feat(dotnet11): add process-api-net11 skill for new process APIs

[shaikhsharukh]

Resolves issue #649 by adding a new skill process-api-net11 for the new System.Diagnostics.Process APIs introduced in .NET 11.

This skill guides coding agents on using the new convenience one-liners, reliable deadlock-free output reading, and parent-child process lifetime properties. Evals have been included and verified using the skill-validator.

[github-actions]

Note

This PR is from a fork and modifies infrastructure files (eng/ or .github/).

Changes to infrastructure typically need to be submitted from a branch in dotnet/skills (not a fork) so that CI workflows run with the correct permissions and secrets.

Please consider recreating this PR from an upstream branch. If you don't have push access to dotnet/skills, ask a maintainer to push your branch for you.

[shaikhsharukh]

@dotnet-policy-service agree

[rickgonzalez]

A short narrated explainer of this PR (dev testing my tool of public repos) - I hope you find it helpful. https://www.lenzon.ai/viewer/cmrb1zapc000o14os7auramuh?voice=google-chirp3

[AbhitejJohn]

🤖 Posted by GitHub Copilot on behalf of the repo maintainers.

Thanks for this contribution, @shaikhsharukh! 🎉 I verified every API in this skill against dotnet/runtime main (which is the .NET 11 branch — eng/Versions.props shows ProductVersion 11.0.0; no release/11.0 branch exists yet), and the good news is these are all genuinely shipping .NET 11 APIs — Process.Run/RunAsync, RunAndCaptureText/Async, StartAndForget, the ReadAll* family, and the new ProcessStartInfo properties. Nice job picking a real, useful new surface. 👍

Before we run evals and move toward merge, though, the SKILL.md documents several of these APIs inaccurately versus the public ref assembly. Since a skill trains agents to emit code, these would lead to non-compiling output. Could you correct the following?

1. ReadAll* return types are wrong. They return tuples / streaming types, not plain strings/arrays:

Documented	Actual
string ReadAllText()	(string StandardOutput, string StandardError) ReadAllText(TimeSpan? timeout = default)
byte[] ReadAllBytes()	(byte[] StandardOutput, byte[] StandardError) ReadAllBytes(...)
string[] ReadAllLines()	IEnumerable<ProcessOutputLine> ReadAllLines(...)
Task<string[]> ReadAllLinesAsync()	IAsyncEnumerable<ProcessOutputLine> ReadAllLinesAsync(...)
Task<string> ReadAllTextAsync()	Task<(string, string)> ReadAllTextAsync(...)

2. InheritedHandles — invented type. The doc uses ProcessHandleInheritance InheritedHandles, but that type doesn't exist. The real property is IList<SafeHandle>? InheritedHandles.

3. Static method signatures. Run, RunAndCaptureText, and StartAndForget don't take (string arguments = "", ProcessStartInfo? startInfo = null). Actual: arguments is IList<string>? with an optional TimeSpan? timeout (Run also has a bool silent parameter); StartAndForget has no startInfo trailing parameter.

4. Example 1 won't compile. Process.RunAndCaptureTextAsync("git", "--version") — a bare string can't bind to IList<string>?. Use ["--version"] (or pass a ProcessStartInfo).

5. Example 3 won't compile. string[] lines = await process.ReadAllLinesAsync(); foreach (var line in lines) — ReadAllLinesAsync() returns IAsyncEnumerable<ProcessOutputLine>, so it needs await foreach, and each item is a ProcessOutputLine, not a string.

For reference, Example 2 (KillOnParentExit = true) and the ProcessTextOutput/ProcessExitStatus usage in Example 1 are correct as written. ✅

Once the signatures and examples are aligned with the ref assembly, we'll kick off evals and continue the review. Thanks again for contributing! 🙏

[shaikhsharukh]

I have aligned the skill definitions and examples with the dotnet/runtime ref assembly:

1. ReadAll* Return Types: Corrected the signatures to match the exact tuple return types ((string StandardOutput, string StandardError) for ReadAllText/Async, (byte[] StandardOutput, byte[] StandardError) for ReadAllBytes/Async) and streaming types (IEnumerable<ProcessOutputLine> / IAsyncEnumerable<ProcessOutputLine> for ReadAllLines/Async).
2. InheritedHandles Property: Changed the property type from the hypothetical ProcessHandleInheritance to IList<SafeHandle>?.
3. Static Method Signatures: Aligned the static Run, RunAndCaptureText, and StartAndForget methods to accept IList<string>? arguments and appropriate timeout/silent parameters.
4. Example 1 Compilation: Modified the call to pass ["status"] as a collection expression instead of a bare string parameter.
5. Example 3 Compilation: Updated the ping example to use await foreach over IAsyncEnumerable<ProcessOutputLine> and access line.Content / line.StandardError.

[AbhitejJohn]

Follow-up review — API accuracy re-verified + eval-hardening suggestions

Thanks for the fixes, @shaikhsharukh. I re-checked everything and ran a deeper evaluation pass. Summary below; concrete suggestions are inline on eval.yaml.

1. API accuracy — all confirmed ✅ (proven)
I re-verified every signature in SKILL.md against the current dotnet/runtime main reference assemblies. All five items from the earlier review are now correct: the tuple-returning ReadAllText/ReadAllTextAsync → (string StandardOutput, string StandardError), streaming ReadAllLinesAsync → IAsyncEnumerable<ProcessOutputLine>, RunAndCaptureText(Async) → ProcessTextOutput, and KillOnParentExit. Content looks accurate.

2. Does the skill actually help? — yes, in plugin/deployment mode (proven, scoped)
I hardened the eval to use intent-only prompts (never naming the API) and ran it with skill-validator (--runs 5, judge = claude-opus-4.6). In the whole-plugin configuration (how the skill actually ships), it produced large, low run-to-run variance (CV 1–6%) gains on its target APIs vs a no-skill baseline:

Scenario	Baseline	With skill
Run + capture (RunAndCaptureText)	1.8 / 5	4.2 / 5
Full read (ReadAllText tuple)	1.0 / 5	4.2 / 5
Stream lines (ReadAllLinesAsync)	1.4 / 5	4.2 / 5

The no-skill baseline reached for the pre-.NET-11 pattern (StandardOutput.ReadToEndAsync() + Task.WhenAll); the judge noted it "didn't even attempt to research what new APIs .NET 11 offers." So the skill is teaching genuinely non-obvious APIs — that's its value.

3. Honest caveat — the blended eval verdict is not green
Being transparent: the validator's official pass metric doesn't clear the 0.10 bar (this skill scored 0.065). The plugin-mode wins above are real, but the isolated single-skill arm shows only small lift and the skill roughly doubles output tokens — both of which the weighted score penalizes. So it's a strong deployment-mode signal with a failing blended score; flagging for maintainers to weigh rather than claiming a clean pass.

4. Suggested eval improvements (inline suggestions)

• Strengthen the run+capture assertions to require both .StandardOutput and .StandardError and reject the old event model.
• Add two scenarios that hit the harder, wrong-signature-prone surface — the tuple-returning ReadAllText and streaming ReadAllLinesAsync / await foreach. These are where the skill adds the most measurable value.
• Add a subtle net10.0 non-activation negative (the new APIs are net11-only).
• The existing KillOnParentExit scenario is fine as coverage, but note it shows ~0 marginal lift — the base model already emits it correctly — so it isn't what demonstrates the skill's value.

---

cc @lewing @SamMonoRT — would appreciate your read on point 3: the deployment-mode gains are real and stable, but the blended eval score doesn't clear the gate. Is the plugin-mode evidence enough to accept, or do you want the eval reworked to a clean green first?

[shaikhsharukh]

I have applied all of the suggested evaluation improvements:

1. Hardened run+capture assertions: Updated to verify that both .StandardOutput and .StandardError are accessed, and blocked usage of the legacy event-based BeginOutputReadLine / OutputDataReceived patterns.
2. Added ReadAllText (tuple return) scenario: Created a new scenario verifying that agents properly destructure the (StandardOutput, StandardError) tuple when reading process output via instance methods.
3. Added ReadAllLinesAsync scenario: Added a streaming scenario verifying that agents consume IAsyncEnumerable<ProcessOutputLine> using await foreach and read .Content / .StandardError properly.
4. Broadened non-activation guards: Updated the .NET 8 negative scenario to block leakages of any of the new .NET 11 APIs (RunAndCaptureText, ReadAllTextAsync, ReadAllLinesAsync, KillOnParentExit).
5. Added .NET 10 negative scenario: Added a subtle negative scenario verifying that agents correctly fallback to pre-.NET 11 redirect and stream-reader APIs when targeting net10.0.

[AbhitejJohn]

/evaluate

[github-actions]

❌ Evaluation did not complete successfully (the evaluate job reported failure). Check the workflow run logs, then re-post /evaluate to try again.

[AbhitejJohn]

@eiriktsarpalis could you please review this PR? You did the deepest review on the closely analogous #535 (new dotnet11 plugin skill + evals), so your feedback here would be much appreciated.

[eiriktsarpalis]

It might be best if @adamsitnik took a look.