[infra] agentic tool-use and token-efficiency guidance

.agents/AGENTS.md

@@ -0,0 +1,11 @@
+# Custom Rules for AI Coding Assistants
+
+To ensure all agentic tasks in this repository are executed as accurately, securely, and token-efficiently as possible, you must strictly follow these rules:
+
+## 1. Tool Selection & Efficiency
+* **Prefer Domain-Specific MCP Tools:** Before running generic terminal commands (e.g. `grep`, `find`, or compiler check scripts), always check if a domain-specific MCP tool (such as `lsp`, `analyze_files`, or `rip_grep_packages` from `dart-mcp-server`) is available and use it instead.
+* **Limit Command Output:** Any command-line tool execution (`run_command`) that might print large datasets, build logs, or file listings must have its stdout restricted using filters or limits (e.g. `head -n 50`, `git log -n 5`) to prevent context bloat.
+* **Targeted File Reading:** When reading source files, specify precise line boundaries with `StartLine` and `EndLine` parameters in `view_file` to read only the code you need. Avoid loading entire files over 100 lines.
+
+## 2. Task Delegation & Subagents
+* **Isolate Exploratory Research:** For open-ended research, codebase discovery, or scanning third-party docs, delegate the subtask to the `research` subagent. The subagent will run in a separate workspace context and return a concise summary, keeping the main conversation's history clean and token-efficient.

.claudeignore

@@ -0,0 +1 @@
+.geminiignore

.geminiignore

@@ -0,0 +1,8 @@
+# Exclude baseline output logs and reports from AI scanning
+third_party/tool/baseline/
+**/build/reports/
+**/build/test-results/
+
+# Exclude large test fixture mock directories
+**/testData/
+**/sample_tests/

.gitignore

@@ -1,4 +1,20 @@
 # Mac Finder settings
 .DS_Store
 
+
+# IntelliJ and IDEs
+.idea/
+**/.idea/
+*.iml
+third_party/.settings/
 /.vscode
+
+# Gradle
+.gradle/
+**/build/
+out/
+
+# Dart / Flutter
+.dart_tool/
+.packages
+pubspec.lock