[Idea] On-Demand Skill Retrieval & Recovery Script to prevent context crash loops on Windows

[BibekG1]

What problem are you trying to solve?

When running a full installation via npx agentic-awesome-skills --antigravity, the library copies all 1,980+ skills directly into the active .agents/skills folder. In Antigravity IDE, this causes the agent to attempt to parse every single skill on startup, immediately exhausting the token context window. The IDE crashes on launch and enters an infinite restart loop with the error: "TrajectoryChatConverter: could not convert a single message before hitting truncation." Currently, once caught in this loop, the app cannot start, and users are stuck. Furthermore, loading all metadata upfront leads to "context rot" and distraction (false positive tool activations).

Proposed idea

We propose two changes:

1. On-Demand Skill Retrieval Strategy:
   Introduce an --on-demand installer flag. This clones the full library to a static reference path but keeps the active ~/.agents/skills directory empty. The agent's instructions (using rules in GEMINI.md) look up skills_index.json to find the correct skill ID and read only that single SKILL.md dynamically when needed, releasing it afterward. This results in zero token overhead on startup.

2. Windows Recovery Utility:
   Provide a standardized PowerShell recovery script (e.g. Fix-AntigravityCrash.ps1) in the repository. This script clears the corrupted browser session/local storage databases causing the startup loop without touching project files, allowing users to recover instantly.

Real use cases

• Who needs this? Any Windows user utilizing Antigravity or a similar IDE with global skills.
• What task are they doing? Coding with the skills catalog without burning their entire token/instruction budget on startup.
• What does success look like? The IDE starts instantly with zero context overhead, dynamically queries skills only when requested, and has a one-click recovery path if a session database gets corrupted.
  aas-feedback-report.txt

Related skills, tools, or prior art

• Similar skill: Antigravity IDE global rules configuration (GEMINI.md).
• Related doc: see the attached report detailing our custom GEMINI.md implementation and the recovery script code.

Are you willing to help build this?

No, just sharing the idea

Before posting

• I searched existing discussions and issues for similar ideas.

[BibekG1]

First of all, a huge thank you to the maintainers and contributors for building and keeping this project active! This library is a massive upgrade for agent capabilities and helps a lot of developers streamline their coding loops. I'm really grateful for your hard work and dedication.

This should resolve several critical technical trade-offs that occur when loading a library of this size:

• Context Rot & Distractions: Releasing/deactivating skills immediately after a task is done prevents the agent from getting distracted by unused tools in subsequent follow-up prompts, leading to higher quality outputs.
• Instruction Budget Exhaustion: Since we keep the active directory empty and load only the matching SKILL.md dynamically, it prevents the entire token and instruction budget from being burned on startup metadata.
• False Positive Activations: Searching the index cleanly first eliminates the likelihood of the agent getting confused and triggering irrelevant tools.
• Hallucinations & Overcomplication: Clean lifecycle management ensures the agent doesn't carry over instructions from a previous planning phase (like brainstorming) into a coding or refactoring phase.

Moving forward, I think it would be amazing if the installer could support this on-demand retrieval flow automatically under the hood next time we run the install command.

For example:

1. The installer could detect the host platform (like Antigravity) and automatically configure the empty active directory / static storage structure.
2. It could automatically append the GEMINI.md/AGENTS.md rules during the install process so users don't have to write them manually.
3. Alternatively, exploring if the local MCP server can serve as the primary on-demand provider natively across all IDE configurations might be a great long-term addition to the roadmap.

I hope this provides useful insights for the project's design! Keep up the great work! ❤️

[sickn33]

Thanks for reporting this and for the detailed reproduction. The failure mode is real, and we have implemented the smallest safe fix on main in #1047.

For Antigravity only, a bare install or --antigravity without a selection now stops before cloning or changing files. The message:

• explains the context-exhaustion and crash-loop risk;
• recommends asking a Codex or Claude agent with the read-only AAS Core MCP to inspect the project and choose exact skill IDs;
• provides a safe --skills ... --dry-run command and a copyable agent prompt;
• keeps exact selections and metadata filters working normally;
• requires an explicit --all override to install the complete catalog.

The MCP deliberately selects and validates skills but does not install them. This keeps the flow inspectable: the agent chooses the minimal set, the user reviews the dry run, and the existing installer performs the exact installation.

We did not add automatic GEMINI.md mutation or another static-copy lifecycle at this stage. AAS Core already provides on-demand catalog search without loading every skill into the watched directory, so the guard solves the immediate problem without introducing another installation mode or silently changing host instructions.

The error message also links the existing Windows recovery guide.

The change is merged and fully green on main; it will become available through npx agentic-awesome-skills with the next npm release. Thanks again — this report directly led to the safer Antigravity default.