The review panel

A panel is the set of enabled reviewers in your config.json. Each reviewer reads the same plan independently, produces its own review, and those reviews are merged by the synthesizer into a single consolidated output.

Why more than one reviewer

One LLM reviewing its own family’s output is not a review. It tends to agree, miss the same blind spots, and pattern-match to the same training biases. A single Claude reviewer giving a ship verdict tells you Claude thinks the plan is shippable; it doesn’t tell you whether the plan is actually sound.

The value of a panel comes from genuine disagreement between models that were trained differently. When three independent reviewers from different families agree on a finding, that finding is real. When they disagree, the synthesizer flags it as contested and you (the human) decide.

Composition advice

• Mix model families, not wrappers. Two Claude-based tools and one Gemini gives you one Claude review plus one Gemini review, not three reviews.
• At least three reviewers to get usable consensus signal. Two reviewers can only agree or disagree, with no tiebreaker.
• Include a local model (OpenCode + Ollama, Aider + Ollama) if you can. It’s free, runs offline, and produces genuinely different reviews than hosted models.
• Avoid stacking paid reviewers for routine runs. Reserve the full panel for plans that warrant the cost.

Failure isolation

One reviewer crashing, timing out, or returning unparseable output never aborts the run. The failed reviewer is recorded in manifest.json and excluded from synthesis; the others continue. See How synthesis works for what counts as “usable.”