feat(hook): tag skills loaded by reading their SKILL.md, not only Skill tool invokes

[lilpacy]

Problem

collect_skill_tags covers two skill trails: a Skill tool_use block, and attributionSkill on assistant rows (slash commands). There is a third trail it misses: dependency sub-skills loaded by reading their SKILL.md directly with the Read tool, which is how a skill referenced from another skill's body ("REQUIRED SUB-SKILL: …", "See skills/foo/SKILL.md") typically gets loaded.

Measured on my local transcripts over 4 weeks: a commit-workflow skill that marks a shell-quoting sub-skill as required was invoked 33 times, while the sub-skill collected 0 skill: tags — it was only ever loaded via Read. A sibling Codex setup that counts raw SKILL.md reads saw the same sub-skill in 32 sessions over the same period. Tag-based usage analytics therefore undercount exactly the skills that only ever run as dependencies, and "which skills can I safely disable?" queries give a false zero for them.

Change

add_skill_tags_from_rows now recognizes the third trail: a Read tool_use whose file_path matches (^|/)skills/<name>/SKILL.md$ adds the same skill:<name> tag.

• The regex is anchored on the SKILL.md filename, so reads of skills/<name>/references/*.md or other files inside a skill directory do not tag.
• Dedup is unchanged (a turn that both invokes and reads the same skill yields one tag).
• Subagent rows flow through the same helper, so the subagent-skill: namespace picks the trail up as well.
• Still gated behind the existing CC_LANGFUSE_SKILL_TAGS option; no new configuration.

Verification

Three new unit tests in tests/unit/test_skill_tags.py:

Test	Pins
test_skill_loaded_via_read_of_skill_md_is_tagged	the new trail
test_read_of_non_skill_file_is_not_tagged	ordinary reads and references/* reads stay untagged
test_read_and_invoke_of_same_skill_dedupes	one tag per skill per turn

Full suite: 223 passed.

End to end: a headless claude -p session instructed to only Read a skill's SKILL.md (no Skill invoke) produced a trace tagged ['claude-code', 'skill:goal-first'], confirmed through the metrics API on the tags dimension against a live Langfuse project.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}