MetalHexx/rad-code-review

Code Review

Three modes. Identify yours from the spawn context fields you received, then follow the matching workflow file end-to-end. Each mode's workflow is fully self-contained — do not load any other review doc or cross-reference between modes.

| Your context includes… | Mode | Scope | Status Enum | Workflow | Template | |--------------------------------------------------------------|-------|----------------------------------------------------------------|---------------------------------|--------------------------------------------------------|--------------------------------------------------------| | task_number (and task_id, head_sha) | Task | The task's diff vs. its Task Handoff contract | on-track \| drift \| regression | task-review/workflow.md | task-review/template.md | | phase_first_sha (and phase_head_sha) | Phase | The phase's cumulative diff vs. its Phase Plan contract | on-track \| drift \| regression | phase-review/workflow.md | phase-review/template.md | | project_base_sha (and project_head_sha); no task / phase | Final | The project's cumulative diff vs. the Requirements doc | met \| missing | final-review/workflow.md | final-review/template.md |

Every mode writes a per-requirement audit table. Verdict enum is unchanged across all three: approved | changes_requested | rejected. Each workflow runs the conformance pass first, then a lean quality sweep; findings merge and highest severity wins.

Evidence-Not-Intent Charter

These rules apply to every review mode. Violations invalidate the review.

1. The diff is truth. The handoff is intent. Execution notes are not evidence. If a sentence in your review would read identically whether you had run the diff or not, delete it.
2. No copy-paste from the Task Handoff, Phase Plan, or prior reviews. Summaries of prescribed content are not review output. If the handoff prescribes file content verbatim, your conformance claim must show the comparison mechanism (diff command, byte-range reference), not restate the prescription.
3. Every finding carries evidence. Evidence is one of: a quoted code line with File:Line, a diff excerpt, captured test output, or a grep result. An audit row with status=on-track still requires evidence (the proof that the slice is correct).
4. Positive observations never cushion deferred or silent behavior. If the code silently does nothing where a requirement is deferred to a later task, that is a carry-forward note or quality finding — not a positive.
5. Verdicts cite the driving finding. An approved verdict cites "no findings ≥ low severity, all audit rows on-track". A changes_requested verdict names the finding ID(s) that drove it.
6. Ran it yourself. Test counts, build status, and diff stats come from commands you executed in this review session, not from any upstream report.

Finding-ID Scheme

Every finding in every finding-bearing table gets a stable F-N identifier, numbered sequentially per review doc starting at F-1. Conformance-pass findings (drift / regression rows in the audit table) and quality-sweep findings share the same F-N space. IDs reset per review document — a corrective review starts fresh at F-1. The orchestrator's corrective-playbook addendum keys its disposition table off these IDs; missing or duplicate IDs break the mediation contract.

Evidence Contract

Every row in every finding-bearing table (per-requirement audit, quality sweep) carries these fields:

• F-ID — stable finding identifier (see Finding-ID Scheme).
• File:Line — concrete source pointer (e.g., src/cli.js:42 or src/cli.js:42-58). Use — only when the finding genuinely spans no specific location (rare; most NFRs still have representative sites).
• Evidence — quoted code, diff excerpt, test output, or grep result that establishes the finding. Paraphrase is not evidence. Use — only for on-track rows where the evidence is the absence of drift across the entire listed file range; cite the range in File:Line in that case.

The per-mode templates enforce these columns. A review doc that lacks F-ID, File:Line, or Evidence in any finding row is structurally invalid.