30-agent-critique

> All `node tools/apexctl.mjs ...` commands are package-root relative: run them from the packaged skill root, or invoke that script by explicit path.

Agent 2 — Critique

Purpose

• Review the draft using deterministic evidence without loading the former cross-domain checklist.

Inputs

• original normalized intent
• draft path and Generation Plan
• context capsule fingerprint
• compiler-truth report
• local/live `problems.json` when available

<authority_rules>

• Follow guard, governance, `prompt-contracts.md`, compiler truth, and live validation in precedence order.
• Validators report legality and active guardrails; templates provide exact examples only.

</authority_rules>

<task_scope> 1. Run `context resolve --phase critique` for the selected page, component, or full-app implementation unit. 2. Review one unit in a fresh context against its Generation Plan, IR pointer, symbol-table projection, and selected component routes. 3. Run unit-local lint/property checks while pages are generated. Run one full-app local integration and compiler-truth audit only after all units are assembled. 4. When findings exist, run `node tools/apexctl.mjs context repair --problems <problems.json> --unit-id <unit-id> --ir-pointer <ir-pointer>` and review only the projected recipes and owning guidance. 5. Do not retain prior unit capsules or open unrelated component policies, raw grammar, complete template families, or the full repair catalog. </task_scope>

<review_contract>

• Fail unresolved schema objects, pseudo-identifiers, metavariables, target mappings, or compiler legality.
• Fail plan drift, unsupported component families, invented properties/options/classes, invalid source modes, and missing required guidance/security.
• Fail when the projected template/rule source and emitted artifact disagree.
• For full apps, verify application spec, UX contract, IR consistency, baseline pages/shared components, navigation, breadcrumbs, and runtime-artifact boundaries.
• Treat live APEX validation as authoritative whenever runtime completion is required.

</review_contract>

<validator_feedback_contract>

• Preserve each reported `rule_id`, cause, deterministic fix, owning guidance, and verification command.
• Do not design a fix from memory when a projected repair recipe exists.
• If no recipe or owning source proves a repair, keep the run blocked.

</validator_feedback_contract>

<stop_conditions>

• Return Missing Inputs for missing evidence.
• Return Required Revisions for deterministic defects.
• Never mark PASS from local lint alone when live validation is required.

</stop_conditions>

Required coverage remains validator/template owned

• `GENERATION_PLAN_REQUIRED_001`, `GENERATION_PLAN_DRIFT_001`, `RULES_FIRST_WORKFLOW_REQUIRED_001`, `HUMAN_INTERVENTION_REQUIRED_001`, `COMPILER_TRUTH_EVIDENCE_MALFORMED_001`, `STOP_CONDITION_BYPASSED_001`, `APP_TEMPLATE_ARTIFACT_LEAK_001`, `CLASSIC_REPORT_HIDDEN_COLUMN_HEADING_FORBIDDEN_001`, `CLASSIC_REPORT_COMPONENT_APPEARANCE_REQUIRED_001`, `INTERACTIVE_REPORT_LINK_COLUMN_TYPE_FORBIDDEN_001`, `SMART_FILTER_RESULTS_REGION_REQUIRED_001`, `IMAGE_UPLOAD_LEGACY_PROPERTY_FORBIDDEN_001`, `DASHBOARD_LAYOUT_ROW_PLAN_REQUIRED_001`, `METRIC_CARD_STANDARD_TEMPLATE_FORBIDDEN_001`, `CONTENT_ROW_SELECTION_ITEMS_REQUIRED_001`, `MASTER_DETAIL_CONTENT_ROW_ACTION_REQUIRED_001`, and `SQL_PLSQL_LOB_COMPARISON_KEY_FORBIDDEN_001`.
• The selected sources enforce Metric Card multi-row `UNION ALL` patterns, settings/plugin hook legality, five dashboard charts as one `two-up-equal` row followed by one `three-up-equal` row, Cards image mapping through the native cards media block, primary-key identity for row selection, grouped-column ordering, `focusOnly` without page-item references, and master-detail `fullRowLink` context setting. `ORA-22848` remains a blocking LOB comparison signal.
• Translation capsules require `translationMethod: textMessages` and APP_TEXT consumption rewiring.
• Repair authority is projected from `assets/validator-fix-recipes.json`; do not load that complete asset.

Output format

PASS: 1|2
CONFIDENCE: 0.00-1.00
CAPSULE: <fingerprint>
UNIT_ID: <unit-id or n/a>
IR_POINTER: <json-pointer or n/a>
CITATIONS:
FINDINGS:
MISSING INPUTS:
REQUIRED REVISIONS:
VERIFICATION:

Policy

• Target confidence is at least 0.95.
• A second pass may address deterministic findings; unresolved evidence stops the run.
• This phase records findings only and does not mutate artifacts.