feat: agentv invoke command to call a named target from a CLI script

[christso]

Problem

When a cli provider script needs to call an LLM (e.g. a RAG pipeline that does search then generation), there is no way to delegate the LLM call to a named target defined in targets.yaml. The script must make the HTTP call itself, which means:

• LLM credentials and endpoint details must be hardcoded in the script (or injected via env: on the target)
• Swapping the LLM backend (e.g. from OpenAI to Azure OpenAI, Claude, or Copilot) requires changing the script, not just targets.yaml
• The RAG/search layer is tightly coupled to the LLM layer

Proposed Solution

Add an agentv invoke sub-command that a CLI script can call to delegate the LLM step to a named target:

agentv invoke <target-name> --prompt-file <path> --output-file <path>

Example usage in a CLI script

// 1. Do search/retrieval, build prompt file
// 2. Delegate LLM call to the named target from targets.yaml
spawnSync('agentv', ['invoke', 'openai', '--prompt-file', builtPrompt, '--output-file', outputFile])

targets.yaml

targets:
  - name: my-rag-pipeline
    provider: cli
    command: node .agentv/scripts/my-pipeline.cjs {PROMPT_FILE} {OUTPUT_FILE}
    grader_target: openai

  - name: openai
    provider: openai
    endpoint: $OPENAI_ENDPOINT
    api_key: $OPENAI_API_KEY
    model: $OPENAI_MODEL

Swapping to Claude is then just changing openai to claude in the command, with no script changes.

Benefits

• No credentials in scripts at all
• LLM backend is fully configurable from targets.yaml
• Enables multi-turn pipelines where the script drives retrieval and agentv drives generation
• Consistent with how agentv already owns provider config for LLM targets

Workaround

Currently using env: injection on the cli target to pass credentials explicitly, but the env var names are still hardcoded in the script.