_base

Base Agent Structure

Sections inherited by all subagents. Do not invoke this file directly.

Consult Skills

For framework-specific patterns, consult the corresponding skill in .ai/skills/{name}/SKILL.md before implementing.

Where You Operate (Permissions)

Each agent defines where it can act. Always respect these limits:

Permission	Meaning
Can write	The agent can create and modify files in these paths
Read only	The agent can read but NOT modify
Cannot touch	The agent must NOT read or modify these paths

If an agent does not define its own "Where You Operate" section, it inherits the permissions from its tools in the frontmatter.

Branch Strategy

Agents respect the project's branch strategy:

Branch	Agents
main	Cannot merge or commit directly
staging	Can open PRs
feature/*	Can commit

If an agent makes a mistake, it only affects a feature branch, never production.

Lesson Learned → Rule

Every Fatal Restriction and Quality Gate exists because its absence caused a real problem. When documenting a new restriction, use this format:

**Restriction:** [what not to do]
**Origin:** [what happened when it was done]
**Impact:** [real consequence of the error]

Restrictions are not bureaucracy. They are lessons learned turned into rules.

Continuous Verification (optional)

By default, verification is done at the end (checkpoint). For critical features, continuous verification can be enabled:

• Checkpoint (default): verify Quality Gates when finished. Low token usage.
• Continuous (optional): self-evaluate at each intermediate step. 3-5x more tokens but catches errors earlier.

Use continuous verification only when:

• The feature touches multiple layers simultaneously
• There is high security risk
• The context is very large and information could be lost

To enable it, tell the agent: "Use continuous verification for this task."

Final Verification