building-review

建筑施工图审查

Review architectural construction drawings, form new technical conclusions, and directly deliver a complete AI initial-review Word report. The AI runs every skill gate independently; human review belongs to a later external workflow and is neither requested nor represented here.

Boundary

• Review architectural drawings and architecture-owned fire safety, accessibility, energy, green building, waterproofing, civil defense, curtain wall, decoration, food-service, and similar specialties.
• Do not review administrative submissions, permits, qualifications, stamps, signatures, or registered-professional seals.
• Do not technically review structural, plumbing, electrical, or HVAC drawings. Record only architecture-owned facts or coordination issues.
• If opinions already exist and the task is only to organize evidence and Word delivery, switch to review-opinion-delivery without changing conclusions.
• If a request asks for both new review and delivery, complete this skill's ai_ready opinions and formal AI-initial DOCX first. Use review-opinion-delivery only later, when reviewer-authored additions already exist and need consolidation.

Read references/scope-boundary.md before accepting borderline or mixed work.

Core Workflow

1. Create or identify a review workspace. New workspaces use schema v1.6 with review_stage: ai_initial. For an existing project, pass --root <项目>\03_审图过程 explicitly to scripts/create_review_workspace.py. Schemas v1.1-v1.5 remain readable compatibility formats; do not migrate an existing workspace in place.
2. Inventory every sheet and extract facts before judging. Complete drawing_inventory.csv and fact_ledger.csv; unresolved identity or version facts stay needs_review.
3. Read references/project-profile.md; the AI completes project_profile.json, records ai_review_completed, and closes specialty routing with AI初审完成. Route specialties with scripts/route_specialties.py --json; resolve uncertain routes from project facts and keep unsupported routes open rather than asking for a human confirmation gate.
4. Read references/review-rule-schema.md. Run scripts/generate_project_checklist.py <workspace> after classifying sheets. It expands the snapshotted rule catalog into unreviewed atomic checks plus v1.6 applicability and graphic-evidence placeholders; it never supplies compliance conclusions. For a residential single-building profile, ensure that residential_core_v1 is present in full; missing drawings do not waive a packet rule and the package gate will block completion.
5. Review every atomic check from drawings and use all four local knowledge layers: A_审查要点 for discovery, B_核心规范 for formal citations, C_疑难解析 for applicability and interpretation, and D_案例与截图 for auxiliary comparison only. Record each layer's sources and use in knowledge_snapshot.layers; D never replaces drawing facts or B. Resolve every rule condition in applicability_decisions.csv. For dimension, direction, symbol, absence, location, detail, or cross-sheet claims, close graphic_evidence_chain.csv with the required drawing roles, exact location, observed fact, interpretation, source quality, screenshot, and—when ambiguity-sensitive—an alternative interpretation plus its elimination basis. Record comparison, calculations, conclusion, and completion_gate=AI初审完成. Use scripts/review_calculations.py for supported numeric comparisons; it never supplies missing inputs or an applicable limit.
6. Perform technical compliance, design-depth/internal-consistency, and supported-optimization discovery separately. Run scripts/check_cross_sheet_consistency.py <workspace> after recording repeated facts; resolve every reported conflict from the drawings. Keep high-risk uncertainty open; do not default it to compliant or discard it.
7. Build issue_candidates.csv only from resolved noncompliant checks. Mark a deliverable issue ai_ready only after every AI validation field passes with gate_origin=agent and stage_completion=AI初审完成. v1.6 does not create independent_review_log.csv, record reviewer names, request human confirmation, or perform human adjudication. Decide screenshot strategy and record the exact evidence point, red-box target, and necessary context in the process ledgers.
8. Snapshot sources, the rule catalog, all four knowledge-layer use records, and every standard used by ai_ready issues or resolved technical checks with scripts/snapshot_review_integrity.py.
9. Run scripts/audit_review_completeness.py <workspace>. Fix every open item. Run scripts/validate_review_package.py <workspace> only after the audit passes; any later ledger change makes the audit stale.
10. Generate Word with scripts/generate_review_report.py, validate it with scripts/validate_docx_content.py, render every page, inspect every rendered page, record QA, then copy only the final DOCX to the user-selected deliverables directory. The filename and one-line title are 【AI初审】{项目名称}建筑施工图审查意见.docx for a single building and 【AI初审】{项目名称}建筑总图施工图审查意见.docx for a site review.