# Conversion (#1539): the agentic `flow-deploy-to-production` skill expressed as a # declarative YAML Flow. This is the largest SDLC flow and uses the multi-agent # documentation pattern in two places — Step 2 (3 parallel validation agents → # readiness synthesis) and Step 4 (4 parallel monitoring agents → production # regression-gate synthesis). Both are encoded as `fanout` steps (agentic-step # extension, #1547 / adr-flow-agentic-steps.md). The SKILL.md remains the thin # wrapper / discoverable trigger surface; this playbook is the orchestration # source of truth for the step sequence, the regression gates, and the # strategy-specific deployment path. # # Schema-fit note (#1539): `inventory`/`targets` are omitted — this flow acts on # a single project/deployment context rather than iterating an inventory. # # Faithful to the prose flow's structure: # Step 1 strategy selection → strategy-selection (deployment-manager) # (strategy confirmation gate) → strategy-confirmation (kind: gate) # Step 2 pre-deploy validation (×3) → predeploy-validation (FANOUT panel + readiness synthesis) # (GO / CONDITIONAL_GO / NO-GO) → readiness-gate (kind: gate) # Step 2.5 deploy to staging → staging-deploy (devops-engineer) # Step 2.5 staging regression check → staging-regression-check (regression-analyst) # (staging regression decision) → staging-regression-gate (kind: gate) # Step 3 execute deployment → execute-deployment (devops-engineer; strategy-specific) # Step 3 smoke tests on candidate → smoke-tests-candidate (qa-engineer) # Step 4 capture production baseline→ capture-baseline (reliability-engineer) # Step 4 monitoring panel (×4) → production-monitoring (FANOUT panel + regression-gate synthesis) # (production regression decision) → production-regression-gate (kind: gate) # Step 5 rollback (conditional) → rollback-execute (devops-engineer; strategy-specific) # Step 6 deployment report + archive→ deployment-report (deployment-manager) # Step 7 stakeholder notification → stakeholder-notification (support-lead) # Step 7 runbook updates → runbook-updates (deployment-manager) apiVersion: flow.aiwg.io/v1 kind: FlowPlaybook metadata: name: flow-deploy-to-production labels: category: sdlc-orchestration domain: deployment spec: vars: regression_threshold: 0 rollback_on_regression: true strategy: "" # blue-green | canary | rolling (set by strategy-selection / guidance) steps: # Step 1: recommend a deployment strategy from project context + guidance. - id: strategy-selection capability: deploy-strategy-selection inputs: - { name: regression_threshold, from: "vars.regression_threshold" } outputs: - name: recommended_strategy # Human gate: present the recommended strategy + trade-offs and confirm. - id: strategy-confirmation kind: gate description: | Human gate (confirm strategy): present the deployment-manager's strategy recommendation — Blue-Green | Canary | Rolling — with rationale, trade-offs (pros/cons), infrastructure requirements, expected duration, and monitoring period. Proceed only on explicit "yes". A "no" returns to strategy-selection to re-recommend. depends_on: [strategy-selection] # Step 2: pre-deployment validation. Three reviewers run in parallel # (quality gates, environment health, rollback readiness); the # deployment-manager synthesizes the Deployment Readiness Report. - id: predeploy-validation fanout: strategy: parallel agents: - deploy-gate-validation # project-manager - deploy-env-health # devops-engineer - deploy-rollback-validation # reliability-engineer synthesize: deploy-readiness-synthesis # deployment-manager depends_on: [strategy-confirmation] outputs: - name: readiness_decision # Human gate: GO | CONDITIONAL_GO | NO-GO on the synthesized readiness report. - id: readiness-gate kind: gate description: | Human gate (GO / CONDITIONAL_GO / NO-GO): review the Deployment Readiness Report synthesized from the three validation reviewers. GO proceeds to the staging regression gate; CONDITIONAL_GO surfaces the conditions that must be addressed before proceeding; NO-GO stops the deployment and reports the gaps + remediation. depends_on: [predeploy-validation] # Step 2.5a: deploy the new version to staging. - id: staging-deploy capability: deploy-staging-deploy depends_on: [readiness-gate] outputs: - name: staging_status # Step 2.5b: staging regression check (staging vs production baseline, full scope). - id: staging-regression-check capability: deploy-staging-regression-check depends_on: [staging-deploy] inputs: - { name: threshold, from: "vars.regression_threshold" } - { name: scope, value: full } outputs: - name: staging_regression_rate # Staging regression decision gate. On FAIL, the regression-analyst impact # assessment is surfaced here and the deployment is BLOCKED (hard stop). - id: staging-regression-gate kind: gate description: | Staging regression gate (hard stop): if staging regression_rate > threshold, BLOCK promotion to production — surface the regression-analyst impact assessment (categorized regressions, root cause, severity, remediation) and STOP. If regression_rate <= threshold, proceed to deployment execution. depends_on: [staging-regression-check] # Step 3a: execute the deployment using the confirmed strategy. The capability # carries the Blue-Green (deploy green → cutover) / Canary (1-5% → 25% → 50% # → 100% progressive) / Rolling (node-by-node) variants from the prose. - id: execute-deployment capability: deploy-execute-deployment depends_on: [staging-regression-gate] inputs: - { name: strategy, from: "vars.strategy" } outputs: - name: deployment_status # Step 3b: smoke tests on the candidate (green env / canary / rolled node) # before/while cutting over real traffic. - id: smoke-tests-candidate capability: deploy-smoke-tests-candidate depends_on: [execute-deployment] outputs: - name: smoke_status # Step 4a: capture the pre-deployment production baseline for regression compare. - id: capture-baseline capability: deploy-capture-baseline depends_on: [smoke-tests-candidate] outputs: - name: baseline_snapshot # Step 4b: production monitoring panel. Four agents run in parallel (SLO # monitoring, production smoke tests, infrastructure health, production # regression check); the regression-analyst synthesizes the production # regression-gate decision. Any BREACH / FAIL / UNHEALTHY routes to rollback. - id: production-monitoring fanout: strategy: parallel agents: - deploy-slo-monitor # reliability-engineer - deploy-prod-smoke-tests # qa-engineer - deploy-infra-health # devops-engineer - deploy-prod-regression-check # regression-analyst synthesize: deploy-regression-gate-synthesis # regression-analyst depends_on: [capture-baseline] inputs: - { name: threshold, from: "vars.regression_threshold" } outputs: - name: production_regression_decision # Production regression decision gate. On FAIL with rollback_on_regression # enabled, the rollback step runs; otherwise the human chooses # rollback | accept | investigate. - id: production-regression-gate kind: gate description: | Production regression gate: if any monitoring agent reports BREACH / FAIL / UNHEALTHY, or production regression_rate > threshold, then — when --rollback-on-regression is enabled (default) — trigger the automated rollback step. When disabled, the human decides: rollback | accept risk | investigate. If all checks pass for the monitoring duration, the deployment is successful and proceeds to reporting. depends_on: [production-monitoring] # Step 5: strategy-specific rollback (conditional). Executes only when the # regression gate routes here; verifies rollback, re-checks regression, and # declares the incident + RCA. Not a hard dependency on success of the gate — # the executor invokes it on the failure branch. - id: rollback-execute capability: deploy-rollback-execute depends_on: [production-regression-gate] inputs: - { name: strategy, from: "vars.strategy" } outputs: - name: rollback_status # Step 6: synthesize the Deployment Summary Report from all artifacts and # archive the working deployment directory + audit trail. - id: deployment-report capability: deploy-deployment-report depends_on: [production-regression-gate] outputs: - name: report_path # Step 7a: draft the stakeholder notification (support-lead). Human sends it. - id: stakeholder-notification capability: deploy-stakeholder-notification depends_on: [deployment-report] # Step 7b: capture lessons-learned runbook updates (deployment-manager). - id: runbook-updates capability: deploy-runbook-updates depends_on: [deployment-report]