that skill.invoked never fires for a skill whose SKILL.md body is 8,450 estimated tokens (57 KB), even with 15+ keyword overlaps in the prompt. After splitting the same content into two skills (~3,500 and ~5,270 tokens), both activate correctly.

The agentskills.io spec says "< 5000 tokens recommended" under Progressive Disclosure but doesn't mention enforcement. The exact cutoff is somewhere between 5,270 and 8,450 tokens — we couldn't narrow it further.

This is the same silent-rejection pattern as #3494 (description > 1024 chars silently dropped). Could the body-size threshold be documented, and ideally logged as a warning when a skill exceeds it?