ljwx-deploy/PLAN.md

Bid-MVP Factory Control Planes Plan (ljwx-deploy, refreshed on 2026-03-03)

Phase 0 Inventory

Existing

• Workflows exist under .github/workflows/:
  • nightly-evidence.yml
  • nightly-auto-repair.yml
  • deploy-repo-gate.yml
  • build-promoter-image.yml
• Evidence control plane exists:
  • evidence/schema/evidence-record.schema.json
  • evidence/schema/evidence.schema.json
  • evidence/records/ (directory tracked)
  • scripts/evidence/collect.py
  • scripts/evidence/validate.py
  • evidence/index.json, evidence/summary/latest.md
• Pages dashboard source exists in both pages/ and site/ (needs consolidation decision).
• Release queue + promoter exists:
  • release/queue.yaml, release/services.yaml
  • scripts/promoter/promote.py, scripts/promoter/deploy_promoter.py, scripts/promoter/validate_queue.py
  • CronJob manifests: infra/factory/deploy-promoter-cronjob.yaml, k8s/promoter/deploy-promoter-cronjob.yaml
• Smoke loop exists:
  • scripts/smoke/run_smoke.py, scripts/smoke/targets.json
  • infra/factory/smoke-runner-cronjob.yaml
• Auto-repair loop exists:
  • repairs/recipes.yaml
  • scripts/repair/run.sh, scripts/repair/run_repair.py, scripts/repair/diagnose.py
  • orchestration docs/scripts: infra/factory/INTEGRATION_RUN_ALL.md, scripts/factory/run_all_dev.sh, TEST.md

Current Gap / To Improve

• Two-cluster invariant hardening (required):
  • Same GitOps code must support both local server k3s and China mainland dev server OrbStack k3s.
  • Differences must be represented via overlays/values/env only, no forked logic or duplicated pipelines.
• Promoter implementation appears duplicated (promote.py and deploy_promoter.py):
  • choose one canonical entrypoint to avoid drift and behavioral mismatch.
• Queue state alignment:
  • ensure docs/code agree on pending/promoted/failed/superseded semantics and data placement.
• Pages source alignment:
  • choose one source directory (pages/ or site/) and document workflow contract.
• Deploy gate hardening:
  • verify gate blocks unauthorized file changes and validates queue/evidence schemas deterministically.

Existing promoter/test-runner/evidence/pages status

• Promoter: present.
• Test-runner/smoke runner: present.
• Evidence model/feed: present.
• Pages/gh-pages publishing: present.
• Conclusion: Phases 2-5 are largely implemented; next step is consistency hardening and elimination of drift.

Planned File Changes (next deploy iterations)

• scripts/promoter/* (canonicalize single promoter path and behavior)
• release/README.md, release/queue.yaml (shape/field consistency)
• .github/workflows/nightly-evidence.yml + docs (single pages source contract)
• docs/pages.md, docs/ops-runbook.md (two-k3s same-code rule and runbook precision)
• argocd-apps/ and/or apps/*/overlays/* (if overlay naming normalization is needed for dual k3s)

Local Validation Commands

• test -f PLAN.md && sed -n '1,260p' PLAN.md
• python3 scripts/evidence/validate.py
• python3 scripts/evidence/collect.py --out evidence/index.json --summary evidence/summary/latest.md
• python3 scripts/promoter/validate_queue.py --queue release/queue.yaml
• python3 scripts/promoter/promote.py --dry-run --skip-registry-check
• python3 scripts/smoke/run_smoke.py --dry-run
• python3 scripts/repair/diagnose.py --help
• python3 scripts/repair/run_repair.py --help

Commit Plan

• phase0(deploy): refresh inventory plan
• phase3(deploy): promoter/queue canonicalization + dual-k3s overlay rule hardening
• phase4(deploy): smoke/evidence consistency hardening
• phase5(deploy): auto-repair determinism and gate tightening