Plan SDD corrections

1. Fix one tasks/prd-[slug]/codereview_[num]/ by argument or context. When tasks/prd-[slug]/context-snapshot.md exists, load it through the Load branch of .agents/skills/sdd-snapshot/SKILL.md first. Search only for missing values; multiple candidates require a choice. Require and read its codereview.md once; preserve it. Output: exact, readable report, without mixing IDs between reports.
2. Classify every item as actionable, informational, or pending. In APPROVED, plan only what was requested; in APPROVED WITH RESERVATIONS, improvements require authorized scope; in REJECTED, cover violations, incompleteness, and failures. An unknown status allows only explicitly actionable findings. Preserve CR-NN; in legacy reports without IDs, assign a local ID by order and section. Inventory root and done/ metadata to reuse tasks; finding identity is report path + ID. Output: every item assigned; no duplicate finding or invented decision.
3. Check evidence in the smallest necessary code or TechSpec section; send read-only explorers in parallel only when several findings each need a sweep across many files. Group only the same cause with a reviewable result. Model an acyclic DAG, limits, files, and tests; number new tasks after the largest number in the root and done/. For desktop C#/.NET, omit E2E; preserve the obligation with relevant unit, integration, or manual acceptance. A missing environment or decision becomes an explicit pending item, not a discarded finding. Output: every actionable finding has a new or existing task, acceptance, and verification; no orphan task.
4. Read references/TEMPLATE_TASK.md in full when writing. Create reviewable task_[num].md files with at least two digits in the report folder, without reusing numbers or overwriting existing tasks. Reference sources by ID and section; preserve handoffs. Do not create a correction manifest: dependencies belong in tasks. Output: contracts on disk, report intact, and no code changes.
5. Check coverage, traceability, DAG, atomicity, commands, and idempotency. Report files, reuse, pending items, and impact. Return to the caller; request execution approval only where scope is not already authorized. In standalone use, run the session pause from that reference with sdd-execute-corrections as next step. Output: concrete plan for HIL or already authorized execution; contradictions linked to affected tasks.

A task in done/ whose finding persists is an incomplete correction: keep history and return to the caller to create work in the new report. An unreadable report, conflicting numbering, or cause without evidence blocks only the dependent planning.