[Github Agent] add prFiles and prFailedChecks read actions

[George Ng (GeorgeNgMsft)]

When a pull request is failing, you usually need two things before you can say anything useful about it: what the PR actually changed, and what the red checks are complaining about. Today the github-cli agent can tell you a PR exists and that its checks are red, but not much more, so answering either question means leaving the agent and running gh by hand.

This adds two read-only actions that answer those two questions directly.

prFiles lists the files a pull request changes, with each file's status and line counts, and optionally a short excerpt of each file's diff.

prFailedChecks explains why the checks are red. For each failing check it reports the workflow, timing, and a link, plus the error annotations GitHub recorded against it — usually the actual compiler or test error, with its file and line.

Highlights:

• Deterministic fast actions. Both are typed, read-only, and handled directly: they run gh, parse its JSON, and return a stable rawData payload. No reasoning step is involved, so an external MCP client can discover and call them on its own.
• Bounded output, honestly labelled. File counts, patch sizes, check counts, and annotation counts all have caps, and every result says explicitly what got truncated instead of quietly dropping it.
• Diffs are opt-in. When patches aren't requested, they're stripped server-side via --jq, so listing the files of a large PR doesn't download its whole diff.
• Works on other repos, including by link. Omit repo to use the current checkout, or pass an OWNER/REPO slug — or just the pull request's web link, which gh --repo itself rejects. The host is carried through, so GitHub Enterprise links work too.
• Degrades instead of failing. Missing logs, in-progress checks, checks from non-GitHub providers, forks, and permission errors each produce a clear partial result and a hint, rather than an error.
• Natural language too. Grammar rules are added so phrases like "what files did PR 123 change" and "why is PR 123 failing" work, but they're additive — the actions stand on their own.

Tests cover argument construction, JSON parsing and structured results, repo/link normalization, pagination and truncation boundaries, grammar, and the error paths.

[robgruen]

If you're in here adding fun features I would ask you to add "review PR for me" and "investigate PR failure" and variations thereof. (although the latter requires copilot cli).

[George Ng (GeorgeNgMsft)]

If you're in here adding fun features I would ask you to add "review PR for me" and "investigate PR failure" and variations thereof. (although the latter requires copilot cli).

I can add this in a follow up PR later on, but right now I'm generally covering common Git related actions that are deterministic with no model dependency. Trying to make sure our list of useful fast actions is well curated for the average dev scenario.