This repository was archived by the owner on Dec 23, 2025. It is now read-only.
Repository navigation
This repository was archived by the owner on Dec 23, 2025. It is now read-only.
Path resolution may be too strict #2
Description
consider this simple test skill:
---
# ~/.opencode/skills/klorpify/SKILL.md
name: "klorpify"
description: "Use this when the user says 'klorpify'"
---
# klorpify guidelines
When the user asks to "klorpify" something, follow these guidelines:
1. create a klorp.txt file. write the current date to it.
2. say "klorpified!"when you ask please klorpify, OpenCode will try to write to ~/.opencode/skills/klorpify/klorp.txt instead of the current project directory. I'm not sure if this is intended!
This can be worked around by changing the instructions to "create a klorp.txt file in the current working directory."
I've done a quick fork with an idea on how to solve this: https://github.com/rstacruz/opencode-skills/blob/main/index.ts#L16-L26
Activity
malhashemi commentedon Oct 21, 2025
I am working on an improvement here,
I was doing some investigative work into how Claude Code uses skills, it seems that the SKILL.md get loaded as a user message with very minimal instructions. It looks something like this.
The "{{skill_name}}" skill is running {{skill_name}}Launching skill: {{skill_name}}I think this makes way more sense than having SKILLS.md being sent as response. from my understanding OpenCode dynamically deletes tools responses from the context so user message could be more persistent.
I think I will replicate this behavior and do some testing before releasing. I have not played with inserting message into the session using the SDK not sure if we can make that text invisible as it is too verbose, or if OpenCode support inserting hidden user messages via the SDK. I am doing user message insertion via the SDK on another yet to be published plug-in so I know that this could be achieved here, however that one does it via the TUI so could be slightly different.
malhashemi commentedon Oct 26, 2025
@rstacruz This now matches Anthropic responses and uses user message rather than tool response which prevents tool response purging, I am a bit nervous about being way less verbose with instructions on how to resolve paths, but this is how Claude Code does it, if the community finds the new approach to cause any issues we can revert back to being verbose.
airtonix commentedon Nov 5, 2025
Does anthropic claude code make bash/read calls outside of its allowed fs boundary (does it even have an allowed fs boundary) ?
If we want to enforce that boundary in opencode, we might want to think about adjusting the injected skill:
and then provide the tool
skill_read_skill_resource