feat: support LLM-free command execution (follow upstream #28292)

[Svtter]

Background

The /review:auto command toggles auto-review on/off. Currently it works by having an LLM agent parse the user's argument and call the toggle_auto_review tool. This means:

• Wastes an LLM call just to flip a boolean (unnecessary tokens + latency)
• Logic depends on the LLM's interpretation, which is fragile
• State is in-memory only, lost on restart

Desired behavior

The command.execute.before hook should be able to:

1. Parse the argument directly (on / off)
2. Update the state
3. Return a response to the user
4. Skip the LLM call entirely

Upstream dependencies

• anomalyco/opencode#28292: Allow plugins to intercept slash commands and return results directly (skip LLM)
• anomalyco/opencode#18554: Add cancellation support for command.execute.before hook
• PR anomalyco/opencode#27521: Adds noReply field to command schema

Temporary workaround (until upstream merges)

• Persist autoEnabled state to a file (.opencode/.review-state.json) so users can also edit it directly
• Pre-process the toggle logic in command.execute.before, so the LLM only echoes the result