Skip to main content
Dynamic skill injection architecture

​
Resolution pipeline

On every chat start, the server calls skill_manager.get_user_skills(email). The result resolves through three fallback levels:
  1. Memory cache — 60-second TTL. Same (email) hit twice in a minute = dict lookup.
  2. Settings Wrapper API — if MCP_TOKENS_URL is set, fetches the user’s enabled skills.
  3. Disk cache — if the API is down but we cached last known good, reuse it.
  4. Hardcoded defaults — all 13 public skills.
Without the Settings Wrapper, step 4 always wins — everyone gets the full public set.

​
System prompt injection

The skill list becomes an <available_skills> XML block injected into the system prompt: 
<available_skills>
  <skill>
    <name>docx</name>
    <description>Create and edit Word documents with tracked changes</description>
    <path>/mnt/skills/public/docx/</path>
  </skill>
  <!-- ... -->
</available_skills>
The model reads this once per chat and then consults SKILL.md files on demand.

​
User skill mounts

User-uploaded skills (ZIPs from the Settings Wrapper) are:
  1. Downloaded and extracted atomically: unzip to temp, rename to final.
  2. Cached under /data/skills-cache/{name}/ on the host.
  3. Tracked in .manifest.json (name → SHA-256 + timestamp).
  4. Bind-mounted into the sandbox at /mnt/skills/user/{name}/ — read-only.
Stale cache reuse: if the API or download fails but old cache exists, use it. Docker daemon needs host paths for bind mounts — configured via SKILLS_CACHE_HOST_PATH.

​
Core functions (skill_manager.py)

FunctionPurpose
get_user_skills(email)Async — full resolution with all fallbacks
get_user_skills_sync(email)Sync — reads cache only (called from the Docker-creation thread)
ensure_skill_cached(skill)Download + extract user-uploaded ZIP
build_available_skills_xml(skills)<available_skills> for the system prompt
build_sub_agent_skills_text(skills)Skill list for the sub-agent prompt
get_skill_mounts(skills)Dict of bind mounts for user skills

​
HTTP endpoints

EndpointReturns
GET /system-prompt?user_email=X&chat_id=YRendered per-user prompt
GET /skill-mounts?user_email=XBind-mount descriptors for user skills
GET /skill-list?user_email=XFormatted skill list for sub-agent prompts

​
Configuration

VarDefaultPurpose
MCP_TOKENS_URL(empty)Settings Wrapper URL
MCP_TOKENS_API_KEY(empty)Internal API key
SKILLS_CACHE_DIR/data/skills-cacheIn-container cache path
SKILLS_CACHE_HOST_PATH/tmp/skills-cacheHost path for Docker mounts

​
Verify

# System prompt should list skills
curl "http://localhost:8081/system-prompt?user_email=admin@open-computer-use.dev&chat_id=test" | head -50

# Bind mounts
curl "http://localhost:8081/skill-mounts?user_email=admin@open-computer-use.dev"

# Skill list
curl "http://localhost:8081/skill-list?user_email=admin@open-computer-use.dev"