Enrich dependency alerts with Copilot compatibility assessments

[harder]

What changed

• Enrich critical-dependency issues with upstream release highlights, skill-specific notes, explicit verification status, and focused follow-up checks. Retry missing or malformed Copilot assessments on existing issues.
• Run pinned Copilot CLI with read-only tools for each new issue or manual reassessment. Validate required sections, compatibility status, and the 500-word limit before a separate least-privilege publisher updates its own marked comment.
• Harden gh skill install calls: put selectors after --, pass --pin=<ref> safely, treat --upstream as a boolean flag, and place @VERSION on a skill selector rather than a repository. Update TUI and CLI call sites.
• Add a live gh skill list --json contract check and update automation documentation.

Verification

• Node monitor and assessment tests: 9 passed.
• Real gh 2.102.0 contract tests: 7 passed; all 48 install-agent IDs match the catalog.
• Focused install/parser tests: 44 passed. Release build: zero warnings or errors.
• Live workflow dispatch for issue GitHub CLI v2.102.0 compatibility review #31 on the latest commit: monitor, Copilot assessment, and comment publishing all passed.
• Latest PR CI: Linux/macOS/Windows tests, four AOT targets, CodeQL, workflow lint, and site check passed; remaining checks in progress.
• Copilot PR review requested again for the latest commit after addressing prior findings.

Related: #31

[Copilot]

Copilot review overview

🟡 Changes recommended

Release filtering can include unrelated notes, and one failed matrix leg suppresses all successful assessments.

Review effort: Balanced
Findings: 2

Open (2)

• Avoid copying unrelated Markdown list items into skill-related release notes · New
• Publish successful assessment artifacts despite other matrix leg failures · New

What changed in this PR

Adds automated Copilot compatibility assessments to critical-dependency alerts while hardening gh skill install argument handling.

Changes:

• Adds read-only Copilot assessment and publishing jobs.
• Enriches dependency issues with release highlights and verification status.
• Places untrusted install selectors after -- and adds contract coverage.

File	Description
AGENTS.md	Documents automation and argument-safety contracts.
.github/​actionlint.yaml	Scopes the new permission lint exception.
.github/​copilot-instructions.md	Adds dependency-assessment guidance.
.github/​scripts/​critical-dependency-prompt.md	Defines the assessment prompt.
.github/​scripts/​critical-dependencies.js	Generates enriched alerts and reassessment outputs.
.github/​scripts/​critical-dependencies.test.js	Tests alerts, summaries, and reassessment validation.
.github/​workflows/​README.md	Documents the assessment workflow.
.github/​workflows/​critical-dependencies.yml	Runs and publishes Copilot assessments.
src/​SkillView.Core/​Gh/​GhSkillInstallService.cs	Adds the positional option boundary.
tests/​SkillView.Tests/​Gh/​GhCliContractTests.cs	Adds live JSON inventory coverage.
tests/​SkillView.Tests/​Gh/​GhSkillInstallServiceTests.cs	Tests safe argument ordering.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🟡 Changes recommended

Version refs are still attached to repository arguments incorrectly, and assessment bounds and upstream mention handling need enforcement.

Review effort: Balanced
Findings: 1

Open (1)

• Enforce the advertised 500-word assessment limit · New

Resolved since last review (2)

• Publish successful assessment artifacts despite other matrix leg failures
• Avoid copying unrelated Markdown list items into skill-related release notes

Previously missed (2)

In code that hasn't changed since last review

Sanitize upstream @​mentions in generated issue highlights

.github/​scripts/​critical-dependencies.js:63

Unlike releaseOverview, this copies matching upstream text into an issue without neutralizing @mentions. Release notes commonly credit contributors, so a relevant paragraph can make the monitor notify unrelated upstream users or teams from the generated SkillView issue. Apply the same mention sanitization before storing the highlight.

Pass repository refs via --pin instead of positional parsing

src/​SkillView.Core/​Gh/​GhSkillInstallService.cs:116

gh skill install does not parse a ref from the repository positional; refs are accepted as SKILL@ref or via --pin. Consequently, a versioned listing passes owner/repo@ref to repository parsing and can fail instead of listing that ref. Since this path has no skill positional, pass --pin=<ref> before the separator and keep the repository unchanged.

This issue also appears on line 302 of the same file.

[Copilot]

Copilot review overview

🟡 Changes recommended

Versioned installs remain incompatible with gh’s argument contract, and assessment publication has validation and reliability gaps.

Review effort: Balanced
Findings: 1

Open (1)

• Issue creation before dependency checks can permanently skip assessment · New

Resolved since last review (1)

• Enforce the advertised 500-word assessment limit

Previously missed (3)

In code that hasn't changed since last review

Verdict validation scans all text instead of compatibility section

.github/​scripts/​critical-dependency-assessment.js:32

This checks for a verdict anywhere in the generated text, not specifically in ### Compatibility assessment, and does not enforce the requested single verdict. For example, a release-summary sentence containing **Likely compatible** lets an unsupported assessment such as “Probably fine” pass validation and be published. Validate exactly one allowed verdict and require it to occur in the compatibility section.

Marker matching accepts comments from non-workflow authors

.github/​workflows/​critical-dependencies.yml:133

The marker alone does not prove that the existing comment belongs to this workflow. Any issue participant can post or quote this public marker first, causing a rerun to target an unrelated user's comment (either overwriting it or failing authorization) instead of maintaining the workflow's own assessment. Restrict the match to the GitHub Actions bot author.

Versioned discovery passes ref as repository instead of --pin

src/​SkillView.Core/​Gh/​GhSkillInstallService.cs:116

Versioned discovery still puts @<ref> on the repository argument. In both supported gh 2.97.0 and 2.102.0, the first positional is parsed strictly as the repository, while a ref is selected via --pin <ref> (or via skill@version when a skill exists). Thus ListRepoSkillsAsync(..., "v1.2.0") asks gh for a repository literally named repo@v1.2.0 and fails instead of listing that ref. Pass the version as the trusted --pin value before the separator and keep the repository unchanged.

This issue also appears on line 301 of the same file.

[Copilot]

Copilot review overview

🟡 Changes recommended

Versioned discovery, assessment validation, comment ownership, and release-note mention handling have unresolved defects.

Review effort: Balanced
Findings: 2

Open (2)

• Reuse only marker comments authored by the Actions bot · New
• Pin repository refs with --pin instead of appending @​ref · New

Resolved since last review (1)

• Issue creation before dependency checks can permanently skip assessment

Previously missed (2)

In code that hasn't changed since last review

Validate verdict only within the compatibility assessment section

.github/​scripts/​critical-dependency-assessment.js:32

The verdict regex scans the entire response, so malformed output is accepted when an allowed phrase appears in another section while ### Compatibility assessment contains no required verdict. Validate exactly one allowed verdict at the start of the extracted compatibility section before publishing.

Sanitize contributor mentions in copied release highlights

.github/​scripts/​critical-dependencies.js:68

Relevant excerpts are copied verbatim, so a normal generated release-note bullet such as ... by @contributor pings that upstream user when this workflow opens its issue. releaseOverview already neutralizes these mentions; apply equivalent sanitization here before embedding highlights.

[Copilot]

Copilot review overview

🟡 Changes recommended

Credential exposure, comment-marker spoofing, and install argument validation defects remain unresolved.

Review effort: Balanced
Findings: 2 · 1

Open (3)

• Disable credential persistence before running untrusted assessments · New
• Reuse only marker comments authored by the Actions bot
• Accept assessment markers only from the workflow bot · New

Resolved since last review (1)

• Pin repository refs with --pin instead of appending @​ref

Previously missed (4)

In code that hasn't changed since last review

Require the verdict at the start of the compatibility section

.github/​scripts/​critical-dependency-assessment.js:32

The verdict regex scans the whole assessment rather than the compatibility section. For example, an **Unknown** quoted under “What changed” lets an assessment whose compatibility section has no verdict pass validation and be published. Require exactly one allowed verdict and require it at the start of sections[2].

Only edit bot-authored comments; otherwise create a new comment

.github/​workflows/​critical-dependencies.yml:133

The first marked comment may be user-authored. In that case the workflow tries to edit a comment it does not own, receives a permission error, and never publishes the assessment. Restrict replacement to the workflow bot's marked comment; otherwise create a new one.

Reject mutually exclusive flags before environment probing

src/​SkillView.Core/​Cli/​CliDispatcher.cs:751

The boolean parser allows --from-local and --upstream together, and the updated parser test even constructs that combination. Every supported gh version rejects these flags as mutually exclusive, so the CLI performs environment/inventory work only to fail in the subprocess. Return invalid usage before probing, and keep this validation aligned with the TUI/service path.

Disable upstream mode when local mode is selected

src/​SkillView.Core/​Ui/​InstallScreen.cs:313

This checkbox value is passed independently of FromLocal, so selecting both produces a command that gh rejects because --upstream and --from-local are mutually exclusive. Disable and clear the upstream checkbox while local mode is selected (and enforce the same invariant at the service/CLI boundary) instead of allowing an install that is guaranteed to fail.

[Copilot]

Copilot review overview

🟡 Changes recommended

Malformed marked assessments can suppress retries, and invalid compatibility statuses can pass validation.

Review effort: Balanced
Findings: 1 · 2

Open (3)

• Reuse only marker comments authored by the Actions bot
• Malformed marked comments incorrectly suppress assessment retries · New
• Accept assessment markers only from the workflow bot

Resolved since last review (1)

• Disable credential persistence before running untrusted assessments

Previously missed (1)

In code that hasn't changed since last review

Status validation accepts allowed values from unrelated sections

.github/​scripts/​critical-dependency-assessment.js:38

The allowed-status regex scans the entire output, so **Unknown** in “What changed” lets an arbitrary value such as “Probably compatible” pass in the actual compatibility section. Validate that exactly one allowed status occurs and that it belongs to sections[2] before publishing.

[Copilot]

Copilot review overview

🔵 Needs a closer look

It combines model-driven issue publishing with user-visible install security changes, warranting final human validation.

Review effort: Balanced
Findings: None

Resolved since last review (3)

• Reuse only marker comments authored by the Actions bot
• Malformed marked comments incorrectly suppress assessment retries
• Accept assessment markers only from the workflow bot