odin-gemini-cli-extension AGENTS.md

Gemini Code Agent Manual

<agent_selection> Selection Algorithm:

1. Scope (PRIORITY): simple/trivial -> Fast | default -> Tidy-First | tricky/other perspective -> Smartass | challenge/stress-test/pre-mortem -> Advocate | unknown lib -> Librarian | exploration needed -> Explore
2. Domain: security|auth|jwt -> Security-Auditor | thread|async|mutex -> Concurrency-Expert | alloc|heap|leak -> Memory-Expert | component|css|a11y -> UI-UX-Designer | trading|order|latency -> Trading-System-Architect
3. Risk: production-critical -> Perfectionist/High-Fidelity
4. Review: pre-merge/PR -> Reviewer

Delegation Enforcement (MANDATORY):

• Always plan and write todos in consideration of launching and utilizing multiple agents; Include the multiple agents utilization in the plans while reasoning.
• Default to spawning multiple agents in parallel for any multi-step task, multi-file analysis, or parallelizable subtasks. [MANDATORY]
• Use Explore agents for codebase discovery and unknown code paths; use domain experts when keywords match.
• Provide precise instructions: scope, files/paths, success criteria, and whether to write code or only research.
• Do not delegate for trivial single-file edits, known file reads, or direct short answers.

Subagents: Spawn for complex tasks benefiting from independent execution. Own context, can't communicate mid-task, main agent receives final summary only.

• Multi-step tasks divisible into independent parts
• Operations producing extensive output not needed after
• Parallel work across different code areas
• Keep main context clean while coordinating complex work

Second Opinion (Oracle/Deep-Reasoner/Advocate): Extended reasoning for complex analysis.

• Debug complex multi-file issues
• Review intricate logic or refactoring plans
• "Ask Oracle whether there's a better solution"
• "Use Deep-Reasoner to analyze how X and Y interact"
• "Have the Advocate stress-test this design before we commit"
• "Use Advocate for a pre-mortem on this migration plan"

Librarian: Cross-repository research and external code search.

• Search remote codebases (GitHub public/private)
• Read framework/library source code
• Investigate upstream changes or API differences

Pre-Delegation: Always run git branchless init before delegating. </agent_selection>

Oracle

Senior engineering advisor for planning, reviewing, analyzing, and debugging complex tasks.

WHEN TO USE:

• Code reviews and architecture feedback
• Debugging multi-file issues
• Planning complex implementations or refactoring
• Analyzing code quality and suggesting improvements
• Answering complex technical questions requiring deep reasoning

WHEN NOT TO USE:

• Simple file reading or searching (use Read directly)
• Codebase searches (use Explore)
• Basic code modifications (do directly or use Execution agents)
• Web browsing/searching (use web tools)

USAGE GUIDELINES:

1. Be specific about what you want reviewed, planned, or debugged
2. Provide relevant context and list specific file paths
3. Mention oracle invocation to user ("consulting the Oracle...")

Deep-Reasoner

Extended reasoning specialist with large context window for complex multi-file analysis.

WHEN TO USE:

• Multi-file analysis requiring large context windows
• Complex dependency mapping across modules
• System-wide refactoring planning
• Extended reasoning chains with many steps

WHEN NOT TO USE:

• Quick lookups (use Explore)
• Single-file changes (use Execution agents)
• Simple searches (use Grep/Read)