feat(langfuse): add CI/CD experiment gate guidance

[wochinge]

Summary

• Add a Langfuse CI/CD reference for setting up experiment regression gates with langfuse/experiment-action.
• Cover the practical setup flow: CI platform detection, dataset shape validation, evaluator selection, threshold calibration, workflow setup, secrets, verification, and common issues.
• Link the reference from the existing langfuse skill and README, and bump both plugin manifests to 1.1.0 for the new published capability.

Linear

• LFE-9689

Review Focus

• Whether the plugin version bump from 1.0.1 to 1.1.0 matches the repo's published-skill rules.

[wochinge]

@Lotte-Verheyden: @annabellscha suggested creating a skill for the new CI/CD action so here it is :-)
Would appreciate your feedback 🙌🏻

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0afd6ff390

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[Lotte-Verheyden]

Thanks @wochinge! I reviewed this and pushed one commit (1e1dde9). Reasoning below, plus a bigger question for the remaining part:

I made the following changes (I also added these as general guidelines in the agents.md file):

• Reverted the top-level description edit, because that field only controls invocation, and a CI/CD request likely already mentions Langfuse/eval so it would trigger anyway.
• Dropped the broad github.com allowed-tools. When an action isn't on it, the agent just gets a one-time “always allow” prompt the first time. We keep the list to read-only LF-specific commands so this list does not become a point of hesitation for people who want to install the skill.
• Removed the hardcoded CLI/gh blocks + de-pinned SDK versions so that the skill doesn't go stale when we update anything to our SDK/CLI. 2 of the hardcoded commands were in fact already wrong (scores list doesn’t exist; get-get-runs takes a positional, not --dataset-name).

For the part that remains, it looks very long for what it needs to do, knowing that the agent is also already fetching the docs. I didn't test it completely, did you try a test flow with vs. without the skill update on your side? I wanted to trim it to information that's not on the docs page, and what was left wasn’t very long.

--> If the workflow works as well without the skill update as with, I would say we don't include this as a specific use case. If it doesn't, we trim it to the necessary lines and add it to the skill :)