agent skills #22

Closed
opened 2026-07-03 22:23:01 +02:00 by ffaerber · 12 comments
Owner

i have a quation about the ai agents skills and context window.
if a agent1 has some skill like to read github. this will be also included always in the system prompt of a agent2 that dont has the skills.
i think its better to only include the simple info like agent1 can search and read github issues and repos
but how the api calles are done should not be included in agent2 system prompt only in agent1 system prompt.
if agent2 needs some info about a github issue he se that agent1 has some way todo that. he dont know how that is be done but he can ask.

i have a quation about the ai agents skills and context window. if a agent1 has some skill like to read github. this will be also included always in the system prompt of a agent2 that dont has the skills. i think its better to only include the simple info like agent1 can search and read github issues and repos but how the api calles are done should not be included in agent2 system prompt only in agent1 system prompt. if agent2 needs some info about a github issue he se that agent1 has some way todo that. he dont know how that is be done but he can ask.
Author
Owner

@pm go

@pm go

🤖 @pm

The user wants to optimize the system prompts of AI agents by preventing the full technical definitions of skills from leaking into the prompts of agents that do not possess those skills.

Currently, it seems all skill definitions are shared globally. The goal is to change this so that agents only see a high-level summary of what other agents can do (e.g., "@agent1 can search GitHub"), while the detailed API implementation remains exclusive to the agent that actually uses the skill.

Plan:

  1. Research how skills are currently loaded into the system prompts of agents.
  2. Modify the skill loading mechanism to differentiate between "full definition" (for the agent with the skill) and "summary definition" (for other agents).
  3. Ensure that agents without a specific skill only receive the summary, reducing context window usage and noise.

I believe @senior is best suited for this as it involves modifying the core prompt/skill orchestration logic across multiple files.

@ffaerber ready to start building? reply yes to proceed.

🤖 **@pm** The user wants to optimize the system prompts of AI agents by preventing the full technical definitions of skills from leaking into the prompts of agents that do not possess those skills. Currently, it seems all skill definitions are shared globally. The goal is to change this so that agents only see a high-level summary of what other agents can do (e.g., "@agent1 can search GitHub"), while the detailed API implementation remains exclusive to the agent that actually uses the skill. **Plan:** 1. Research how skills are currently loaded into the system prompts of agents. 2. Modify the skill loading mechanism to differentiate between "full definition" (for the agent with the skill) and "summary definition" (for other agents). 3. Ensure that agents without a specific skill only receive the summary, reducing context window usage and noise. I believe @senior is best suited for this as it involves modifying the core prompt/skill orchestration logic across multiple files. @ffaerber ready to start building? reply yes to proceed.
Author
Owner

@pm please call the gitea api with your token and find the your roles of waht you can do with the token. like manage git repos or create issues

@pm please call the gitea api with your token and find the your roles of waht you can do with the token. like manage git repos or create issues

🤖 @pm

Based on my gitea-api skill definition and the API results, the AGENT_TOKEN has read and write access to the following scope groups:

  • issue: Manage issues, PRs, comments, labels, and milestones.
  • repository: Manage repo contents and Actions runs/jobs/logs.
  • organization: List and read organization repositories.
  • misc: General miscellaneous API access.

It does not have access to admin, user, notification, package, or activitypub.

🤖 **@pm** Based on my `gitea-api` skill definition and the API results, the `AGENT_TOKEN` has **read and write** access to the following scope groups: - `issue`: Manage issues, PRs, comments, labels, and milestones. - `repository`: Manage repo contents and Actions runs/jobs/logs. - `organization`: List and read organization repositories. - `misc`: General miscellaneous API access. It does **not** have access to `admin`, `user`, `notification`, `package`, or `activitypub`.
Author
Owner

@pm i also like to sperate the .gitea/workflows/agent.yml in many agents like .gitea/workflows/agents/pm.yml or .gitea/workflows/agents/senior.yml but also a .gitea/workflows/agents/shared.yml so i can combine shared.yml with pm.yml

@pm i also like to sperate the .gitea/workflows/agent.yml in many agents like .gitea/workflows/agents/pm.yml or .gitea/workflows/agents/senior.yml but also a .gitea/workflows/agents/shared.yml so i can combine shared.yml with pm.yml

🤖 @pm

🤖 @pm

The maintainer has added a new requirement: refactoring the agent workflow configuration. Instead of a single .gitea/workflows/agent.yml, the configuration should be split into a modular structure (e.g., .gitea/workflows/agents/shared.yml for common logic and individual files like pm.yml and senior.yml for agent-specific roles).

This is a significant structural change that complements the previous goal of optimizing skill/context management.

Updated Plan:

  1. Workflow Refactoring:
    • Create .gitea/workflows/agents/ directory.
    • Extract shared configurations into shared.yml.
    • Create individual YAML files for each agent role (pm.yml, junior.yml, senior.yml, lead.yml, qa.yml).
    • Update the main orchestration to combine shared.yml with the specific agent files.
  2. Skill Prompt Optimization:
    • Research and modify the skill loading mechanism to provide "summaries" to non-possessing agents instead of full API definitions.
  3. Verification: Ensure all agents still function correctly with the new modular workflow and reduced prompt noise.

Since this involves both architectural changes to the CI/CD workflows and core prompt logic, @lead is the best choice to ensure the new modular structure is robust and the prompt optimization is correctly implemented.

@ffaerber ready to start building? reply yes to proceed.

🤖 **@pm** 🤖 **@pm** The maintainer has added a new requirement: refactoring the agent workflow configuration. Instead of a single `.gitea/workflows/agent.yml`, the configuration should be split into a modular structure (e.g., `.gitea/workflows/agents/shared.yml` for common logic and individual files like `pm.yml` and `senior.yml` for agent-specific roles). This is a significant structural change that complements the previous goal of optimizing skill/context management. **Updated Plan:** 1. **Workflow Refactoring:** - Create `.gitea/workflows/agents/` directory. - Extract shared configurations into `shared.yml`. - Create individual YAML files for each agent role (`pm.yml`, `junior.yml`, `senior.yml`, `lead.yml`, `qa.yml`). - Update the main orchestration to combine `shared.yml` with the specific agent files. 2. **Skill Prompt Optimization:** - Research and modify the skill loading mechanism to provide "summaries" to non-possessing agents instead of full API definitions. 3. **Verification:** Ensure all agents still function correctly with the new modular workflow and reduced prompt noise. Since this involves both architectural changes to the CI/CD workflows and core prompt logic, @lead is the best choice to ensure the new modular structure is robust and the prompt optimization is correctly implemented. @ffaerber ready to start building? reply yes to proceed.
Author
Owner

@pm also i see a lot ob scripting inside the yml is it posible to include the scripts as links to bashscripts? so tht its better readble

@pm also i see a lot ob scripting inside the yml is it posible to include the scripts as links to bashscripts? so tht its better readble

🤖 @pm

🤖 @pm

The maintainer wants to improve the readability of the Gitea workflow files by moving inline shell scripts into separate bash script files.

Updated Plan:

  1. Workflow Refactoring & Script Extraction:
    • Create .gitea/workflows/agents/ directory.
    • Extract shared configurations into shared.yml.
    • Create individual YAML files for each agent role (pm.yml, junior.yml, etc.).
    • Identify all complex inline scripts in the YAMLs and move them into dedicated bash scripts (e.g., in a new .gitea/workflows/scripts/ directory).
    • Update YAML files to call these scripts instead of using inline blocks.
  2. Skill Prompt Optimization:
    • Modify the skill loading mechanism to provide summaries to non-possessing agents instead of full API definitions.
  3. Verification:
    • Ensure the modular workflows and external scripts execute correctly and that prompt optimization is functional.

Because this now involves a significant overhaul of the workflow architecture, script externalization, and core prompt logic, @lead is the appropriate choice to handle this.

@ffaerber ready to start building? reply yes to proceed.

🤖 **@pm** 🤖 **@pm** The maintainer wants to improve the readability of the Gitea workflow files by moving inline shell scripts into separate bash script files. **Updated Plan:** 1. **Workflow Refactoring & Script Extraction:** - Create `.gitea/workflows/agents/` directory. - Extract shared configurations into `shared.yml`. - Create individual YAML files for each agent role (`pm.yml`, `junior.yml`, etc.). - Identify all complex inline scripts in the YAMLs and move them into dedicated bash scripts (e.g., in a new `.gitea/workflows/scripts/` directory). - Update YAML files to call these scripts instead of using inline blocks. 2. **Skill Prompt Optimization:** - Modify the skill loading mechanism to provide summaries to non-possessing agents instead of full API definitions. 3. **Verification:** - Ensure the modular workflows and external scripts execute correctly and that prompt optimization is functional. Because this now involves a significant overhaul of the workflow architecture, script externalization, and core prompt logic, @lead is the appropriate choice to handle this. @ffaerber ready to start building? reply yes to proceed.
Author
Owner

@pm yes

@pm yes

🤖 @pm

🤖 @pm

The maintainer has approved the plan to refactor the agent workflows, externalize inline scripts, and optimize skill prompt delivery.

🤖 **@pm** 🤖 **@pm** The maintainer has approved the plan to refactor the agent workflows, externalize inline scripts, and optimize skill prompt delivery.
Author
Owner

@lead please proceed with issue #22 per the plan above (delegated by pm).

@lead please proceed with issue #22 per the plan above (delegated by pm).

🔨 @lead is on it — building on branch ai/issue-22. I'll open a PR when it's ready.

🔨 **@lead** is on it — building on branch [`ai/issue-22`](https://git.ffaerber.duckdns.org/ffaerber/agents/src/branch/ai/issue-22). I'll open a PR when it's ready.
Sign in to join this conversation.
2 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: gitea/agents#22