Convene a multi-lens GLM review panel over a spec or plan and synthesize one report. You (the main model) orchestrate — GLM agents cannot fan out themselves.
-
Pick the reviewer. A spec/*-design.md → glm-review-spec. A plan/*-plan.md → glm-review-plan. (For code or an implementation-vs-spec check, use glm-review-code / glm-review-implementation with the same lens structure.)
-
Preflight: is the proxy up? Run:
bash "${CLAUDE_PLUGIN_ROOT}/hooks/proxy-ready.sh"
If it exits non-zero, STOP and tell the user "cc-proxy is not running — start it, then re-run the panel." Do not dispatch agents against a down proxy.
(If ${CLAUDE_PLUGIN_ROOT} is unset in your shell, the script is at <plugin-root>/hooks/proxy-ready.sh — the plugin root is two directories above this SKILL.md file. A "No such file or directory" error means the variable was unset, NOT that the proxy is down.)
-
Dispatch N reviewers in parallel, one Agent call per lens, in a single message. Each call uses the matching reviewer agent and a prompt of the form:
Review ONLY through lens B — contradictions & feasibility. Ignore the other dimensions your instructions mention; another reviewer covers them. Target: <artifact path>. [paste or point to the artifact]. Check every load-bearing claim against the actual code and cite path:line.
Use the lens A / B / C definitions above verbatim so the three passes are genuinely distinct. (If a reviewer ignores the narrowing and returns its full default review, that is acceptable — worst case is overlap the synthesis dedups — but prefer the narrowed prompt.)
-
Synthesize (you, the strong model). Collect the three reports. For each distinct finding, score 0–100 for "is this real": 0 = false positive · 25 = maybe · 50 = real but minor · 75 = real and important · 100 = certain. Drop everything below 50. Dedup overlapping findings across lenses. Group the survivors:
- must-resolve (≥ 75)
- should-clarify (60–74)
- consider (50–59)
-
Clarify (interactive). The should-clarify (60–74) findings ARE the open questions — they are the ambiguous/underspecified items by definition. (must-resolve are contradictions/bugs to fix, not "what do you want?" choices; consider are minor. Neither is asked.)
- If the
should-clarify bucket is empty, skip this step entirely — go to step 6.
- Otherwise turn each
should-clarify finding into an AskUserQuestion item: the finding's open question as the question, plus 2–4 concrete options derived from the finding's own suggested directions (the tool always adds "Other"). Carry the finding's lens in the question text.
- Cap: ≤4 questions per round.
AskUserQuestion allows at most 4. Ask the top 4 by score in one call. If more remain, do not chain rounds — note the leftover count in the step-6 report. One round, not an interrogation.
- Collect the answers; do NOT write them to the artifact yet — that happens in step 7, after the run report exists (ordering matters, see step 7).
-
Write the run report (this file IS the panel-ran marker — its presence means the panel ran). Create <artifact-dir>/.review-panel/<artifact-basename>.md (a .review-panel/ directory next to the reviewed artifact):
# Panel run — <artifact-basename> (<YYYY-MM-DD HH:MM>)
- **artifact:** <path> **reviewer:** glm-review-spec **N:** 3
- **lenses:** A ambiguity · B contradictions/feasibility · C testability
- **per-lens:** A → 4 findings · B → 2 · C → 3 (tokens: A ~Xk · B ~Yk · C ~Zk)
- **buckets:** must-resolve 1 · should-clarify 3 · consider 2 · dropped <50: 4
- **asked:** 3 should-clarify → answers in the artifact's Clarifications section
- **verdict:** <one line>
## Findings
### must-resolve
- [score] <finding> (lens B)
### should-clarify (→ asked)
- [score] <finding> → **A:** <answer>
### consider
- [score] <finding>
Per-agent token counts come from each Agent result's <usage>subagent_tokens</usage> block (the main agent receives it). Record them if present; omit the tokens line if not — never fabricate.
-
Append the clarifications to the reviewed artifact (its own path, at end of file). Append-only — never rewrite the body:
## Clarifications (YYYY-MM-DD)
- **Q (lens A):** <question> → **A:** <user's answer>
This step runs AFTER step 6 on purpose: this Edit re-fires the PostToolUse hook, and the run report you just wrote is the marker that tells the hook the panel already ran. If the hook still emits a suggestion here, ignore it — you ARE the panel; do not convene another one on this artifact this round.
-
Report to the user: the must-resolve + consider findings (and any un-asked should-clarify, with the leftover count), then a one-line pointer — "N clarifications recorded → <artifact>; full run report → <marker path>". Note this was a GLM first pass — confirm before acting.
