Back to list
abdelrahman146

tool-minimization

by abdelrahman146

Kyora is a SAAS service that helps social media store owners manage their business and finances

0🍴 0📅 Jan 26, 2026

SKILL.md


name: tool-minimization description: 'MCP/tool selection rules and lane-default toolchains for Kyora Agent OS. Use when selecting tools for a task, configuring agent tool lists, or understanding when to use MCP servers. Triggers: tools, MCP, tool selection, least-privilege, Context7, Playwright, GitHub MCP.'

Tool Minimization Workflow

Operational guide for selecting and minimizing tools in Kyora Agent OS. Implements least-privilege tool selection and safe MCP usage.

When to Use This Skill

  • Selecting tools for a new task
  • Configuring agent tool lists
  • Deciding when to use MCP servers
  • Reducing tool overload in agents/prompts
  • Understanding lane-default toolchains

Prerequisites

  • Understanding of current task scope and risk level
  • Knowledge of available MCP servers (if considering MCP)
  • Access to agent/prompt files for tool list auditing

Core Principle: Least Privilege

Start with the minimum tools needed. Add more only when clearly necessary.

Tool overload hurts reliability:

  • More tools = more confusion for the model
  • More tools = higher chance of incorrect tool selection
  • More tools = longer context

Tool Selection Protocol

Step 1: Start with Workspace Read/Search

Default starting tools:

  • read — Read file contents
  • search — Find files and text (grep, glob)
  • semantic_search — Fuzzy concept matching

These solve most discovery and planning tasks.

Step 2: Add Only What's Needed

NeedTool to Add
Modify filesedit
Run commandsexecute
Run teststests / runTests
Track taskstodo
Invoke agentsagent

Step 3: Consider MCP Only When Beneficial

MCP tools add latency and complexity. Use only when they clearly reduce work or increase certainty.

Lane-Default Toolchains

LaneDefault ToolsAdd Only If Needed
Discoveryread, searchGitHub MCP (repo context), Context7 (library uncertainty)
Planningread, searchContext7 (API usage), diagrams only if asked
Implementationread, search, edit, execute, testsPlaywright (UI), Context7 (new dep)
Validationexecute, testsPlaywright (UI smoke), DevTools (perf)
Recoveryread, search, get_changed_filesNone unless blocked

Role-Default Tool Lists

Orchestrator (Least Privilege)

tools: ['read', 'search', 'todo']
# edit only for plans/notes, NEVER production code

Leads

tools: ['read', 'search', 'edit']  # edit for specs only
# execute optional for validation

Implementers

tools: ['read', 'search', 'edit', 'execute']
# tests tool added as needed

QA/Test Specialist

tools: ['read', 'search', 'edit', 'execute']
# Playwright preferred for UI flows

Security Reviewer (Read-Only)

tools: ['read', 'search']
# NEVER edit or execute

MCP Server Policy

Approved MCP Servers

ServerPurposeWhen to Use
Context7Up-to-date library docs/snippetsNew dependency; unfamiliar API
PlaywrightUI automation, screenshotsUI flows, RTL verification, E2E
Chrome DevToolsLayout/perf/network debugPerformance issues, layout bugs
GitHub MCPIssues/PRs/policiesRepo context (NOT local-only tasks)

When NOT to Use MCP

  • Simple local refactors
  • Tasks solvable via workspace search + reading existing code
  • Anything requiring secrets or sensitive logs
  • Local-only tasks (don't use GitHub MCP)

MCP Safety Rules

See MCP Safety Checklist for detailed rules.

Key points:

  • Treat local MCP servers as code execution
  • Never paste secrets; use VS Code input variables
  • Keep enabled tools under model/tool limits

Tool Correctness Rules

  1. Prefer workspace tools for local codebase truth
  2. Use Context7 only to confirm third-party API usage; stop after minimum needed
  3. Use Playwright/DevTools when visual/RTL/layout correctness matters
  4. Disable entire servers if not relevant to current task

Operational Notes

Tool List Caching

  • Tool lists are cached by VS Code
  • Reset cached tools when server tools change
  • Use "Copilot: Reset Chat" or restart session

Tool Sets

  • Group related tools into tool sets
  • Keep enabled tools under model limits
  • Disable unused servers to reduce picker noise

Success Checklist

  • Started with read/search only
  • Added tools only when clearly needed
  • MCP tools have clear justification
  • No unnecessary tools in final list
  • Tool list matches role defaults
  • Safety rules followed for MCP
  • No surprise docs generated
  • Stop-and-ask triggers checked

Stop-and-Ask Triggers

MUST ask before proceeding if:

  • Adding MCP server not in approved list
  • Tool requires secrets or sensitive data access
  • Tool list exceeds role defaults significantly
  • Uncertain if tool is necessary for task

Troubleshooting

IssueCauseSolution
Tool not availableNot in enabled listCheck agent frontmatter tools: list
MCP server slowNetwork latency or server loadConsider if workspace tools suffice
Tool list too longScope creepReset to lane defaults, add only essentials
Cached tools staleServer tools changedUse "Copilot: Reset Chat" or restart

Validation Commands

No specific commands for tool selection, but verify:

# Check agent tool lists
grep -r "tools:" .github/agents/

# Check prompt tool lists
grep -r "tools:" .github/prompts/

References

SSOT References

Score

Total Score

50/100

Based on repository quality metrics

SKILL.md

SKILL.mdファイルが含まれている

+20
LICENSE

ライセンスが設定されている

+10
説明文

100文字以上の説明がある

0/10
人気

GitHub Stars 100以上

0/15
最近の活動

3ヶ月以内に更新がある

0/10
フォーク

10回以上フォークされている

0/5
Issue管理

オープンIssueが50未満

+5
言語

プログラミング言語が設定されている

+5
タグ

1つ以上のタグが設定されている

0/5

Reviews

💬

Reviews coming soon