--- id: PX-SPEC-O6OLUJJO type: spec title: Peer Review Workflow --- ## Intent Make peer review a deep investigative task delegated through general primitives, not a dedicated review engine. A fresh reviewer challenges completed work without inheriting the author's full reasoning. ## Scope In: discoverable peer-review skill, deliberate diverse reviewer choice, investigative access, evidence-backed findings, targeted re-review. Out: new review extension, mandatory reviewer pool, maintained model-alias registry, perfect sandboxing, delivery authorization, automatic publication. Code, documents, configuration, research, and other deliverables are eligible; Git is optional. ## Reviewer choice Choose a reviewer for the task, not a fixed universal reviewer. Use deliberately expressed user preferences; credentials, environment keys, saved logins, and catalog availability do not establish preference. Choose a substantially different model from the author, preferably from another provider; a different endpoint serving the same model is not diversity. Report selected harness/provider/model and what is known about actual model identity without requiring exhaustive identity mappings. When no suitable preference is known, ask rather than silently selecting an authenticated provider. Existing Dilemma advisor preferences may inform choice when deliberately applicable; Dilemma remains an independent advice interface, not a required review dependency. Unavailable reviewers or transports produce an explicit blocker, not an undisclosed substitution. ## Behavior The author supplies intended outcome, constraints, work or revision, relevant initial evidence, and requested depth. The reviewer can expand investigation to callers, tests, dependencies, documentation, and other relevant context instead of being confined to author-picked excerpts. Use focused reproduction and actual rendering where needed and authorized; distinguish observed results from static inference and author-reported checks. Large files and multi-file work remain reviewable through targeted incremental inspection rather than fixed snapshot eligibility limits. The reviewer reports inspected scope, concrete findings, uncertainty, verification gaps, and any incomplete work. Each actionable finding identifies location or artifact, consequence, supporting evidence, and a correction or decisive verification need. Reading a file, matching a schema, or returning no findings does not prove adequate coverage or correctness. The executor evaluates findings and owns fixes; re-review checks revised evidence and relevant new effects rather than accepting claimed fixes. Completion, cancellation, and explicitly agreed bounds stop the task; do not recreate Relop's persisted convergence bureaucracy. Neither review outcome nor reviewer agreement authorizes delivery or replaces required tests. ## Interaction model Delegation and findings form an ordinary peer task; fixes remain with the executor. ```mermaid sequenceDiagram participant A as Author participant P as Peer A->>P: Outcome, constraints, work, review assignment P->>P: Inspect context and verify consequential claims P-->>A: Findings, evidence, scope, limitations opt Findings warrant correction A->>A: Evaluate and fix A->>P: Revised work and verification P-->>A: Rechecked findings and remaining gaps end ``` ## Independence The supported root installation makes the skill usable through PiPi without Relop, Hornets, Ask, or Dilemma. [Hornets](../PX-SPEC-WQR3XWPQ-hornets-cross-harness-delegation/index.md) is an optional alternative transport, not a prerequisite. Review procedure does not own harness lifecycle, tools, credentials, or provider-selection infrastructure. Fresh context and model diversity are separate claims; granting command execution is not a mechanical read-only guarantee. ## Acceptance criteria - [ ] A PiPi-only installation can conduct useful review of a multi-file change, including omitted callers and large-file context. - [ ] Reviewer selection follows explicit user preference, uses a substantially different model with cross-provider preference, and never treats credentials or aliases of the author's model as diversity. - [ ] A review traces relevant evidence beyond the author's supplied files and labels static, reproduced, visual, and author-reported evidence accurately. - [ ] Findings include location, impact, evidence, and actionable correction or verification; uncertainty and incomplete scope remain visible. - [ ] Re-review examines actual revised work and consequences; caller disagreement or claimed fixes alone do not erase concerns. - [ ] The skill works without Hornets, Ask, Dilemma, or Relop and does not duplicate their responsibilities. - [ ] Missing preferred reviewer, unavailable transport, cancellation, or insufficient investigation cannot be reported as completed review.