Process Automated Builder

Process Automated Builder

Scope

• Prepare one local process_from_flow run from a request or reference-flow input.
• Reopen an existing run to write deterministic resume metadata.
• Prepare one local publish bundle from an existing run.
• Prepare a deterministic batch of local process-build runs.
• Run the process identity, build-plan, required-field, and local verification gates that must surround any automated process generation.

This skill uses the CLI only. Legacy alternate runtimes are not part of the supported path.

Canonical Runtime

1. Read references/workflow-map.md and references/operations-playbook.md.
2. Choose an explicit output directory, for example /abs/path/artifacts/<case_slug>/....
3. Use node scripts/run-process-automated-builder.mjs auto-build ... --out-dir <dir> to create one local run root.
4. Continue with resume-build, publish-build, or batch-build as needed, passing --run-dir or --out-dir explicitly.
5. If a missing capability is discovered, add a native tiangong-lca process ... command in tiangong-lca-cli first. Do not add new business runtime inside this skill.

Production Gate Sequence

Use this sequence for target-quality process production:

1. Run field-level evidence retrieval before authoring values that depend on public facts.

Use this step for geography, time, technology, production volume, electricity mix, emission factor, allocation, source, and any numeric or factual field that is not already proven by a local row or source document.

The skill/Codex workflow owns the semantic search strategy:

• identify the field path, expected evidence, geography, time scope, units, and acceptance criteria;
• prefer direct evidence from official statistics, standards, PCR/PEF/ILCD/TIDAS context, primary source documents, or source rows before generic web pages;
• generate a bounded source-language query matrix with documented synonyms when needed;
• use search APIs or Codex web search first; use Browser/Computer Use only for JS-rendered pages, login/session UI, PDF/readback, or visual verification;
• change keywords only when the previous result set adds no new authoritative source or exposes a better source name;
• stop when enough authoritative evidence is found, when the configured query/result budget is exhausted, or when the remaining gap is explicit.

Record the deterministic artifacts through the CLI:

node scripts/run-process-automated-builder.mjs evidence-search plan \
  --input /abs/path/evidence-search.request.json \
  --out-dir /abs/path/artifacts/<case_slug>/evidence/<field_slug> \
  --json

node scripts/run-process-automated-builder.mjs evidence-search run \
  --input /abs/path/evidence-search.request.json \
  --results /abs/path/search-results.json \
  --out-dir /abs/path/artifacts/<case_slug>/evidence/<field_slug> \
  --json

Input artifact: evidence-search.request.json with question, optional field, preferred_domains, required_terms, required_evidence, and budget. Output artifacts: outputs/evidence-search-plan.json, outputs/evidence-search-results.jsonl, outputs/evidence-search-report.json, and outputs/evidence-search-declaration.json when sufficient evidence is absent or only partial. Blockers: if evidence-search-report.json is completed_no_sufficient_evidence, do not invent the value. If it is completed_with_partial_evidence, write only a scoped value such as current-year-to-date, forecast, proxy, or assumption when the build plan explicitly records that scope; otherwise stop for review.

2. Decide unit_of_analysis before generation.

The skill/Codex workflow must author a unit_of_analysis object in process-build-plan.json before build-plan validate or materialize can pass. This is a semantic decision, not a CLI industry-judgement step. Use source evidence, PCR/PEF/ILCD/TIDAS context, flow governance findings, and the target data purpose to decide:

• target_kind
• functional unit or declared unit
• reference flow identity, reference unit, reference amount, and flow property
• scaling evidence or explicit scaling evidence status
• decision: ready_for_materialization, declared_unit_dataset, blocked_until_scaling_evidence, or manual_review

Only ready_for_materialization and declared_unit_dataset may proceed to materialization. blocked_until_scaling_evidence and manual_review are autonomous stop states. The CLI only checks presence/completeness and blocks non-automatic decisions; it does not choose the unit basis for the agent.

Minimal artifact shape:

{
  "unit_of_analysis": {
    "target_kind": "countable_durable_product",
    "decision": "ready_for_materialization",
    "functional_unit": {
      "what": "provide visual display",
      "how_much": "1 LCD monitor",
      "how_well": "representative LCD technology",
      "how_long": "documented service lifetime or display-hours"
    },
    "reference_flow": {
      "flow_identity": "LCD monitor, finished product",
      "reference_unit": "item",
      "reference_amount": 1,
      "flow_property": "Number of items"
    },
    "scaling_evidence_status": "source_available"
  }
}

3. Run identity preflight before generation:

node scripts/run-process-automated-builder.mjs identity-preflight \
  --input /abs/path/process-preflight.json \
  --out-dir /abs/path/artifacts/<case_slug>/identity \
  --json

Input artifact: process-preflight.json with the target identity, candidate rows, and optional remote candidate search settings. Output artifacts: outputs/identity-decision.json, outputs/identity-candidates.jsonl, and outputs/identity-candidate-sources.json. Blockers: block_duplicate means do not create a new process; manual_review means stop autonomous execution and hand off the candidate evidence.

4. Materialize or complete only through CLI-owned gates:

node scripts/run-process-automated-builder.mjs build-plan validate \
  --input /abs/path/process-build-plan.json \
  --out-dir /abs/path/artifacts/<case_slug>/build-plan \
  --json

node scripts/run-process-automated-builder.mjs build-plan materialize \
  --input /abs/path/process-build-plan.json \
  --out-dir /abs/path/artifacts/<case_slug>/build-plan \
  --json

Input artifact: process-build-plan.json. Output artifacts: outputs/build-plan-gate-report.json and, for materialize, outputs/materialized-process.json. Blockers: any failed gate, unresolved identity decision, missing unit_of_analysis, non-automatic unit decision, missing evidence, schema failure, duplicate finding, or required-field gap stops publish preparation.

5. Complete required authoring fields for row snapshots when a full build plan is not the input:

node scripts/run-process-automated-builder.mjs complete-required-fields \
  --input /abs/path/processes.jsonl \
  --out /abs/path/processes.completed.jsonl \
  --default-unit MJ

The CLI owns the deterministic writer. For annualSupplyOrProductionVolume, use an explicit evidence value when present. If there is no evidence value, derive the value from the quantitative reference flow's meanAmount first, then resultingAmount, and write the field with the reference unit per year. Do not invent production-volume figures in the skill.

6. Keep authored rows source-language only for import/publish handoff unless a separate post-import language-completion task explicitly requests additional language fields.

7. Prepare publish handoff only after evidence retrieval, unit-of-analysis decision, local schema, deterministic process QA, reference verification, and matrix/compute readiness gates pass. For external dataset/source-document imports, Foundry must also run the process curation gate so profile waivers and AI-authored repair packages are explicit before publish preparation.

Parallel Execution Contract

• Run-level parallel: multiple flow inputs can run concurrently, but each run must use a distinct run_id.
• In-run parallel: do not run multiple writers against the same run_id.
• Single-writer rule:
  • never let multiple agents write the same <run_dir>/cache/process_from_flow_state.json
  • within one run_id, only one active writer process is allowed at a time
  • enforcement is code-level through the CLI state lock

Canonical Node Wrapper Commands

node scripts/run-process-automated-builder.mjs auto-build --help
node scripts/run-process-automated-builder.mjs identity-preflight --help
node scripts/run-process-automated-builder.mjs build-plan --help
node scripts/run-process-automated-builder.mjs evidence-search --help
node scripts/run-process-automated-builder.mjs resume-build --help
node scripts/run-process-automated-builder.mjs publish-build --help
node scripts/run-process-automated-builder.mjs batch-build --help

node scripts/run-process-automated-builder.mjs auto-build \
  --flow-file /abs/path/reference-flow.json \
  --operation produce \
  --out-dir /abs/path/artifacts/<case_slug>/process_from_flow/<run_id> \
  --json

node scripts/run-process-automated-builder.mjs resume-build --run-dir /abs/path/artifacts/<case_slug>/process_from_flow/<run_id> --run-id <run_id> --json
node scripts/run-process-automated-builder.mjs publish-build --run-dir /abs/path/artifacts/<case_slug>/process_from_flow/<run_id> --run-id <run_id> --json
node scripts/run-process-automated-builder.mjs batch-build --input /abs/path/batch-request.json --out-dir /abs/path/artifacts/<case_slug>/process_batch/<batch_id> --json
node scripts/run-process-automated-builder.mjs verify-rows --rows-file /abs/path/process-list-report.json --out-dir /abs/path/artifacts/<case_slug>/process-verify
node scripts/run-process-automated-builder.mjs evidence-search plan --query "中国2026年电力结构数据" --out-dir /abs/path/artifacts/<case_slug>/evidence --json

Runtime Requirements

• The wrapper runs the exact published CLI by default through pnpm dlx --package=@tiangong-lca/cli@0.1.18 tiangong-lca.
• Set TIANGONG_LCA_CLI_DIR or pass --cli-dir only when you need a local CLI working tree for dev/CI.
• The wrapper requires explicit output paths instead of relying on cwd/artifacts/... defaults.
• For repeatable runs, use an explicit output root such as /abs/path/artifacts/<case_slug>/....
• The current canonical commands prepare local run outputs and do not depend on legacy private runtimes.
• If a future native CLI command needs additional env, document it in tiangong-lca-cli first and keep this skill as a thin caller only.

Process Name Contract

• Any generated process payload must preserve the four-part process name object: name.baseName, name.treatmentStandardsRoutes, name.mixAndLocationTypes, name.functionalUnitFlowProperties.
• baseName, treatmentStandardsRoutes, and mixAndLocationTypes are schema-required in current TianGong process payloads. Keep the keys even when one field is semantically empty; do not collapse the whole reference-flow short description back into baseName.
• When name splitting is ambiguous, align with ../lifecycleinventory-review/profiles/process/references/process-review-rules.md instead of inventing a one-off local convention.

Fast Troubleshooting

• Local CLI override issues: set TIANGONG_LCA_CLI_DIR or pass --cli-dir only when you intentionally need an unpublished working tree.
• Missing --out-dir or --run-dir: the wrapper requires an explicit output path such as /abs/path/artifacts/<case_slug>/....
• Missing --input / --flow-file: new runs need one explicit request or reference-flow input.
• Run-level conflicts: do not reuse the same run_id across concurrent writers.
• Publish preparation issues: inspect stage_outputs/10_publish/ and cache/agent_handoff_summary.json before touching downstream publish flow.
• Identity blockers: block_duplicate and manual_review are stop states, not warnings.
• Unit-of-analysis blockers: missing unit_of_analysis, manual_review, and blocked_until_scaling_evidence are stop states. Fix the skill-authored decision artifact; do not patch around the CLI gate.
• Evidence blockers: completed_no_sufficient_evidence is a stop state for factual field authoring; completed_with_partial_evidence may proceed only when the partial scope is written into the field evidence and build plan.
• Build-plan blockers: inspect outputs/build-plan-gate-report.json; do not patch around a failed CLI gate inside the skill.
• Process QA findings from qa process are not a substitute for Foundry profile curation. If the rows came from an external import, route schema/QA findings through Foundry curation instead of asking the CLI to make semantic authoring decisions.
• If a required step is missing, add it as a native tiangong-lca process ... command instead of reintroducing a legacy runtime here.

Load References On Demand

• references/workflow-map.md: current CLI-only execution map and output layout.
• references/operations-playbook.md: concise command examples and troubleshooting.