Process Dedup Review

process-dedup-review

Use this skill when the input is grouped JSON listing suspected duplicate process rows, especially groups generated from identical exchange fingerprints.

Workflow

1. Create an output directory such as artifacts/<case_slug>/.
2. Freeze the grouped JSON input under inputs/.
3. Run the bundled wrapper:

node process-dedup-review/scripts/run-process-dedup-review.mjs \
  --input /abs/path/duplicate-groups.json \
  --out-dir /abs/path/artifacts/<case_slug> \
  --json

4. Read outputs/duplicate-groups.json and outputs/delete-plan.json.
5. Apply the decision rules in references/review-rules.md.
6. Write the review report and required summary/verification reports, and explicitly note any unresolved limitations in the review artifacts.

The wrapper delegates to the canonical CLI command:

tiangong-lca process dedup-review --input /abs/path/duplicate-groups.json --out-dir /abs/path/artifacts/<case_slug>

Canonical Input Contract

{
  "source_label": "duplicate-processes-export",
  "groups": [
    {
      "group_id": 1,
      "processes": [
        {
          "process_id": "proc-1",
          "version": "01.00.000",
          "name_en": "Example process",
          "name_zh": "示例过程",
          "sheet_exchange_rows": [
            {
              "flow_id": "flow-1",
              "direction": "Input",
              "mean_amount": "1",
              "resulting_amount": "1"
            }
          ]
        }
      ]
    }
  ]
}

If the source begins as a spreadsheet, convert it into grouped JSON before using this skill. The skill no longer ships a workbook parser, because the default form must remain a thin CLI wrapper.

Remote Enrichment

• If TIANGONG_LCA_API_BASE_URL, TIANGONG_LCA_SUPABASE_PUBLISHABLE_KEY, and TIANGONG_LCA_OAUTH_CLIENT_ID are available, run tiangong-lca auth status --json; the CLI then enriches each candidate with:
  • state_code
  • created_at
  • modified_at
  • remote exchange flow short descriptions
  • reference scan results within the authenticated user's accessible scope across processes and lifecyclemodels
• The CLI loads .env from the current working directory before reading TIANGONG_LCA_*.
• If status is login-required, stop and ask the human to run auth login in a trusted terminal. Never request a password, code, token, or legacy API key. Use a separate private session file per account; headless tokens may come only from an approved orchestrator and never enter argv or artifacts.
• If remote enrichment fails, continue with local grouped-JSON evidence and note the limitation in the review output.

Canonical Outputs

• inputs/dedup-input.manifest.json
• inputs/processes.remote-metadata.json when remote enrichment succeeds
• outputs/duplicate-groups.json
• outputs/delete-plan.json

Decision Boundary

• Treat a group as an exact duplicate only when the normalized exchange multiset is identical after ignoring exchange row order and exchange internal IDs.
• Keep/delete is a review recommendation, not an automatic remote delete.
• If you cannot verify downstream references, label the delete list as priority delete candidates rather than claiming the rows are globally safe to remove.
• If the upstream identity workflow returned block_duplicate, preserve that decision as the reason to reuse an existing row instead of creating another process.
• If any group requires semantic judgment beyond exact exchange equality, mark it manual_review in the handoff and stop autonomous delete planning for that group.
• Do not convert delete-plan.json into remote mutation without a separate CLI-backed write command, dry-run artifact, and post-write verification.

Required Handoff Artifacts

• Input artifact: grouped duplicate-process JSON.
• CLI output artifacts: inputs/dedup-input.manifest.json, optional inputs/processes.remote-metadata.json, outputs/duplicate-groups.json, and outputs/delete-plan.json.
• Review output: a concise report that states exact duplicate evidence, chosen keeper, reference-verification coverage, and whether each delete candidate is safe or only prioritized for manual review.

Load On Demand

• Read references/review-rules.md for the evidence hierarchy and keep/delete tie-break rules.