Skip to content
APEX

Dev Container Hygiene

This guide explains how APEX keeps GitHub Copilot Chat context lean in the dev container, and what contributors can do locally when extension-contributed chat customizations inflate every turn.

The largest avoidable source is extension-contributed customization: chat skills, chat agents, chat prompt files, and chat participants registered by VS Code extensions. These can be loaded into every Copilot Chat turn even when they are unrelated to the APEX workflow.

Repository Mitigations

Section titled “Repository Mitigations”

User-Scope Discovery Is Disabled

Section titled “User-Scope Discovery Is Disabled”

The workspace settings disable user-profile customization discovery for this repository:

  • chat.instructionsFilesLocations disables ~/.copilot/instructions and ~/.claude/rules.
  • chat.agentFilesLocations disables ~/.copilot/agents and ~/.claude/agents.
  • chat.agentSkillsLocations disables ~/.copilot/skills and ~/.claude/skills.
  • chat.useClaudeMdFile is set to false because this repository uses AGENTS.md.

The same settings are mirrored into the dev container VS Code customizations. They are workspace-scoped, so personal prompts and instructions still work in other repositories.

The Dev Container Extension List Is Curated

Section titled “The Dev Container Extension List Is Curated”

The dev container extension list excludes extensions that contribute heavy Copilot Chat customizations without serving the APEX workflow. The policy is recorded next to the customizations.vscode.extensions array in .devcontainer/devcontainer.json.

Unwanted Extension Recommendations Warn Contributors

Section titled “Unwanted Extension Recommendations Warn Contributors”

The workspace uses .vscode/extensions.json unwantedRecommendations to flag extensions that commonly add large chat customization payloads:

  • ms-azuretools.vscode-azure-github-copilot: duplicates APEX’s end-to-end agent set.
  • ms-windows-ai-studio: AI Toolkit is not used by the APEX flow.
  • teamsdevapp.vscode-ai-foundry: AI Foundry is not used by the APEX flow.

When one of these extensions is installed, VS Code shows a workspace-specific recommendation dialog. Accepting the prompt removes the extension from this workspace environment.

CI Rejects Denylisted Extensions

Section titled “CI Rejects Denylisted Extensions”

npm run validate:extension-bloat rejects changes that add denylisted extensions to the dev container extension list. The denylist lives in tools/scripts/validate-extension-bloat.mjs.

Borderline extensions can remain as unwantedRecommendations only. That keeps the warning visible without blocking contributors who deliberately need a tool for work outside the APEX flow.

Contributor Cleanup

Section titled “Contributor Cleanup”

Acknowledge the VS Code Dialog

Section titled “Acknowledge the VS Code Dialog”

When the workspace opens, accept the unwanted extension recommendation dialog for flagged extensions. This is the simplest way to reduce per-turn context in this repository.

Remove Extensions Globally When Appropriate

Section titled “Remove Extensions Globally When Appropriate”

To remove flagged extensions from every workspace on your machine, run these commands from your host VS Code environment:

Terminal window
code --uninstall-extension ms-azuretools.vscode-azure-github-copilot
code --uninstall-extension ms-windows-ai-studio
code --uninstall-extension teamsdevapp.vscode-ai-foundry

Reinstall an extension later with code --install-extension <id> if you start using it in another workspace.

Trim User-Profile Prompt Files

Section titled “Trim User-Profile Prompt Files”

Personal *.instructions.md and *.prompt.md files in your VS Code user profile load globally by default. The workspace suppresses them for APEX, but they can still add context in other repositories.

Common locations:

  • Windows: %APPDATA%\Code\User\prompts\
  • macOS: ~/Library/Application Support/Code/User/prompts/
  • Linux: ~/.config/Code/User/prompts/

Inspect What Loaded

Section titled “Inspect What Loaded”

Right-click in the Copilot Chat view and select Diagnostics. The diagnostics view lists every active agent, skill, instruction, prompt, and hook, including where each one came from.

Use this when you need to confirm that workspace mitigations are active or when a new extension appears to be adding unexpected context.

Adding Dev Container Extensions

Section titled “Adding Dev Container Extensions”

Before adding an extension to .devcontainer/devcontainer.json, inspect the extension package.json for these contributes keys:

  • chatSkills
  • chatAgents
  • chatPromptFiles
  • chatParticipants

If the extension contributes chat customizations that overlap with APEX’s agent and skill model, prefer a per-developer install. If the extension should be blocked for everyone, update the denylist in tools/scripts/validate-extension-bloat.mjs and document the reason.

Parallel Chat Retry Race

Section titled “Parallel Chat Retry Race”

VS Code Copilot Chat can occasionally issue the same model request twice in parallel, usually during slow or rate-limited turns. The second response may replace the first in chat history. Both requests can count against input tokens, but only the later response is visible.

This is a client-layer behavior, not an APEX agent behavior. Agent prompts cannot reliably prevent it because the retry happens outside the agent’s execution context. Saved telemetry can reveal the pattern by grouping chat: spans with the same gen_ai.request.id within a short time window.

When filing an upstream issue, use the Copilot Chat feedback issue template in .github/ISSUE_TEMPLATE/copilot-chat-feedback.md and include the relevant saved telemetry details.

Section titled “Related”