global-context-management

Manages long agent sessions with durable task state, read-only subagents, and focused validation.

7|5|Updated Jul 22, 2025
One-click install
npx skills add https://github.com/nebius/nebius-ps-services --skill global-context-management-nebius
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: global-context-management
Source: https://github.com/nebius/nebius-ps-services/tree/main/skills/global-context-management
Command: npx skills add https://github.com/nebius/nebius-ps-services --skill global-context-management-nebius

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve? Long or complex coding sessions in Codex or Claude Code lose earlier decisions, flood the parent thread with raw logs and broad file dumps, and hit context compaction. This Skill keeps the parent thread concise by separating runtime policy, durable task-state continuity, and bounded read-only exploration helpers. ## Core Features & Use Cases - Durable task state: Hooks advertise a private per-session current.md path under the agent home so decisions, validation status, and next actions survive compaction without re-explaining context in chat. - Bounded read-only subagents: Optional repo_mapper, test_strategist, and risk_reviewer helpers handle broad repo mapping, test discovery, and near-final risk review without polluting the parent thread. - Parent-thread discipline: Enforces targeted reads, focused patches, narrow-first validation, and a final lifecycle sweep that closes every spawned subagent handle. - Use Case: During a multi-file refactoring that triggers CI failures, the agent reads the injected task-state file, delegates repo mapping to a read-only helper, implements focused patches, runs narrow validation, and records the current status before the final summary. ## Quick Start Ask the agent to use the global-context-management skill for a complex multi-file task so it maintains durable task state and uses bounded read-only helpers where authorized.

Frequently Asked Questions about global-context-management

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I prevent AI coding agents from losing context in long sessions?

Use durable task-state files that record the objective, decisions, changed files, validation status, and next action. Hooks inject a private per-session path, and the agent reads and updates that rolling summary instead of re-explaining context in chat.

How do I use read-only subagents for code exploration in Codex?

Delegation requires authorization from the current prompt or a user-enabled local hook policy, plus runtime multi-agent support. Once authorized, the parent agent spawns bounded read-only roles like repo_mapper or test_strategist, waits for summaries, and closes each handle.

Does this context management workflow work with Claude Code?

Yes, it supports both Codex and Claude Code through native invocation syntax and per-host homes. Claude uses CLAUDE_CONFIG_DIR with roles named repo-mapper, test-strategist, and risk-reviewer, configured through the config-claude setup skill.

Why is my AI agent still hitting context limits with hooks installed?

Hooks and skills cannot remove context already in the thread, such as prior turns, tool outputs, or system instructions. Broad shell output, large file dumps, and repeated exploration in the parent thread still consume context regardless of setup.

When should I not use subagent delegation for coding tasks?

Skip delegation for tiny single-file edits, when the runtime lacks multi-agent tools, or when active instructions forbid it. Do not spawn every configured role by default; reserve risk review for non-trivial or risky near-final changes.