consult-zai

Ask your code question and invoke this Skill to dispatch it to both z.ai GLM 5.2 and Claude's code-searcher in parallel. You receive both raw responses plus a comparison table, agreement level, and a synthesized recommendation.

Both agents receive the same enhanced prompt requiring file:line citations and structured findings. The Skill compares their file paths, line numbers, code snippets, and unique findings, then buckets each finding as corroborated, solo, or disputed.

You need jq installed for parsing z.ai JSON output (a hard dependency) and a zai shell function configured in your zshrc or bashrc pointing at the z.ai endpoint. GNU timeout via coreutils is recommended to bound hung dispatches.

The Skill still presents the successful agent's response and labels the report a degraded single-AI run with no cross-comparison. On z.ai failure it quotes the stderr log tail for diagnostics before cleaning up temp files.

No. The z.ai dispatch enforces read-only access with allowedTools restricted to Read, Grep, and Glob, plus an explicit disallowedTools list blocking Bash, Edit, and Write. The orchestrator is also instructed not to mutate the reviewed tree while agents run.

Skip it for simple syntax questions, basic file lookups, or straightforward documentation queries where a single AI suffices. The Skill itself notes a direct read or lighter path is cheaper for low-complexity questions.