benchmark-opencode-modelslisted

# Benchmark OpenCode Models **Announce at start:** "I'm using the benchmark-opencode-models skill to deep-test which OpenCode models actually work." This codifies the manual model-testing process (ping sweep, then a real fake-feature dispatch, then independent verification) into a repeatable skill, and extends it: instead of one throwaway "add fibonacci" prompt, every model gets 5 prompts that simulate the shape of a real `superpowers:subagent-driven-development` implementer task — full context/acceptance-criteria/constraints/definition-of-done, not just a one-liner — across a feature direction, a bugfix direction (each at both a short and a detailed prompt length), and a dedicated TDD red-to-green prompt. This is the **deep, expensive** test — it dispatches a real OpenCode run per model per prompt (5 real dispatches per model). For a quick "is this model even reachable right now" check before committing to this, run `bridge:check-opencode-models` first. ## Step 1: Determine which models to test Ask the user (via AskUserQuestion if not already stated) which models to test, or default to: ```bash cat ~/.opencode-bridge/config.json # default_model + fallback_models opencode models # full catalog, to pick a provider/tier (litellm, openrouter free, opencode zen free, etc.) ``` Build a comma-separated model list. Keep the list to what's actually in question — this dispatches 5 real tasks per model, so testing 10 models means 50 real OpenCode runs (plus