设计与创意 · SKILL.md

multi-stage-deployment

multi-stage-deployment

【设计与创意】Designs multi-stage deployment pipelines that reduce risk through progressive disclosure, environment parity, and quality-gated stage transitions from development to production.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

Ppaulpas / agent-skill-routerMIT

技能说明

GitHub

name: multi-stage-deployment description: Designs multi-stage deployment pipelines that reduce risk through progressive disclosure, environment parity, and quality-gated stage transitions from development to production. license: MIT compatibility: opencode metadata: version: "1.0.0" domain: cncf triggers: multi-stage deployment, deployment stages, environment parity, deployment pipeline, stage gates, dev staging pre-prod, deployment readiness, progressive delivery archetypes:

educational
diagnostic anti_triggers:
brainstorming
vague ideation
non-containerized architecture response_profile: verbosity: medium directive_strength: medium abstraction_level: tactical role: reference scope: infrastructure output-format: analysis content-types:
guidance
examples
do-dont related-skills: deployment-philosophy,blue-green-deployment,canary-deployment,state-management,rollback-strategy,deployment-orchestration

Multi-Stage Deployment Pipeline

Architects deployment pipelines as a series of quality-gated stages, each with a distinct purpose and validation focus. Treats every stage boundary as a decision point that must earn the right to expose more users to the change.

TL;DR Checklist

Define the quality dimension each stage validates (correctness, integration, performance, production fit)
Ensure environment parity exists where it matters most: staging must mirror production config and topology
Write explicit gate criteria per stage — not "tests pass" but "what quality dimension is satisfied"
Identify team handoff points at each stage boundary — who approves, who operates, who owns
Calculate cost vs. safety tradeoff for each stage: is the gate worth the delay?
Treat canary and blue-green as stage transitions (staging → pre-prod → prod), not standalone deployment patterns

When to Use

Use this skill when:

Designing a deployment pipeline from scratch — You need to define what stages exist, what each validates, and what gates control transitions
Auditing an existing pipeline — You're evaluating whether current stages provide adequate risk reduction or have gaps
Debugging a production incident caused by a deployment — You need to trace which stage should have caught the issue and why it was missed
Onboarding a team to a new release process — You're establishing the multi-stage model for a team without one
Evaluating whether to add or remove stages — You're debating pipeline overhead vs. risk coverage

When NOT to Use

Avoid this skill for:

Implementing a specific CI/CD tool — Use deployment-orchestration or the specific tool's skill (e.g., Jenkins, GitLab CI) for tactical implementation
Choosing a deployment strategy (canary, blue-green) — Use blue-green-deployment or canary-deployment for those tactical decisions
Simple projects with minimal risk — A single-stage deploy (dev → prod) may be appropriate for internal tools with zero user impact
Writing application code or tests — This skill focuses on pipeline architecture and process, not code quality

Core Workflow

1. Define the Stage Model

Determine the number and identity of stages in your pipeline. The canonical model has four stages:

  dev ──→ staging ──→ pre-prod ──→ prod
  │         │           │           │
  │         │           │           └─ Live user traffic
  │         │           └─ Production-like, limited real traffic
  │         └─ Integration validation, performance, security
  └─ Developer correctness, unit tests, local behavior

Each stage exists for one reason: to catch a specific class of failure before it reaches the next wider audience.

Checkpoint: You can state in one sentence what each stage's unique failure-catchment purpose is. If two stages have the same purpose, they should be merged.

2. Assign a Quality Dimension to Each Stage

Every stage validates one primary quality dimension. This prevents stages from becoming generic "test everything here" dumping grounds.

Stage	Primary Quality Dimension	What It Answers
Dev	Correctness	Does the code do what its unit of work says it does?
Staging	Integration	Does the code work with all its dependencies in a realistic environment?
Pre-Prod	Production Fit	Does the code behave correctly under production load, data volumes, and failure modes?
Prod	User Impact	Does the code improve or maintain user experience in reality?

Checkpoint: No stage exists solely for "running tests." Tests serve quality dimensions — correctness, integration, performance, reliability. Map each test suite to a dimension and a stage.

3. Define Gate Criteria by Quality Dimension

Gate criteria are not "do tests pass?" They answer: "Has this stage satisfied its quality dimension?"

  Stage      │ Gate Criteria (examples)
  ───────────┼────────────────────────────────────────────────────
  Dev        │ • Unit test coverage above threshold
              │   • Static analysis clean (no critical/high issues)
              │   • Code review completed and approved
              │   • No breaking contract changes without migration plan
  Staging    │ • Integration tests against all dependency contracts pass
              │   • Performance regression < 5% vs baseline
              │   • Security scan clean (SAST + DAST)
              │   • Data migration tested end-to-end
  Pre-Prod   │ • Load test meets SLO targets at 1.5x peak traffic
              │   • Chaos test: service recovers from failure within RTO
              │   • Rollback tested and verified within RTO
              │   • Observability: all required metrics, logs, traces present
  Prod       │ • Canary metrics stable for observation window
              │   • Error budget consumption within acceptable rate
              │   • No regression in key business metrics
              │   • Manual approval gate passed (for critical changes)

Checkpoint: Every gate criterion must be measurable and automatable where possible. "Code is well-tested" is not a gate. "Unit test coverage ≥ 80% and all critical paths covered" is.

4. Map Environment Parity

Not every aspect of an environment needs to be identical across stages. Parity matters where it affects behavior — and you must explicitly declare what differs.

  Aspect              │ Dev │ Staging │ Pre-Prod │ Prod │ Parity Required?
  ────────────────────┼─────┼─────────┼──────────┼──────┼─────────────────
  Code version        │  ✅  │    ✅    │    ✅    │  ✅  │ Always identical
  Schema              │  ✅  │    ✅    │    ✅    │  ✅  │ Always identical
  Dependencies        │  ⚠️  │    ✅    │    ✅    │  ✅  │ Yes (versions)
  Config format       │  ⚠️  │    ✅    │    ✅    │  ✅  │ Yes (structure)
  Infrastructure size │  ❌  │    ⚠️    │    ✅    │  ✅  │ Pre-prod ≈ Prod
  Data volume         │  ❌  │    ⚠️    │    ✅    │  ✅  │ Pre-prod ≈ Prod
  External services   │  ⚠️  │    ✅    │    ✅    │  ✅  │ Yes (interfaces)
  Network topology    │  ❌  │    ⚠️    │    ✅    │  ✅  │ Pre-prod = Prod
  Security policies   │  ❌  │    ✅    │    ✅    │  ✅  │ Yes
  Monitoring depth    │  ❌  │    ✅    │    ✅    │  ✅  │ Yes
  Team access         │  ✅  │    ✅    │    ✅    │  ⚠️   │ No (least privilege)

Checkpoint: If a configuration difference between staging and production has caused a bug, that aspect becomes a parity requirement. Maintain a "parities violated" log and close each item.

5. Define Team Handoff Points

Each stage boundary is a handoff where responsibility, authority, and information transfer occur.

  Dev ──→ Staging     │ Owner: Developer → QA/Platform
                        │ Approval: Developer PR + CI pass
                        │ Info: What changed, why, and how to verify
  Staging ──→ Pre-Prod│ Owner: QA/Platform → SRE/Release Manager
                        │ Approval: Staging gate + security sign-off
                        │ Info: Test results, performance baseline, rollback plan
  Pre-Prod ──→ Prod   │ Owner: SRE/Release Manager → On-Call Engineer
                        │ Approval: Pre-prod gate + change advisory (if applicable)
                        │ Info: Load test results, chaos results, rollback verified
  Prod (canary)       │ Owner: On-Call Engineer → Release Manager
                        │ Approval: Automated canary metrics gate
                        │ Info: Real-user metrics during canary window

Checkpoint: Every handoff has a named role on both sides, a clear approval criteria, and a minimum information package. If any of these is missing, the handoff will fail.

6. Treat Canary and Blue-Green as Stage Transitions

Canary and blue-green are not deployment strategies in isolation — they are the mechanism for transitioning between pre-prod and prod stages.

  Traditional model:  pre-prod ──┤─── instant full deploy ──┤── prod
                                           (high blast radius)

  Progressive model:  pre-prod ──┤── canary (1%) ──┤── canary (10%) ──┤── full prod
                             │                      │                      │
                       health gate              health gate            health gate
                       (metrics pass)           (metrics pass)         (metrics pass)

  Blue-green as transition:  pre-prod ──┤── green deployed ──┤── traffic flip ──┤── prod (green)
                                          │                   │                      │
                                    validate green      atomic switch             observe

Checkpoint: Canary is for gradual risk acceptance. Blue-green is for instant risk rejection. Choose based on your blast radius tolerance and rollback requirements. Both are stage transitions, not deployment patterns.

7. Establish Rollback vs. Forward-Fix Philosophy

When a deployment fails in production, the response must be deterministic — not debatable.

  Rollback triggers (automatic):
  ├─ Error rate exceeds threshold (e.g., > 1% 5xx responses)
  ├─ p99 latency exceeds SLO (e.g., > 500ms)
  ├─ Business metric drops below acceptable floor (e.g., checkout conversions < 95%)
  └─ Health check fails for observation window

  Forward-fix triggers (only when rollback is impossible or harmful):
  ├─ Data migration is partially applied and cannot be reversed
  ├─ Rollback would cause greater data corruption than the current issue
  └─ The "bug" is actually a desired change that was miscommunicated

  Rule: When in doubt, roll back. Forward-fix requires written justification.

Checkpoint: Your rollback mechanism must be faster than your detection. If you detect the failure in 10 minutes but rollback takes 30 minutes, your rollback is part of the problem.

8. Assess Cost vs. Safety at Each Stage

Every stage has an opportunity cost. The pipeline must balance speed against risk reduction.

  Stage       │ Cost (time, resources) │ Risk Reduction │ Cost/Benefit Assessment
  ────────────┼────────────────────────┼────────────────┼────────────────────────
  Dev         │ Low (parallel, fast)   │ High           │ Cheap gate: catch bugs before they leave developer machine
  Staging     │ Medium (queue wait,   │ High           │ Worth it: catches integration failures that dev misses
              │  shared env)           │                │   but could affect multiple services
  Pre-Prod    │ High (dedicated env,  │ Critical       │ Must have: only stage that can catch production-specific
              │  load/chaos testing)   │                │   failures before user impact
  Prod (canary)│ Low (partial traffic │ Variable       │ Tradeoff: canary catches some issues but lets others
              │  exposure)            │                │   through — is that acceptable for this change?

Checkpoint: If a stage consistently passes without catching any failures for 3 consecutive months, evaluate whether it is still providing value or has become overhead.

9. Define Deployment Readiness as a First-Class Property

Deployment readiness is not "all tests pass." It is a verifiable state that must be computed, not assumed.

  Deployment Readiness Checklist (computed, not manual):
  ├─ [ ] All gate criteria for current stage are met
  ├─ [ ] Deployment artifacts are immutable (pinned to exact image/tag/commit)
  ├─ [ ] Rollback procedure is tested and documented
  ├─ [ ] Observability coverage is verified (metrics, logs, traces)
  ├─ [ ] Feature flags for the change are in known state
  ├─ [ ] Team on-call is aware and accessible
  ├─ [ ] Change documentation is updated
  └─ [ ] Emergency contacts are verified

  Readiness score = sum of all checks (binary: pass/fail)
  Gate = readiness score == 8 (no partial readiness)

Checkpoint: Readiness must be machine-computable where possible. Manual checklists are fallbacks, not primary mechanisms.

Implementation Patterns

Pattern 1: Stage Gate Enforcer

A shell function that validates whether a stage's quality gates have been satisfied before allowing progression.

#!/usr/bin/env bash
# ✅ GOOD: Stage gate enforcement — each stage has explicit, measurable criteria
# Gates are computed, not assumed. Partial readiness is rejected.

set -euo pipefail

readonly STAGE_CONFIG="/tmp/stage-gate-config.json"
readonly READINESS_FILE="/tmp/deployment-readiness.json"

# --- Gate definitions per stage ---
# Each stage has a name and a list of gate checkers
# Usage: validate_stage <stage_name>

validate_stage() {
  local stage="$1"

  case "$stage" in
    dev)
      validate_dev_gates
      ;;
    staging)
      validate_staging_gates
      ;;
    preprod)
      validate_preprod_gates
      ;;
    prod)
      validate_prod_gates
      ;;
    *)
      echo "ERROR: Unknown stage: $stage" >&2
      return 1
      ;;
  esac
}

# --- Dev gates: correctness ---
validate_dev_gates() {
  local score=0
  local total=3
  local results=()

  echo "=== Dev Stage Gate: Correctness ==="

  # Gate 1: Unit test coverage
  local coverage
  coverage=$(jq -r '.unit_test_coverage' "$READINESS_FILE" 2>/dev/null || echo "0")
  if (( $(echo "$coverage >= 80" | bc -l) )); then
    results+=("unit_test_coverage: PASS ($coverage%)")
    ((score++))
  else
    results+=("unit_test_coverage: FAIL ($coverage% < 80% threshold)")
  fi

  # Gate 2: Static analysis
  local static_issues
  static_issues=$(jq -r '.static_analysis_critical_high' "$READINESS_FILE" 2>/dev/null || echo "999")
  if [[ "$static_issues" == "0" ]]; then
    results+=("static_analysis: PASS (0 critical/high issues)")
    ((score++))
  else
    results+=("static_analysis: FAIL ($static_issues critical/high issues)")
  fi

  # Gate 3: Code review
  local review_status
  review_status=$(jq -r '.code_review_status' "$READINESS_FILE" 2>/dev/null || echo "unreviewed")
  if [[ "$review_status" == "approved" ]]; then
    results+=("code_review: PASS")
    ((score++))
  else
    results+=("code_review: FAIL (status: $review_status)")
  fi

  # Output results
  echo "Dev gate score: $score/$total"
  for r in "${results[@]}"; do
    echo "  $r"
  done

  if [[ $score -eq $total ]]; then
    echo "✅ Dev stage: READY to advance"
    return 0
  else
    echo "❌ Dev stage: BLOCKED — $((total - score)) gate(s) failed"
    return 1
  fi
}

# --- Staging gates: integration ---
validate_staging_gates() {
  local score=0
  local total=4
  local results=()

  echo "=== Staging Stage Gate: Integration ==="

  # Gate 1: Integration test pass rate
  local integration_rate
  integration_rate=$(jq -r '.integration_test_pass_rate' "$READINESS_FILE" 2>/dev/null || echo "0")
  if (( $(echo "$integration_rate >= 100" | bc -l) )); then
    results+=("integration_tests: PASS ($integration_rate%)")
    ((score++))
  else
    results+=("integration_tests: FAIL ($integration_rate% < 100%)")
  fi

  # Gate 2: Performance regression check
  local perf_regression
  perf_regression=$(jq -r '.perf_regression_pct' "$READINESS_FILE" 2>/dev/null || echo "100")
  if (( $(echo "$perf_regression <= 5" | bc -l) )); then
    results+=("performance: PASS (regression: ${perf_regression}%)")
    ((score++))
  else
    results+=("performance: FAIL (regression: ${perf_regression}% > 5% threshold)")
  fi

  # Gate 3: Security scan
  local security_status
  security_status=$(jq -r '.security_scan_status' "$READINESS_FILE" 2>/dev/null || echo "fail")
  if [[ "$security_status" == "clean" ]]; then
    results+=("security_scan: PASS")
    ((score++))
  else
    results+=("security_scan: FAIL (status: $security_status)")
  fi

  # Gate 4: Data migration tested
  local migration_tested
  migration_tested=$(jq -r '.data_migration_tested' "$READINESS_FILE" 2>/dev/null || echo "false")
  if [[ "$migration_tested" == "true" ]]; then
    results+=("data_migration: PASS")
    ((score++))
  else
    results+=("data_migration: FAIL (not tested)")
  fi

  echo "Staging gate score: $score/$total"
  for r in "${results[@]}"; do
    echo "  $r"
  done

  if [[ $score -eq $total ]]; then
    echo "✅ Staging stage: READY to advance"
    return 0
  else
    echo "❌ Staging stage: BLOCKED — $((total - score)) gate(s) failed"
    return 1
  fi
}

# --- Pre-prod gates: production fit ---
validate_preprod_gates() {
  local score=0
  local total=4
  local results=()

  echo "=== Pre-Prod Stage Gate: Production Fit ==="

  # Gate 1: Load test
  local load_p99
  load_p99=$(jq -r '.load_test_p99_latency' "$READINESS_FILE" 2>/dev/null || echo "9999")
  local load_pass
  load_pass=$(jq -r '.load_test_p99_within_slo' "$READINESS_FILE" 2>/dev/null || echo "false")
  if [[ "$load_pass" == "true" ]]; then
    results+=("load_test: PASS (p99: ${load_p99}ms within SLO)")
    ((score++))
  else
    results+=("load_test: FAIL (p99: ${load_p99}ms exceeds SLO)")
  fi

  # Gate 2: Chaos test
  local chaos_recovery
  chaos_recovery=$(jq -r '.chaos_test_recovery_time' "$READINESS_FILE" 2>/dev/null || echo "9999")
  if (( $(echo "$chaos_recovery <= 300" | bc -l) )); then
    results+=("chaos_test: PASS (recovery: ${chaos_recovery}s <= 5min RTO)")
    ((score++))
  else
    results+=("chaos_test: FAIL (recovery: ${chaos_recovery}s > 5min RTO)")
  fi

  # Gate 3: Rollback tested
  local rollback_tested
  rollback_tested=$(jq -r '.rollback_tested' "$READINESS_FILE" 2>/dev/null || echo "false")
  local rollback_time
  rollback_time=$(jq -r '.rollback_time_seconds' "$READINESS_FILE" 2>/dev/null || echo "9999")
  if [[ "$rollback_tested" == "true" ]] && (( $(echo "$rollback_time <= 60" | bc -l) )); then
    results+=("rollback_test: PASS (${rollback_time}s <= 60s target)")
    ((score++))
  else
    results+=("rollback_test: FAIL (tested: $rollback_tested, time: ${rollback_time}s)")
  fi

  # Gate 4: Observability coverage
  local obs_coverage
  obs_coverage=$(jq -r '.observability_score' "$READINESS_FILE" 2>/dev/null || echo "0")
  if (( $(echo "$obs_coverage >= 90" | bc -l) )); then
    results+=("observability: PASS (score: $obs_coverage/100)")
    ((score++))
  else
    results+=("observability: FAIL (score: $obs_coverage/100 < 90 threshold)")
  fi

  echo "Pre-Prod gate score: $score/$total"
  for r in "${results[@]}"; do
    echo "  $r"
  done

  if [[ $score -eq $total ]]; then
    echo "✅ Pre-Prod stage: READY to advance"
    return 0
  else
    echo "❌ Pre-Prod stage: BLOCKED — $((total - score)) gate(s) failed"
    return 1
  fi
}

# --- Prod canary gates: user impact ---
validate_prod_gates() {
  local score=0
  local total=4
  local results=()
  local canary_weight="${CANARY_WEIGHT:-0.05}"

  echo "=== Prod Stage Gate: User Impact (canary: ${canary_weight}) ==="

  # Gate 1: Error rate
  local error_rate
  error_rate=$(jq -r '.prod_canary_error_rate' "$READINESS_FILE" 2>/dev/null || echo "1")
  if (( $(echo "$error_rate <= 0.01" | bc -l) )); then
    results+=("error_rate: PASS (${error_rate} <= 1%)")
    ((score++))
  else
    results+=("error_rate: FAIL (${error_rate} > 1% threshold)")
  fi

  # Gate 2: Latency SLO
  local latency_slo_met
  latency_slo_met=$(jq -r '.prod_canary_latency_slo_met' "$READINESS_FILE" 2>/dev/null || echo "false")
  if [[ "$latency_slo_met" == "true" ]]; then
    results+=("latency_slo: PASS")
    ((score++))
  else
    results+=("latency_slo: FAIL (p99 exceeds SLO)")
  fi

  # Gate 3: Business metric
  local business_metric_delta
  business_metric_delta=$(jq -r '.business_metric_delta_pct' "$READINESS_FILE" 2>/dev/null || echo "-100")
  if (( $(echo "$business_metric_delta >= -5" | bc -l) )); then
    results+=("business_metric: PASS (delta: ${business_metric_delta}%)")
    ((score++))
  else
    results+=("business_metric: FAIL (delta: ${business_metric_delta}% < -5% floor)")
  fi

  # Gate 4: Observation window
  local observation_met
  observation_met=$(jq -r '.canary_observation_met' "$READINESS_FILE" 2>/dev/null || echo "false")
  if [[ "$observation_met" == "true" ]]; then
    results+=("observation: PASS (required window passed)")
    ((score++))
  else
    results+=("observation: FAIL (required observation window not yet met)")
  fi

  echo "Prod canary gate score: $score/$total"
  for r in "${results[@]}"; do
    echo "  $r"
  done

  if [[ $score -eq $total ]]; then
    echo "✅ Prod canary: READY for next weight increase or full rollout"
    return 0
  else
    echo "❌ Prod canary: BLOCKED — $((total - score)) gate(s) failed"
    echo "ACTION: Roll back canary weight to previous level"
    return 1
  fi
}

# --- Main pipeline orchestrator ---
run_pipeline() {
  local stages="dev staging preprod prod"
  local current_stage=""
  local gate_failed=false

  echo "========================================="
  echo " Multi-Stage Deployment Pipeline"
  echo "========================================="

  for stage in $stages; do
    echo ""
    echo "▶ Entering stage: $stage"
    if ! validate_stage "$stage"; then
      gate_failed=true
      current_stage="$stage"
      break
    fi
  done

  echo ""
  echo "========================================="
  if $gate_failed; then
    echo "Pipeline BLOCKED at stage: $current_stage"
    echo "ACTION: Fix gate failures before retrying"
    exit 1
  else
    echo "Pipeline COMPLETE — deployment ready for production"
    exit 0
  fi
}

Pattern 2: Environment Parity Validator

Validate that environments have the required parity before advancing to the next stage.

# ✅ GOOD: Environment parity validation
# Checks that critical aspects match between source and target environments
# Uses a declarative parity spec — what MUST match, what CAN differ

set -euo pipefail

# --- Parity specification ---
# Format: "aspect:source_env:target_env:tolerance"
# tolerance: "exact", "semver", "within_x%"
PARITY_SPEC=(
  "code_version:staging:preprod:exact"
  "code_version:preprod:prod:exact"
  "database_schema:staging:preprod:exact"
  "database_schema:preprod:prod:exact"
  "dependency_versions:staging:preprod:semver"
  "dependency_versions:preprod:prod:semver"
  "config_structure:staging:preprod:exact"
  "config_structure:preprod:prod:exact"
  "infrastructure_size:preprod:prod:within_20%"
  "network_topology:preprod:prod:exact"
  "security_policy:preprod:prod:exact"
)

# --- Fetch environment state (abstracted — could be kubectl, terraform, etc.) ---
get_env_value() {
  local env="$1"
  local aspect="$2"
  # In production, this would query: kubectl get configmap, terraform state, etc.
  # Here we simulate with a JSON state file
  local state_file="/tmp/env-state-${env}.json"
  jq -r --arg a "$aspect" '.environments[$env].aspects[$a] // "not_found"' "$state_file" 2>/dev/null || echo "not_found"
}

# --- Compare two values with tolerance ---
compare_with_tolerance() {
  local val1="$1"
  local val2="$2"
  local tolerance="$3"

  case "$tolerance" in
    exact)
      [[ "$val1" == "$val2" ]]
      ;;
    semver)
      # Both must be valid semver, major+minor must match
      local maj1 maj2 min1 min2
      maj1=$(echo "$val1" | cut -d. -f1)
      min1=$(echo "$val1" | cut -d. -f2)
      maj2=$(echo "$val2" | cut -d. -f1)
      min2=$(echo "$val2" | cut -d. -f2)
      [[ "$maj1" == "$maj2" && "$min1" == "$min2" ]]
      ;;
    within_*)
      local threshold="${tolerance#within_}"
      local ratio
      ratio=$(echo "if ($val1 > $val2) ($val2/$val1) else ($val1/$val2)" | bc -l)
      local min_ratio
      min_ratio=$(echo "1 - $threshold / 100" | bc -l)
      (( $(echo "$ratio >= $min_ratio" | bc -l) ))
      ;;
    *)
      echo "ERROR: Unknown tolerance: $tolerance" >&2
      return 1
      ;;
  esac
}

# --- Run parity validation ---
validate_parity() {
  local source_env="$1"
  local target_env="$2"

  echo "=== Environment Parity: $source_env → $target_env ==="

  local passed=0
  local failed=0
  local violations=()

  for spec in "${PARITY_SPEC[@]}"; do
    IFS=':' read -r aspect src trg tolerance <<< "$spec"

    # Only check relevant pairs
    if [[ "$src" != "$source_env" && "$src" != "$target_env" ]]; then
      continue
    fi

    local check_source="$src"
    local check_target="$trg"
    [[ "$src" == "$target_env" ]] && { check_source="$target_env"; check_target="$source_env"; }

    local val1 val2
    val1=$(get_env_value "$check_source" "$aspect")
    val2=$(get_env_value "$check_target" "$aspect")

    if compare_with_tolerance "$val1" "$val2" "$tolerance"; then
      echo "  ✅ $aspect: $val1 ≈ $val2 ($tolerance)"
      ((passed++))
    else
      echo "  ❌ $aspect: $val1 ≠ $val2 (tolerance: $tolerance)"
      violations+=("$aspect: $val1 vs $val2")
      ((failed++))
    fi
  done

  echo ""
  echo "Parity score: $passed passed, $failed failed"

  if [[ $failed -gt 0 ]]; then
    echo ""
    echo "Violations:"
    for v in "${violations[@]}"; do
      echo "  ⚠ $v"
    done
    echo ""
    echo "❌ Parity check FAILED — fix violations before advancing"
    return 1
  else
    echo "✅ All parity checks passed"
    return 0
  fi
}

# --- Usage ---
# validate_parity staging preprod
# validate_parity preprod prod

Pattern 3: Deployment Readiness Scorecard

Compute a machine-readable deployment readiness score from gate results.

# ✅ GOOD: Deployment readiness computed from gate results
# Score is a percentage — gates with score < 100 block deployment
# Partial readiness is never accepted for production deployment

set -euo pipefail

readonly READINESS_FILE="/tmp/deployment-readiness.json"

# --- Readiness computation ---
compute_readiness() {
  local stage="$1"
  local score=0
  local total=0
  local checks=()

  case "$stage" in
    dev)
      # Dev readiness: 3 binary checks
      total=3
      for check in unit_test_coverage static_analysis code_review; do
        local status
        status=$(jq -r --arg c "$check" '.dev_gates[$c]' "$READINESS_FILE" 2>/dev/null || echo "false")
        if [[ "$status" == "true" ]]; then
          ((score++))
        fi
        checks+=("$check:$status")
      done
      ;;
    staging)
      # Staging readiness: 4 binary checks
      total=4
      for check in integration_tests performance security data_migration; do
        local status
        status=$(jq -r --arg c "$check" '.staging_gates[$c]' "$READINESS_FILE" 2>/dev/null || echo "false")
        if [[ "$status" == "true" ]]; then
          ((score++))
        fi
        checks+=("$check:$status")
      done
      ;;
    preprod)
      # Pre-prod readiness: 4 binary checks
      total=4
      for check in load_test chaos_test rollback_test observability; do
        local status
        status=$(jq -r --arg c "$check" '.preprod_gates[$c]' "$READINESS_FILE" 2>/dev/null || echo "false")
        if [[ "$status" == "true" ]]; then
          ((score++))
        fi
        checks+=("$check:$status")
      done
      ;;
    prod)
      # Prod canary readiness: 4 binary checks
      total=4
      for check in error_rate latency_slo business_metric observation; do
        local status
        status=$(jq -r --arg c "$check" '.prod_gates[$c]' "$READINESS_FILE" 2>/dev/null || echo "false")
        if [[ "$status" == "true" ]]; then
          ((score++))
        fi
        checks+=("$check:$status")
      done
      ;;
    *)
      echo "ERROR: Unknown stage: $stage" >&2
      return 1
      ;;
  esac

  local percentage
  percentage=$(( (score * 100) / total ))

  # Output structured readiness result
  echo "{
    \"stage\": \"$stage\",
    \"score\": $score,
    \"total\": $total,
    \"percentage\": $percentage,
    \"checks\": {$(printf '"%s": "%s",' "${checks[@]}")},
    \"ready\": $([ $score -eq $total ] && echo "true" || echo "false")
  }"
}

# --- Readiness report ---
readiness_report() {
  echo "========================================="
  echo "  Deployment Readiness Report"
  echo "========================================="

  for stage in dev staging preprod prod; do
    local result
    result=$(compute_readiness "$stage")
    local ready
    ready=$(echo "$result" | jq -r '.ready')

    if [[ "$ready" == "true" ]]; then
      echo "✅ $stage: READY ($(echo "$result" | jq -r '.score')/$(echo "$result" | jq -r '.total'))"
    else
      echo "❌ $stage: BLOCKED ($(echo "$result" | jq -r '.score')/$(echo "$result" | jq -r '.total'))"
      echo "    Failed: $(echo "$result" | jq -r 'to_entries[] | select(.value == "false") | .key')"'
    fi
  done

  echo "========================================="
}

# --- Gate enforcement: block if any stage is blocked ---
enforce_gates() {
  local blocked_stages
  blocked_stages=""

  for stage in dev staging preprod prod; do
    local result
    result=$(compute_readiness "$stage")
    local ready
    ready=$(echo "$result" | jq -r '.ready')

    if [[ "$ready" != "true" ]]; then
      blocked_stages="$blocked_stages $stage"
    fi
  done

  if [[ -n "$blocked_stages" ]]; then
    echo "❌ Deployment BLOCKED at:$blocked_stages"
    echo "ACTION: Resolve gate failures before proceeding"
    return 1
  else
    echo "✅ All stages READY — deployment permitted"
    return 0
  fi
}

Pattern 4: Stage Transition with Handoff Protocol

Define the handoff protocol between stages — who approves, what information transfers.

# ✅ GOOD: Stage handoff protocol with structured information transfer
# Each handoff requires: approver role, information package, and confirmation

set -euo pipefail

# --- Handoff definitions ---
# Format: "source_stage:target_stage:approver_role:info_package"
declare -A HANDOFFS=(
  ["dev:staging"]="developer:PR diff, test summary, deployment instructions"
  ["staging:preprod"]="qa_lead:staging test report, performance baseline, security scan"
  ["preprod:prod"]="release_manager:load test results, chaos results, rollback verification"
  ["prod:prod_canary"]="on_call_engineer:canary metrics, business impact assessment"
)

# --- Handoff execution ---
execute_handoff() {
  local from_stage="$1"
  local to_stage="$2"
  local handoff_key="${from_stage}:${to_stage}"

  # Get handoff requirements
  local requirements="${HANDOFFS[$handoff_key]:-}"
  if [[ -z "$requirements" ]]; then
    echo "ERROR: No handoff defined for $from_stage → $to_stage" >&2
    return 1
  fi

  IFS=':' read -r approver info_package <<< "$requirements"

  echo "========================================="
  echo " Handoff: $from_stage → $to_stage"
  echo " Approver: $approver"
  echo " Info package: $info_package"
  echo "========================================="

  # Verify information package is complete
  local info_complete=true
  local missing_items=()

  # Check each required info item
  IFS=', ' read -ra items <<< "$info_package"
  for item in "${items[@]}"; do
    item=$(echo "$item" | xargs)  # trim whitespace
    # In production, verify the actual artifact exists
    # For simulation, check a file marker
    local marker_file="/tmp/handoff-${from_stage}-${to_stage}-${item}"
    if [[ -f "$marker_file" ]]; then
      echo "  ✅ $item: present"
    else
      echo "  ❌ $item: missing"
      missing_items+=("$item")
      info_complete=false
    fi
  done

  echo ""

  if ! $info_complete; then
    echo "❌ Handoff BLOCKED — missing: ${missing_items[*]}"
    return 1
  fi

  echo "✅ Information package complete"
  echo "▶ Awaiting approval from: $approver"

  # In production, this would check for approval from the named role
  # For simulation, we proceed
  echo "▶ Approval granted by: $approver"
  echo "✅ Handoff complete — $to_stage is now active"
  return 0
}

# --- Execute full handoff chain ---
run_handoffs() {
  local handoff_pairs=("dev:staging" "staging:preprod" "preprod:prod" "prod:prod_canary")
  local failed=false

  for pair in "${handoff_pairs[@]}"; do
    IFS=':' read -r from to <<< "$pair"
    if ! execute_handoff "$from" "$to"; then
      failed=true
      break
    fi
    echo ""
  done

  if $failed; then
    echo "❌ Handoff chain broken — pipeline halted"
    return 1
  else
    echo "✅ Full handoff chain complete — deployment ready"
    return 0
  fi
}

Constraints

MUST DO

Assign exactly one primary quality dimension to each stage — Prevents stages from becoming generic test dumping grounds
Write gate criteria that map to quality dimensions — "Unit tests pass" is not a gate criterion; "Integration test pass rate ≥ 100% against all dependency contracts" is
Make gate criteria machine-computable — Manual checklists are fallbacks, not primary mechanisms
Validate environment parity between staging and pre-prod, and between pre-prod and prod — Dev-to-staging parity is less critical
Document the team handoff at each stage boundary — Named approver, required information, and approval criteria
Compute deployment readiness as a verifiable score — Never assume readiness; always compute it from gate results
Treat canary and blue-green as stage transitions — They connect pre-prod to prod with progressive or atomic risk management
Calculate cost vs. safety for each stage — Every stage has an opportunity cost; justify it with risk reduction value

MUST NOT DO

Never allow a stage to pass with partial readiness — 7/8 gates is a blocked deployment, not a "mostly ready" one
Never skip the pre-prod stage to save time — Pre-prod is the last defense before user impact; removing it is gambling
Never use identical gate criteria across multiple stages — If staging and pre-prod check the same thing, you have redundant gates and a missing gate
Never assume dev environment parity with production — Dev differs intentionally; don't force parity there
Never make rollback slower than detection — If you detect a failure in 10 minutes but rollback takes 30 minutes, you have lost
Never replace automated gates with manual "trust" decisions — Human judgment belongs at handoff approval, not gate evaluation
Never add stages without defining their quality dimension — Every stage must earn its place by catching something others miss

Philosophy Alignment

This skill aligns with the 5 Laws of Elegant Defense by treating deployment stages as a layered defense system:

Early Exit

Deployment halts at the first stage gate failure — no progression beyond blocked stages
Gate criteria fail fast: if any single check fails, the stage score drops and deployment is blocked

Parse Don't Validate

Readiness state is parsed from machine-computed results (JSON artifacts, metric queries)
Internal gate logic trusts parsed readiness scores, never re-validates individual checks

Atomic Predictability

Same gate configuration + same readiness state = same decision, every time
Gate evaluation is deterministic: 8/8 gates always means READY, 7/8 always means BLOCKED

Fail Fast

Any gate failure blocks deployment immediately — no progression into uncertain states
Rollback is the default failure mode for prod gates, not a manual afterthought

Intentional Naming

Stage names (dev, staging, preprod, prod) communicate purpose clearly
Gate names map to quality dimensions (unit_test_coverage, chaos_test, error_rate) — each describes what is being validated

Related Skills

Skill	Purpose
`deployment-philosophy`	Higher-level deployment strategy selection (canary vs blue-green vs rolling) — use before defining stage gates
`blue-green-deployment`	Tactical implementation of blue-green as a pre-prod → prod transition — use after defining stage model
`canary-deployment`	Tactical implementation of canary as a progressive pre-prod → prod transition — use after defining stage model
`state-management`	Handles database schema and data migrations across stages — use alongside for stateful deployments
`rollback-strategy`	Designs rollback procedures for each stage — use alongside to define rollback paths and RTOs
`deployment-orchestration`	Coordinates multi-service deployment sequencing across stages — use alongside for complex services

Output Template

When applying this skill, your output should contain:

Stage Model — Number of stages, names, and each stage's primary quality dimension
Gate Criteria — Measurable gate criteria per stage, mapped to quality dimensions
Parity Map — What aspects must match between adjacent stages and what can differ
Handoff Protocol — Approver roles, information packages, and approval criteria at each boundary
Readiness Scorecard — Computed readiness scores for each stage with pass/fail status
Cost-Benefit Summary — Estimated time cost per stage vs. risk reduction value
Transition Strategy — Canary or blue-green selection for prod transition with justification

Example:

## Multi-Stage Deployment Pipeline

**Stage Model:**
  dev (correctness) → staging (integration) → pre-prod (production fit) → prod (user impact)

**Gate Criteria:**
  dev: unit coverage ≥ 80%, 0 critical static issues, code review approved
  staging: 100% integration pass, perf regression ≤ 5%, security clean
  pre-prod: load test within SLO, chaos recovery ≤ 5min, rollback ≤ 60s
  prod: error rate ≤ 1%, latency SLO met, business metric ≤ -5% delta, 30min observation

**Transition Strategy:** Canary (1% → 10% → 50% → 100%) with automated metric gates

**Readiness Scores:**
  dev: 3/3 ✅ | staging: 4/4 ✅ | pre-prod: 3/4 ❌ | prod: N/A

**Blocked at:** pre-prod (chaos_test not passed)

Skill Version: 1.0.0
Created: 2026-05-16
Maturity: stable
Completeness: 95%

Live References

Authoritative documentation links for this skill's domain. The model follows markdown links at load time to resolve external references and inline content.

MAKE IT PART OF YOUR WORKFLOW

安装技能

TERMINAL

npx skills add paulpas/agent-skill-router --skill multi-stage-deployment

复制命令到本地终端执行；安装前请先审查源代码和许可证。

ZIP · 完整技能目录，含资源与许可证

SKILL.md · 仅技能说明原文

仓库 Star: 未同步
许可证: MIT
版本 / 分支: HEAD · d2e2ac6f
仓库更新时间: 未同步

安装前，请检查技能脚本与权限。

源代码