How synthesis works

After every enabled reviewer has produced an output, the synthesizer reads them all and emits a single consolidated review. The synthesizer is one of your configured reviewers (pointed at by synthesizer.reviewer_id), re-invoked with a different prompt. It can be a disabled reviewer, in which case it acts as a neutral arbiter that consolidates the panel without being one of its voices (see the config reference).

What the synthesizer does

The synthesis prompt (in config/synthesis.md) instructs the model to:

1. Collect distinct findings across all reviewers. Two reviewers saying the same thing in different words is one finding, not two.
2. Tag each finding as:
   • consensus — raised by 2+ reviewers with substantively the same point.
   • contested — raised by one and explicitly disputed by another. The synthesis includes a disagreement field summarizing the split.
   • singleton — raised by exactly one reviewer, not contradicted. Kept because lone reviewers are often right.
3. Pick a verdict (ship / iterate / reject) and explain it.
4. Write a short summary of the plan’s overall shape.

What the synthesizer does NOT do

• It does not silently pick a winner when reviewers disagree. Contested findings are surfaced as contested.
• It does not re-review the plan. If a finding wasn’t in any source review, the synthesizer is instructed not to add it.
• It does not invent agreement. If only one reviewer raised something, it stays a singleton.
• The orchestrator (the PHP code) does not do diffing or tagging itself. That’s the synthesizer model’s job, driven by the synthesis prompt.

What gets fed to the synthesizer

Only reviewers with status ok (parsed JSON) or unstructured (returned prose, used as-is) are passed to synthesis. Reviewers that timed out, exited non-zero, or produced empty stdout are recorded in manifest.json but excluded from the synthesis prompt so they don’t pollute it.

If no reviewer produced usable output, synthesis is skipped entirely; the run still persists the manifest so the failures are recorded.

Why this works

The synthesizer being another LLM rather than a hand-written diff algorithm means it can read prose, understand near-duplicates, and distinguish “X is risky because of A” from “X is risky because of B” even when both reviewers used the word “risky.” No structured-matching algorithm in PHP would catch that.

The flip side is that the synthesizer inherits the bias of whichever reviewer plays that role. Using Claude as the synthesizer of a Claude + OpenCode + Gemini panel tilts the consensus toward Claude’s framing. Worth keeping in mind when reading the output.