Code Lens presents a moderately structured tool set with explicit parameter contracts defined in tool-contracts.ts. However, critical gaps emerge: (1) Tool descriptions are present but many lack specificity about WHEN to use them vs. similar tools, e.g., analyze_pr_impact, generate_review_summary, and detect_api_breaking_changes all address code changes but their distinctions are unclear. (2) Parameter constraints are documented as strings (e.g., '2-32 chars') rather than enforced in JSON Schema, forcing LLMs to interpret text constraints. (3) Output schemas are named ('outputShape: string') but never defined, LLMs cannot see the actual structure of response objects. (4) Error handling is mentioned in gotchas but lacks recovery guidance ('what should the LLM do if this fails?'). (5) The tool naming is largely sound (verb-driven: generate_, analyze_, detect_, load_, ask_, refactor_) but some compound concerns: generate_review_summary couples impact assessment + risk rating + merge recommendation into one tool rather than three focused tools. (6) Cross-tool dependencies (requiresDiff, requiresFile) are declared but validation logic is not visible in provided source, creating risk of runtime failures that LLMs cannot anticipate.

• Define output schemas explicitly for all Gemini-backed tools. Create a schema registry mapping tool name to response structure (e.g., analyze_pr_impact returns {severity: 'critical'|'high'|'medium'|'low', categories: string[], breakingChanges: string[], rollbackComplexity: number}). Document in tool-contracts.ts and reference from tool registration.
• Migrate all constraint declarations from human-readable strings to JSON Schema. For 'language' param, use enum: ['typescript', 'python', 'javascript', ...] instead of 'constraints: "2-32 chars"'. For numeric params like maxTestCases, use {type: 'number', minimum: 1, maximum: 30}.
• Expand tool descriptions to include WHEN to use and how tools differ. Example: 'analyze_pr_impact: Identifies risk factors and breaking changes in a diff. Use this to assess rollback difficulty. For a full merge recommendation, pair with generate_review_summary.' Length target: 100-180 chars.
• Split generate_review_summary into three focused tools: (1) generate_pr_summary (returns {summary, changed_files, affected_components}), (2) assess_pr_risk (returns {risk_level, risk_factors, affected_systems}), (3) recommend_merge_action (returns {recommendation, confidence, caveats}). This enables agents to compose analysis as needed.
• Add validation to tool execution that checks requiresDiff and requiresFile preconditions. Return clear error messages: 'generate_diff must be called first to cache a diff. Call generate_diff(mode="unstaged") then retry.' Reference pattern:recovery-guide.

Inferred effective spec: 2025-06-18+.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry