Using Independent Review Synthesis

A skill that synthesizes multiple AI / human review results to support the merge decision. Inspired by Nolan Lawson's Triple Agent skill, but avoiding tool lock-in and restructured to fit River Review's artifact-driven philosophy — a "synthesis pattern".

• Skill: independent-review-synthesis
• Related epic: #911
• Status: Phase 1 (community / recommended: false). Phase 2 will extend the artifact contract; Phase 3 will add a CLI ensemble mode.

When to use

• You want to aggregate multiple AI review results (e.g. run Claude / Codex / Cursor / GitHub Copilot reviews in parallel and consolidate their output into one).
• You want to combine AI and human review (put the AI results in review-external and the author's self-review in review-self).
• You want to decide based on findings from past PRs (pass the history via findings-pool).

When not to use

• A typical single-reviewer PR review → the existing midstream skills are enough.
• You want to add a new review viewpoint (security / a11y / performance) → create a dedicated skill.
• You want to automate the merge decision by majority vote of review results → this skill does not use majority vote. Hard rule 1.

Inputs

Following the repository convention (.river/ etc.), pass the following artifacts:

artifact ID	Format	Role
diff	patch / diff	Required. The diff of the PR
review-self	Markdown	Optional. The implementer's self-review
review-external	Markdown	Optional. External AI / human review
findings-pool	JSON	Optional. Findings history aggregated from past Review Artifacts
fullFile	source	Optional. Improves accuracy in the verification step

For artifact definitions, see Artifact Input Contract.

Output

A fixed 6 sections:

1. Critical Issues
2. Major Issues
3. Minor Issues
4. Dismissed Findings (hallucination / duplicate)
5. Agent Agreement Summary (a list of which reviewer raised what — supplementary info)
6. Merge Recommendation (merge-ready / human-review / block)

Each finding uses the block form Finding: / Evidence: / Reviewers: / Severity: / ValidatedStatus: / Suggestion:.

Design principles

• No majority vote: Reviewers: is supplementary info. Severity / ValidatedStatus are decided by evidence quality.
• Hallucination guard: verify with grep whether each finding's Evidence: exists in the actual code. If not found, mark it dismissed-hallucination.
• Single-reviewer findings are acceptable: even a finding raised by a single reviewer is confirmed if it has evidence.
• No critical without evidence: the severity ceiling is major.
• No provider lock-in: do not write the names Claude / Codex / Cursor in the skill body.

Local evaluation

Verify behavior with promptfoo eval:

cd skills/midstream/independent-review-synthesis/eval
promptfoo eval

OPENAI_API_KEY / ANTHROPIC_API_KEY are required. In Phase 1, golden/ stays empty (hand-written goldens fall under this repository's "posture, not progress" anti-pattern). Save outputs generated from real LLM runs and proceed along the promotion path.

Phase 1 limitations

• The artifact contract reuses the existing review-self / review-external / findings-pool as-is.
• Provenance fields on findings[] (reviewer, agreement[], validatedStatus) are a Phase 2 schema extension.
• CLI shortcuts (--ensemble claude,codex,cursor, etc.) are Phase 3.
• Currently only skill body + fixtures + eval scaffolding exist.