feat(agents): streamline code review across local changes, branch diffs, and pull requests #2783

The Code Review agent will provide one review workflow for two audiences: authors who want grounded feedback before opening a pull request, and reviewers who need to assess an existing pull request. Review target selection will be independent from review profile selection so local changes, branch diffs, and pull requests can share the same orientation, dispatch, evidence validation, and reporting contracts.

Before dispatch, the workflow will present a decision-ready view of the change areas, preliminary signals, evidence references, selected perspectives, omitted perspectives, and recommended depth. Findings will affect the verdict only after their evidence has been validated, and perspectives that return no output will receive one full-context retry before being reported as unavailable.

Acceptance Criteria

• A reviewer can explicitly select local changes, a branch diff, or a pull request as the review target.
• Review target selection remains independent from Standard, Full, and Custom review profile selection.
• Pull request input accepts a URL or number and resolves repository and base-branch context before review.
• Orientation groups the change into coherent areas and includes preliminary signals with concrete evidence references.
• One confirmation presents the proposed areas, perspective ownership, omitted perspectives with rationale, and recommended depth with rationale.
• Accessibility is selected only when the changed surfaces are user-facing or otherwise accessibility-sensitive.
• Findings with missing, out-of-scope, or contradictory evidence remain visible as rejected findings and do not affect finding counts or the final verdict.
• A perspective that produces no output is retried once with the full dispatch context before it is marked unavailable.
• The pull request review prompt delegates to the same Code Review workflow instead of maintaining a separate review implementation.
• Behavior evaluations cover target and profile selection, decision-ready confirmation, evidence rejection, and no-output retry behavior.
• Generated agent documentation and plugin metadata are synchronized with the final artifact set.

Out of Scope

• Reimplementing native GitHub or Azure DevOps line-comment posting; preserve the emission capability completed by feat(agents): add ADO and GitHub line-comment poster capabilities to code-review emission #2091
• Quantitative change-risk scoring
• A separate reviewer-coaching prompt

Related

• Foundation: Builds on feat(agents): orientation-first review integration (Scenario D loop) #2090, which established the orientation-first Scenario D review loop.
• Dependency: Preserves the native GitHub and Azure DevOps emission capability completed by feat(agents): add ADO and GitHub line-comment poster capabilities to code-review emission #2091; this issue changes target resolution and review setup, not the downstream posting contract.
• Documentation follow-up: Enables docs(agents): document orientation-first mode in user-facing docs and code-review skill #2092 to document one user-facing workflow across local changes, branch diffs, and pull requests. This issue will synchronize generated references but will not replace the broader getting-started documentation tracked there.
• Evaluation follow-up: Extends the behavior surface that feat(agents): add end-to-end review-loop behavior eval for orientation-first Scenario D #2093 should exercise. This issue will add focused regression scenarios but will not replace the end-to-end orientation, dispatch, walk-back, and emission evaluation tracked there.
• Coordination scope: Must reconcile its single decision-ready confirmation with the unresolved Scenario D interaction choices in feat(agents): resolve five open Scenario D UX decision points for orientation-first review loop #2094. Any overlapping choice should be recorded there rather than implemented as a competing interaction pattern.
• Complementary scope: Provides a reusable orientation and depth boundary for feat(prompts): Code Review Coach; guides a human through a large PR #2284's reviewer-coaching direction. Coaching behavior remains separate follow-up work and should converge on this workflow rather than fork review-target discovery.
• Future integration: Provides an evidence slot for feat(skills): change-risk model reference for code-review depth selection #2501's future change-risk model to inform depth recommendations. This issue uses qualitative rationale and does not implement quantitative risk scoring.

Evidence Source

• Stakeholder input: support both pre-PR author review and pull request review through a streamlined Code Review experience.
• Prototype evidence: branch feat/code-review-pr-changes demonstrates the intended agent, skill, prompt, evaluation, documentation, and packaging changes.