slash-command-bridge

Lets the model invoke a curated set of slash commands via the run_slash_command tool. Instead of the user manually running /show-system-prompt or /compact, the model can call these commands itself when it determines they’re needed.

Available commands

Command	Description
show-system-prompt	Returns the full system prompt text
context	Returns context window token usage breakdown (or no-data when usage is unknown)
compact	Triggers session compaction to free context space

Usage

The model calls the tool automatically when appropriate:

run_slash_command({ command: "context" })
→ "Context Usage: 45,000 / 200,000 tokens (22.5%)"

run_slash_command({ command: "compact" })
→ "Session compaction will begin after this response completes."

If usage is unknown (tokens: null), context returns the same no-data error path as /context for parity.

Model-invoked compact is deferred until the assistant finishes its current post-tool response. Once compaction actually completes, the bridge schedules a hidden continuation turn only if the session is still idle, which avoids the old dead-end where /compact returned successfully but no compaction or resume happened.

Design

This is Phase 1 of the slash command invocation feature. The approach is pragmatic: commands whose logic is available via ExtensionContext are executed directly (system prompt access, context usage, compaction).

Commands requiring ExtensionCommandContext (session management like /clear, /new, /fork) cannot be invoked by the model yet — the framework doesn’t expose command handler access to tools. Phase 2 will add an allowModelInvocation flag to command registration and a framework-level bridge.

Context injection

The extension injects a hidden message before each agent turn listing the commands the model can invoke via the tool. This ensures the model knows the capability exists without polluting visible conversation.

Extension metadata

Field	Value
Category	tool
Enhances	clear, context-usage, show-system-prompt