Chapter 9: Slash Command System¶
Slash commands are the human-facing command grammar of the harness. In Python, they are often the simplest place to keep structured user intent out of freeform chat.
Why this system exists¶
Once users rely on repeatable workflows, they need more than conversational luck. Slash commands give the harness a discoverable, structured entry surface.
Shared architecture¶
- commands map user intent to runtime actions
- registry and discovery should be centralized
- command execution should pass through the same safety and state boundaries as chat-originated work
Python implementation¶
Use one registry that can dispatch:
- local built-ins like
/help - workflow commands like
/resume - higher-level automation commands that still enter the query loop or tool executor
Keep command handlers small and explicit. They should call runtime systems, not replace them.
OpenAI Responses API mapping¶
Commands are usually a runtime feature, not a Responses API feature. But they often end in one of two actions:
- mutate local runtime state
- call the model through
responses.create
That means command outputs should already be shaped for the same loop and transcript system used by freeform requests.
Failure modes and tradeoffs¶
- command logic bypasses core runtime state
- hidden commands with no discoverability
- command handlers that duplicate query-loop logic
- commands becoming a second permission model
Build-it-yourself checklist¶
- define one command registry
- keep handlers small
- route command effects through existing runtime systems
- make command discovery easy
- avoid a second parallel architecture just for commands
Reference provenance¶
- Open this chapter inside the full blueprint
- the source discovery here centered on command registry patterns and skill loading