[Vsevolod Kukol (sevoku)]

Description

Align Cosmos DB Shell guidance evaluations with explanation-only requests so unavailable MCP tools and equivalent account-identity wording are not treated as execution failures.

• Make the read and deletion prompts explicitly non-executing, with a fallback when MCP tools are unavailable.
• Accept equivalent account/endpoint wording and conditional schema-discovery plans without requiring tool calls or JSON.
• Make the intended database and coverage of Shell matching defaults, unsupported safety options, backups, and uncertain outcomes explicit in the prompts.
• Add concise preflight checklists to the existing Shell rules and regression checks for evaluation scope and retained safety requirements.

Skill activation, strict correctness thresholds, and deletion safeguards remain unchanged. No new rules or live Shell integration are added.

Type of Change

• Rule improvement
• Bug fix
• Documentation

Related Issues

Related to #281 and the Shell guidance failures in Vally run 40371778679.

[Copilot]

🟢 Approval recommended

The focused changes consistently support explanation-only requests while retaining target verification, deletion safeguards, and existing grading requirements.

No open issues 🎉

What changed in this PR

Aligns Cosmos DB Shell evaluations with explanation-only requests, avoiding execution requirements when MCP tools are unavailable while preserving existing safety guidance.

Changes:

• Clarifies prompts and rubrics to accept equivalent account terminology and proposed procedures without tool calls or JSON.
• Adds read and deletion preflight checklists to the existing Shell rules.
• Adds regression assertions for explanation-only scope and retained deletion safeguards.

File	Description
skills/​cosmosdb-best-practices/​rules/​tooling-cosmosdb-shell-writes.md	Adds deletion preflight checks and explanation-only fallback guidance.
skills/​cosmosdb-best-practices/​rules/​tooling-cosmosdb-shell-target.md	Adds read preflight checks and unavailable-MCP guidance.
evals/​cosmosdb-best-practices/​tooling/​eval.yaml	Aligns Shell prompts and rubrics with non-executing explanations.
evals/​cosmosdb-best-practices/​graders.test.mjs	Adds assertions preserving evaluation scope and safety requirements.
CHANGELOG.md	Records the Shell guidance and evaluation updates.

_{Review effort: Balanced}

[Mike Krüger (mkrueger)]

These changes look good: explanation-only requests are now evaluated on the proposed procedure rather than MCP availability or actual execution. Accepting semantically equivalent wording reduces false failures while preserving the existing targeting, ETag, and confirmation safeguards. Offline validation and the live CI evaluations passed.

[Mike Krüger (mkrueger)]

I think it's a good approach - installing cosmosdbshell would be an overkill.