Agent Skills: Kiro Context Management

Kiro Context Management

• Scope: Kiro-specific context management, terminal limitations, tool preferences
• Load if: Using Kiro IDE (ALWAYS load first in Kiro sessions)
• Prerequisites: None (this file loads before other rules)

CRITICAL: Terminal Limitations (Primacy Zone)

Terminal commands that cause Kiro to freeze:

• NEVER use echo with double quotes (causes hang)
• NEVER use heredoc syntax (<<EOF) (fails in Kiro terminal)
• NEVER use complex zsh themes (powerlevel10k, oh-my-zsh cause timeouts)
• NEVER use interactive editors (vim, nano) in automated commands
• NEVER use pagers (less, more) in automated commands
• NEVER chain commands with && or || in complex scripts
• NEVER expect second command in same session to return reliably (zsh bug)

Terminal best practices for Kiro:

• Use Python scripts instead of complex shell commands
• Write output to files, then read files (not echo)
• Use -m flag with git commit (avoid editor)
• Use --no-pager with git commands
• Use simple bash syntax over zsh features
• Run one command per terminal invocation when possible
• Add timeouts to long-running commands

CRITICAL: Serena MCP is Mandatory

When Serena MCP is available, you MUST use Serena tools for all file operations.

Kiro native tools have known reliability issues:

• strReplace fails on duplicate content
• readFile silently truncates large files
• File writes can create duplicate sections
• Mid-operation aborts during file reads/edits

Tool Preference Order (mandatory):

1. Reading code: find_symbol, get_symbols_overview > readFile
2. Editing code: replace_content (regex mode) > strReplace
3. Navigation: find_referencing_symbols > grepSearch
4. Context: write_memory, read_memory for persistent state

• Using Kiro strReplace when Serena replace_content is available
• Using Kiro readFile for large files when Serena tools are available
• Ignoring Serena MCP availability

Serena Activation Workflow

At session start:

1. Check if Serena MCP is available
2. Run activate_project to initialize
3. Run check_onboarding_performed to verify setup
4. Run list_memories to discover available context
5. Run read_memory for relevant project context

During work:

• Use write_memory for important discoveries
• Use semantic tools (find_symbol, replace_content) over text-based

Before compaction (70% context):

• Save session state with write_memory

AGENTS.md Integration

Kiro automatically loads AGENTS.md from workspace root (Kiro v0.5.0+). No separate .kiro/steering/ files needed.

How it works:

• Place AGENTS.md at workspace root
• Kiro reads it automatically at session start
• Include instructions for loading domain-specific files
• Critical rules (terminal limitations) embedded directly with XML tags

• @smith-ctx/SKILL.md - Universal context strategies
• @smith-serena/SKILL.md - Serena MCP tools (mandatory for Kiro)

ACTION (Recency Zone)

Proactive context management:

1. Prompt for context status periodically
2. At 70%: Warn that 80% auto-summarize is approaching
3. Save important context to Serena memory before 80%

Platform auto-summarizes at 80% - agent prepares user by saving state first.

In Kiro sessions:

1. Avoid echo with double quotes
2. Use Python scripts for file generation
3. Use Serena MCP tools (mandatory, not optional)