OMP — Oh My Pi Agent Harness
Oh My Pi (OMP) is the coding harness that AI agents operate within on this server. It provides structured access to the filesystem, tools, sub-agents, and external services.
Core Tools
| Tool | Purpose | When to use |
|---|---|---|
read | Read files, dirs, archives, URLs, SQLite, images, docs | Always for file inspection |
write | Create or overwrite files | New files |
edit | Surgical line-based edits with snapshot tags | Modifying existing files |
bash | Terminal: git, builds, tests, package managers | Compute, not content reads |
search | Regex search across files | Content lookup |
find | Fast glob-based file discovery | Structure mapping |
lsp | Language Server Protocol: definitions, refs, rename | Code intelligence |
ast_grep / ast_edit | Structural AST search and rewrite | Codemods |
browser | Headless Chromium via Puppeteer | JS-dependent pages, UI testing |
eval | Persistent Python/JS kernels | Quick compute, data processing |
task | Spawn parallel sub-agents | Decompose complex work |
irc | Message other agents | Coordination |
web_search | Search the web | External information |
debug | DAP debugger (gdb, lldb, debugpy, dlv) | Step-through debugging |
todo | Phased task management | Multi-step work tracking |
Agent Types
| Agent | Capabilities | Use case |
|---|---|---|
task | Full tools (read, write, bash, edit, etc.) | General multi-step work |
explore | READ-ONLY (no write/exec) | Codebase investigation |
plan | Architecture planning | Complex multi-file design |
oracle | Senior engineer consultation | Debugging, architecture |
designer | UI/UX specialist | Visual implementation |
reviewer | Code review | Quality/security analysis |
librarian | Library research | External API investigation |
quick_task | Minimal reasoning, full tools | Mechanical updates only |
Internal URIs
| URI | Resolves to |
|---|---|
skill://<name> | Skill instructions |
rule://<name> | Rule details |
agent://<id> | Agent output artifact |
artifact://<id> | Full output (long results) |
history://<id> | Agent transcript |
local://<name>.md | Shared content with sub-agents |
memory://root | Persistent memory |
mcp://<uri> | MCP resource |
issue://<N> | GitHub issue |
pr://<N> | GitHub PR |
omp:// | Harness documentation |
Key Conventions
- Snapshot tags: Every
readreturns a[PATH#TAG]header. Edits require the current tag — stale tags are rejected. - No shell for reads:
cat,grep,ls,findare intercepted. Useread,search,findtools instead. - Parallel by default: Independent work should be parallelized via
tasksub-agents. - Verify before yield: Non-trivial changes must be tested before declaring done.
TYPO3 MCP Integration
The test instance (test.freundeskreis.family) runs hn/typo3-mcp-server v0.2.0 exposing 8 tools:
| Tool | Purpose |
|---|---|
GetPageTree | Page hierarchy |
GetPage | Single page data |
ListTables | Available DB tables |
GetTableSchema | Table structure |
ReadTable | Query table data |
WriteTable | Create/update records (workspace-safe) |
Search | Full-text content search |
GetFlexFormSchema | FlexForm field definitions |
All writes go through DataHandler into a TYPO3 workspace — never direct to live tables.
For the full integration matrix showing how TYPO3 MCP connects with other tools, see Skills & Integrations.