[kyle-sexton]

Closes #9898

Summary

Replaces #11953 (lanes App PR), which merges here unchanged, plus fixes for its two open review threads and the incremental review's eval finding. The pr-unblock lane stopped #11953 with feedback-untaken-threads, so no lane would take them.

The change adds documentation and a setup skill only. It provisions nothing and bills nothing: the Azure SQL Database doc describes the free tier with --free-limit-exhaustion-behavior AutoPause, and no workflow or infrastructure file changes.

Fix

• Merges feat(sql-server): add the setup skill and Azure SQL Database docs #11953's head 280084a39fbb03eda32250e2cd2b8ae561bdc717: /sql-server:setup (check and apply), its scripts, tests and evals, reference/azure-sql-database.md, README Setup section, prerequisites, catalog line and changelog fragment.
• skills/setup/scripts/install-server.mjs: builds the copy and runs npm ci in <data-dir>/server.next, then swaps it in with renameSync only once the manifest and dependencies check out, restoring the old copy if the swap fails. A failed npm ci now leaves the live copy that registrations use untouched.
• test/setup.test.mjs: new test that a failing npm ci (a fake npm on PATH) exits 1 and leaves the live copy's manifest in place.
• skills/setup/SKILL.md, README.md, check-setup.mjs comment: the register-by-reference and no-project-scope decisions in our own words; the claude mcp get readback and the grants call are what confirm a credential arrived. New Upstream record pointing at https://code.claude.com/docs/en/mcp#environment-variable-expansion-in-mcp-json with as-of date and recheck trigger.
• Setup evals: cases 1 to 3 state that no config exists and expect the check's stop on the config FAIL; all four cases are converted by convert-evals-json from skills/setup/evals/evals.json into a plugin-level suite under plugins/sql-server/evals/, where the eval lane finds them.
• skills/setup/SKILL.md: no restated plugin-data persistence or MCP load timing; Upstream record pointers to https://code.claude.com/docs/en/plugins/components#path-variables-and-persistent-data and https://code.claude.com/docs/en/mcp#managing-your-servers.

Verification

• npm test in plugins/sql-server: 134 tests pass, including the new install-server test.
• check-skill.sh plugins/sql-server/skills: PASS (1 warning: no Gotchas section).
• validate-cases.py plugins/sql-server/evals: no FAIL (WARNs: judge-only graders and samples awaiting calibration, as for every converted suite).
• The docs anchor was read from the live page this session.

Related

• feat(sql-server): add the setup skill and Azure SQL Database docs #11953 (superseded; closes when this merges)

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-01T16:51:06.199014Z	6b94472	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-security-review reviewed 532780fe6fecfbdc12baa99c8111fb8fbba28edd: read 15 of 15 changed file(s); 0 finding(s), 0 as inline comments.

Security review (incremental, dcab900..532780f)

No security issues found in this change.

What changed and what I checked

• plugins/sql-server/skills/setup/SKILL.md: one sentence added. When the user asks for project scope, the skill now refuses in its first reply, before any step. This makes the existing --scope project refusal and the Boundaries rule ("Never write ... a project-scope registration") stronger. It removes or weakens no standing instruction, so the instruction-surface lens turns up nothing.
• Three setup eval cases (setup-apply-refuses-project-scope-on-config-fail, setup-check-reports-every-row-and-modifies-nothing, setup-literal-password-is-a-config-fail-with-env-remediation): allowed_tools now includes Bash, so the eval run can actually execute check-setup.mjs. The prompts are written in the repository and name only made-up paths (for example ~/sql-server/config.json). They don't pull in any text from outside the repository, so this grant gives no untrusted input a way to reach Bash. The new deterministic tool_used graders don't loosen anything either: no-install and no-register use max: 0 with arm: 'both' to fail a run that calls install-server.mjs or claude mcp add after a config FAIL.
• Grader sample JSON files (added, plus the renamed expectation-*.json): these are labelled data only. The sample commands are never executed.

Not touched by this increment: no trust controls, workflows, token or kill-switch paths, merge holds, or plugin manifests (.claude-plugin/*.json), so there are no manifest <field> lines.

Settled threads: none of them was re-raised. The two SKILL.md threads marked lines changed since resolution: yes were about restated upstream behaviour. The new line doesn't bring that back.

Spend: giving Bash to three eval cases lets eval runs execute a local read-only script. This uses only prepaid eval or subscription usage and is not a hold.

critical-findings: 0

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6b94472509

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".

[github-actions]

claude-review: 1 inline finding(s) on 6b94472509eb65d6d43dfe5889588325fc439032; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

claude-review reviewed 532780fe6fecfbdc12baa99c8111fb8fbba28edd: read 15 of 15 changed file(s); 0 finding(s), 0 as inline comments.

Incremental review, dcab900..532780f: no findings.

I read all 15 files in scope. The incremental diff had no patch for the three renamed expectation-*.json samples, so I read their head contents in the full diff.

What changed

• Deterministic runs-check-script grader. The three setup cases that run check-setup.mjs now have a tool_used grader for it (min: 1, arm: with-only), and Bash is added to their allowed_tools. In setup-check-reports-every-row-and-modifies-nothing, this grader replaces the old LLM expectation-1.
  • The input_match pattern "command":"(?:[^"\\]|\\.)*?check-setup\.mjs is the same form other suites in the repo use (for example plugins/plugin-quality/evals/re-pressed-for-second-parser/case.yaml).
  • Checked against the samples: both pass samples match, the ${CLAUDE_SKILL_DIR} form with its escaped quotes and the expanded form. The cat ~/sql-server/config.json fail sample does not match.
• Renamed case. setup-apply-refuses-project-scope-and-confirms-each-registration is now setup-apply-refuses-project-scope-on-config-fail. It gains deterministic no-install and no-register graders (max: 0, arm: both). Both fit what the case expects: a config FAIL stops apply at step 1, so no registration happens and none needs confirming. Each sample set separates the behaviour correctly: the check-only call passes, and install-server.mjs / claude mcp add fail.
• SKILL.md. The skill now refuses project scope in its first reply, before any step, so a run that stops early on a check FAIL still answers the request. This matches what the renamed case expects (expectation-3). It also stays inside .claude/rules/skill-bodies-state-current-rules.md: it gives our own decision and points to step 3's reason, without restating any upstream product behaviour.

Rules and context

• .claude/rules/eval-case-transcripts.md applies to the eval files. The samples use synthetic paths (/opt/plugins/..., /srv/sql-server/...) and contain no transcript material.
• Settled threads: the three threads on evals.json and SKILL.md stay settled. Nothing in this increment reopens their points: the new SKILL.md lines add no restated upstream behaviour.
• Closing issue feat: add the SQL Server MCP setup skill and Azure SQL Database docs #9898: this increment changes no acceptance criterion's evidence. The check-skill.sh and validate-cases.py results in the PR body are claimed by the author; I did not run them.
• No manifest changes in this increment.

Spend: the three cases now allow Bash, so eval runs can execute check-setup.mjs. That draws only on prepaid API credit or subscription usage; it is not a hold.

[github-actions]

claude-review: 2 inline finding(s) on dcab900ceb5eed9f114a8b019cf6593038964762; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

Post-merge mutation testing

• Range: 3d843d8..2e5f1d6
• Merged pull request: #11994
• Outcome: reported

Killed 3, survived 4, timeout 0, no coverage 0, invalid 0, not run 0.
Covered-code score 43%; mutation score 43%.

File	Killed	Survived	Timeout	No coverage	Invalid	Not run
plugins/sql-server/skills/setup/scripts/check-setup.mjs	3	2	0	0	0	0
plugins/sql-server/skills/setup/scripts/install-server.mjs	0	2	0	0	0	0

Surviving mutants

Location	Operator	Disposition	Cause
plugins/sql-server/skills/setup/scripts/check-setup.mjs:72	ROR	productive	input-gap
plugins/sql-server/skills/setup/scripts/check-setup.mjs:112	ROR	productive	no-assertion
plugins/sql-server/skills/setup/scripts/install-server.mjs:37	LCR	productive	input-gap
plugins/sql-server/skills/setup/scripts/install-server.mjs:56	ROR	productive	no-assertion

The run's summary has the full report.