sdd-code-review

Implementation Code Review

Resources

Before opening shared/..., follow symlinks in this loaded file's path, derive <plugin-root> from <plugin-root>/skills/<name>/SKILL.md, and read shared/agent-runtime.md, shared/path-resolution.md, shared/vcs-detection.md, shared/completion-evidence.md, and shared/review-lanes.md in place. Never copy or symlink plugin material into a working directory, target repository, or planning root.

Review Lanes

Use four independent lenses. When fresh non-inheriting contexts are available, dispatch all four in parallel; otherwise run affected lanes serially and record the accurate mixed or single-agent review mode. Do not claim independent corroboration for serial lanes.

Lens	Dispatch identifier	Inputs	Purpose
Plan drift	review_plan_drift	Diff, plan, phase, prior debriefs	Missing work, scope creep, approach drift
Quality	review_quality	Diff and code only	Correctness, safety, maintainability, tests, needless complexity
Spec compliance	review_spec_compliance	Diff, specs, designs	Requirements coverage and contract violations
Blind spots	review_blind_spots	Diff and changed-code context only	Adversarial edge cases, production failures, security, concurrency

Do not pass plan, spec, or design material to the quality or blind-spot lanes. Use exactly these runtime-neutral identifiers when the runtime exposes a task name or description field; do not request an agent or model.

Synthesis Rules

• Only independently run lanes can corroborate findings; serial lanes do not corroborate, including in mixed or single-agent review mode.
• Surface lane disagreements with their evidence rather than silently resolving them in synthesis.
• Label a finding raised only by the blind-spots lane as blind-spot-only.
• Do not manufacture primary-agent findings in synthesis; consolidate only findings actually reported by a lane.

Process

1. Identify the active plan, phase, target repository, and concrete diff range. A phase-completion gate requires a frozen durable native SCM revision/range; dirty or no-SCM work is not eligible. If synthesis or resolution relies on a decision, run sdd decide current --plan <Name> for the plan's standing decisions. Do not pass that context to any of the four intent-isolated lanes. A resolution that binds work beyond this artifact follows shared/decision-log.md: show the exact statement, then run sdd decide add --plan <Name> --statement "..." once approved.
2. Check completion evidence. Missing, pending, vague, failing, or source-identity-mismatched evidence is a plan-drift finding, not proof.
3. Render and run all four lanes, preserving their input isolation. Consolidate only actual findings and give Aligned, Needs changes, Blocked, or No reviewable diff with actual verification results. For a graph plan's review-role node, every finding left open carries action: revise (with nodes and a revise: block) or action: extend (with a node: fragment), and the verdict is Aligned only when every finding is terminal, else Amend — the frozen findings report sdd graph amend applies (shared/review-artifacts.md).
4. Persist the review using shared/templates/review.md. A phase gate is driven through the binary: `sdd review scaffold --frozen
5. Validate the persisted gate with sdd validate --scope Plans/<PlanName> --format json before handing the result back to sdd-implement, and cite the gate from phase evidence as - Final aligned review: <artifact path>; frozen: <exact rev>. The Git target adapter uses a full immutable <base>..<endpoint> range whose endpoint is the clean phase checkpoint. The planning Git adapter loads phase and plan content at reviewed_planning_revision; only lifecycle-only changes may follow review. Write the review in-flow. At phase close, it belongs unchanged at planning HEAD in the planning-root atomic commit with the debrief, evidence/status, and plan phase-array update. Record every affected plan decisions file in that same planning-root boundary commit (shared/autonomy.md § SCM boundary cadence). In Git, commit only in commit-capable workflows where commits are authorized. Unsupported planning or target SCM adapters keep the phase non-complete.
6. Material findings create new planned tasks. Implement each as a complete, reviewed native revision, then freeze and rerun all four lanes.