claude-api bundled skill (v2.1.169): aggressive auto-trigger on model name mentions causes double context compaction

Summary

The claude-api bundled skill introduced in Claude Code 2.1.169 (installed 2026-06-17) contains an aggressive TRIGGER condition that auto-invokes the skill whenever a user casually mentions Anthropic model names (Sonnet, Opus, Haiku) or asks about model performance. This caused a double context compaction and a killed session in a project where neither the user nor the assistant had any intent to use the Claude API.

Version

Claude Code 2.1.169 (bundled skills folder created: 2026-06-17 10:25 AM local time)

Steps to Reproduce

1. Open Claude Code 2.1.169 on any project.
2. In conversation, casually compare model context window usage -- e.g. "when I load the same file in Sonnet it jumps to 19-20% context, in Opus it was 4-6%."
3. Claude auto-invokes the claude-api skill.

What Happened

• The bundled claude-api skill description includes this TRIGGER text injected into every system-reminder:

> TRIGGER -- read BEFORE opening the target file; don't skip because it "looks like a one-liner" -- whenever: the prompt names Claude/Anthropic in any form (Claude, Anthropic, Opus, Sonnet, Haiku, anthropic, @anthropic-ai, claude-*, us.anthropic.*); the user asks about an LLM (pricing/model choice/limits/caching) -- never answer from memory...

• The trigger fired on a casual remark comparing Sonnet vs Opus context percentages -- a meta/conversational question with no relation to API development.
• The claude-api skill payload is enormous (full API reference, code examples across 9 language directories). Loading it pushed the session context to 100%.
• First compaction ran. The compaction summary serialized all the skill content, producing a 583 KB synthetic user message (line 41 of transcript, 583,685 chars).
• The 583 KB summary immediately re-filled the context on reload. Second compaction ran.
• User killed the session.

Impact

• Session destroyed by double compaction with no user intent to invoke the skill.
• The skill was never invoked before 2.1.169 -- same conversation patterns (comparing Sonnet vs Opus performance) had zero effect on prior versions.
• Workaround applied: local stub skill at ~/.claude/skills/claude-api/SKILL.md + explicit CLAUDE.md rule. The bundled trigger text still appears in the system-reminder because it is baked into the harness, not a file the user can edit.

Expected Behavior

Bundled skills with auto-trigger conditions should not fire on casual mentions of model names in conversational/meta context. The trigger scope ("any mention of Sonnet, Opus, Haiku, Claude") is too broad and does not distinguish between "user is building a Claude API application" and "user is discussing model behavior."

At minimum, the trigger should require an active coding task involving the Claude API -- not a meta-discussion about the tool itself.

Suggested Fix

• Scope the TRIGGER condition to require an open file or explicit coding context involving the Anthropic SDK.
• Or provide a user-level setting to disable specific bundled skill triggers without requiring file hacks.
• The current workaround (local stub + CLAUDE.md rule) is fragile because the bundled trigger text still appears in the system-reminder every session.