Agent CLI Wrappers Guidelines
This skill provides explicit instructions on how to run a prompt using a specific model via the available agent CLI wrappers. It serves as a tactical reference for orchestrating subagents, overriding models, and executing standalone tasks in the terminal without invoking the interactive TUI.
🛠️ Antigravity CLI (agy)
The Google Antigravity CLI allows you to specify models and run single-shot prompts autonomously.
Running a Prompt with a Specific Model
To execute a prompt non-interactively using a specific model, use the --model flag followed by the model alias, and the --print (or -p) flag for the prompt.
Syntax:
agy --model "<model-alias>" --print "<your-prompt-here>"
Example (Enforcing Gemini 3.5 Flash for file grepping tasks):
agy --model "gemini-3.5-flash" --print "Parse the server log and list any critical errors."
Running a Specific Agent
The Antigravity CLI does not support slash commands or standard at-mentions for invoking agents reliably in single-shot mode. To run a specific agent and bypass the default routing, you must be explicit in the prompt payload and instruct it to wrap the output so it can be extracted cleanly.
Syntax:
agy -p "Run <AgentName> agent. Only return the final agent output wrapped in <output>...</output> tags. Do not wrap anything else. <your-prompt-here>" | sed -n '/<output>/,/<\/output>/p' | grep -v 'output>'
Example:
agy -p "Run news-finder agent. Only return the final agent output wrapped in <output>...</output> tags. Do not wrap anything else." | sed -n '/<output>/,/<\/output>/p' | grep -v 'output>'
Listing Available Models:
To see all valid model aliases configured in the environment (models.json), run:
agy models
🛠️ OpenCode CLI (opencode)
OpenCode provides similar capabilities to run one-off tasks using specific providers and models.
Running a Prompt with a Specific Model
To execute a prompt using OpenCode, use the run subcommand along with the -m (or --model) flag. OpenCode models are specified in the format provider/model.
Syntax:
opencode run -m "<provider>/<model>" "<your-prompt-here>"
Example:
opencode run -m "anthropic/claude-3-sonnet-20240229" "Audit the active directory for security flaws."
Running a Specific Agent
OpenCode has built-in parameter support for dispatching prompts directly to a specific custom agent, bypassing the need for natural language routing.
Syntax:
opencode run --agent "<agent-name>" "<your-prompt-here>"
Listing Available Models: To see all valid models and providers configured for OpenCode, run:
opencode models
🧠 Constraints and Routing
- Inheritance vs. Constraints: When executing these wrappers to spawn subagents for specific operations (e.g., file reading, grepping), ensure you explicitly pass the required model. This effectively acts as setting
inherit = falsefor the subagent's context by overriding any inherited master model state. - Model Selection Strategy:
- Use fast, low-latency models (e.g.,
gemini-3.5-flash) for simple extraction, code grepping, or parsing tasks. - Use high-reasoning models for complex architectural reviews or orchestrations.
- Use fast, low-latency models (e.g.,