OpenCode pre-PR review

Drives a headless multi-model reviewer committee against the local diff and returns the chair's consolidated report so you can fix issues before a PR is opened.

Self-contained: the skill ships everything. It drives opencode by model (opencode run --pure --model <id>), embedding each reviewer's persona from prompts/*.md into the message — it does not rely on any agent being defined in the user's opencode.jsonc. (This also sidesteps an opencode footgun: a mode:subagent agent invoked as a top-level --agent run self-replicates and fork-bombs; driving by --model avoids agents entirely.)

You (Claude Code) are the lead/orchestrator. The script fans out to two members in parallel, then hands both reports to the chair, then runs an optional fact-check pass:

• SWE — correctness/bugs/security (prompts/swe.md, default model Muse Spark Contributor)
• Architect — design/architecture (prompts/architect.md, default model GLM Flash)
• Chair — dedupe + verify + fill gaps → final report (prompts/chair.md, default model Qwen Flash)
• Fact-check (optional, phase 3) — prunes only findings the diff can directly falsify (prompts/factcheck.md, default model DeepSeek V4 Flash)