Glm Review

GLM review adapter

Use the deterministic adapter at ${CLAUDE_PLUGIN_ROOT}/scripts/run_glm_review.py. Do not call Z.AI or claude -p directly from CLoClo.

Contract

Receive the same parameters as codex-review:

• review_type: spec, plan, or impl
• session_dir, output_file, and repository root
• spec_path, plan_path, base_ref, and commit_list when applicable
• maturity: spike, dev, or ship
• iterate: optional convergence-oriented pass

Run one foreground adapter process while the pipeline may run this skill and codex-review concurrently:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/run_glm_review.py" \
  --review-type "$review_type" \
  --cwd "$repo_root" \
  --output-file "$output_file" \
  --spec-path "$spec_path" \
  --plan-path "$plan_path" \
  --base-ref "$base_ref" \
  --commit-list "$commit_list" \
  --maturity "$maturity"

Add --iterate only when requested. Omit arguments that do not apply.

Result handling

• Exit 0 plus a non-empty review containing Verdict global means success.
• Exit 2 means GLM was unavailable or the result contract failed. Log the warning and continue with Codex; do not substitute a Claude review.
• Read the review from output_file. Runtime diagnostics live at ${output_file}.runtime.log.
• Record the actual engine. The runtime prefers glm-5.2 and may explicitly report a glm-4.7 fallback when Z.AI rate-limits the preferred model.

The adapter is the only component allowed to translate GLM output into the pipeline review-file contract. Preserve the consensus, severity, escalation, and retry rules from ../pipeline/references/review-chain.md.