code-testing-agent: mention find-untested-sources for C# discovery

[Evangelink]

Summary

Adds a conditional pointer to the find-untested-sources skill in two places inside code-testing-agent:

• SKILL.md Step 3 (Research Phase) — high-level note for C# / .NET multi-file scopes: prefer the helper over manual find/grep/glob walks.
• code-testing-researcher.agent.md Section 7 (Discover Preexisting Tests) — directive instruction telling the researcher to invoke the helper before manually pairing source ↔ test files, and to use its source_to_tests / untested output to fill the research document.

Both callouts are gated on "when available in the workspace" — installations without find-untested-sources continue working via manual discovery. Adds no behavior change for non-C# repos.

Diff

+4 lines, -0 lines across 2 files. No structural changes, no removed content.

Measurement — honest restatement

I originally claimed "−15.67 % input tokens at neutral pass rate" from a 5×136-instance internal msbench experiment. After re-doing the analysis correctly (per-task mean rather than volume-weighted aggregate, with a non-.NET control bucket), the picture is more modest:

Bucket	n tasks	Per-task mean Δ input tokens	Per-task median Δ
.NET tasks	35	−8.73 %	−6.98 %
non-.NET tasks (control)	101	−0.13 %	−1.43 %

Differential (.NET minus non-.NET): −8.60 pp, Welch's t ≈ −2.02 (right at p ≈ 0.05). Per-task std-dev is ~21 % in both buckets — 5 runs per task is not enough to nail down per-task effects precisely, but the across-task pattern is clean: non-.NET tasks are unchanged (which is the predicted no-effect under my hypothesis), .NET tasks shift modestly negative.

Pass rate was neutral (within noise) on both buckets.

Important caveats

• The volume-weighted −15 % total was dominated by a handful of very-high-token .NET tasks (e.g. ocelot-core-gen-detailed went 19M → 11.6M); the per-task picture is the right unit of analysis.
• The helper itself was invoked 0/679 times. The Copilot CLI router did not auto-load it as a sibling skill, so any savings come from the model reading the documentation text in the loaded code-testing-agent SKILL.md, not from the helper executing. The runtime value of the helper is still unmeasured.
• Several .NET tasks regressed (aspnetcore +44 %, maui +35 %, complog-gen-detailed +13 %), consistent with "the model trusts the heuristic and sometimes skips exploration it shouldn't have skipped."

The honest claim is therefore: a modest, marginally significant reduction on .NET tasks, isolated to .NET (control bucket flat), at the cost of higher variance.

Dependency

Depends on #733 (which adds the find-untested-sources skill itself). Safe to merge before #733 lands — the "when available" gating means the pointer is a no-op until the helper is installed.

Test plan

Doc-only change. No code paths affected. Verified diff is the two intended additions only.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[github-actions]

Skill Coverage Report

	Plugin	Skill	Covered	Coverage
✅	dotnet-test	code-testing-agent	4/5	80%

Uncovered: dotnet-test/code-testing-agent

• [WorkflowStep] Step 2: Invoke the Test Generator (line 83)

[Copilot]

Pull request overview

Updates code-testing-agent documentation to recommend the find-untested-sources helper (when installed) for C#/.NET multi-file scopes, so the researcher can build source↔test pairing more efficiently without manual repo-wide discovery walks.

Changes:

• Add a C#/.NET-specific note in SKILL.md Step 3 (Research Phase) pointing researchers to find-untested-sources when available.
• Add a directive in code-testing-researcher.agent.md Step 7 (Discover Preexisting Tests) to invoke find-untested-sources first and use its untested / source_to_tests output to populate the research doc.

Show a summary per file

File	Description
plugins/dotnet-test/skills/code-testing-agent/SKILL.md	Adds a gated C#/.NET pointer to prefer find-untested-sources for pairing over manual find/grep/glob.
plugins/dotnet-test/agents/code-testing-researcher.agent.md	Instructs the researcher to invoke find-untested-sources (when available) before manual source↔test pairing and to use its JSON outputs in research.

Copilot's findings

Tip

Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

• Files reviewed: 2/2 changed files
• Comments generated: 0