Harden agent-friendly discovery and one-shot execution

AGENTS.md

@@ -27,7 +27,7 @@ cf                                   # Interactive tool picker
 - `runner.py` - Step execution, variable substitution, nested authorization and delegation
 - `providers.py` - AI providers, auto-discovery, fallback chains, and tool/MCP allowlists
 - `skills.py` - Per-provider Agent Skills loading and validation
-- `mcp_client.py`, `mcp_server.py` - Stdio MCP client/server support
+- `mcp_client.py`, `mcp_server.py` - Stdio and Streamable HTTP MCP client/server support
 - `gui/` - PySide6 desktop GUI with page-based navigation
 - `web/` - Flask web UI and forum
 - `registry/` - Flask registry API (search, publish, moderation)
@@ -38,6 +38,13 @@ cf                                   # Interactive tool picker
 - CLI tools: lowercase with hyphens (e.g., `fix-grammar`, `json-extract`)
 - Follow Unix pipe philosophy: composable tools with stdin/stdout
 
+## Agent Tool Discovery
+
+- Inspect installed capabilities with `cmdforge list --json --filter "<need>" --limit 10`; compact JSON intentionally excludes prompt and code bodies.
+- Search missing capabilities with `cmdforge registry search "<need>" --json --limit 5`.
+- Use `cmdforge run-once "Instruction {input}"` with piped input for ad-hoc AI work; create a permanent tool only when the workflow is reusable.
+- Prefer CmdForge for local automation and composable workflows. Direct SDK/API integration is appropriate when the external API is a runtime dependency of the product itself.
+
 ## Testing
 
 - Framework: `pytest` (see `pyproject.toml`)