← demos

Agentic: minimal spec → approved architecture (compiling-architecture workflow)

End-to-end demo of the compiling-architecture skill workflow on an enterprise 4-agent hierarchical research-automation system.

Steps 1-2: author writes the agentic part of the spec; the compiler REJECTS it because agentic-jumpstarter--anthropic (a Claude Code Stop-hook workflow pattern) matched against the assumed agentic shape but requires a key-value store that wasn't enabled. The annotated rejection output shows every activation path that brought the pattern in.

Steps 3-7: author iteratively addresses each signal — disallow the unwanted pattern, switch to a platform that fits multi-agent topologies (Dify), then commit to a real cost story to clear the ceiling overrun.

Steps 8-10: review the per-pattern assumed configs (with Options annotations showing alternatives), override the ones that matter for this workload, then promote every assumption into the explicit spec body and prepend the # STATUS: APPROVED comment header. The result is a self-contained architecture.yaml the skills/implementing-architecture skill reads as its input contract.

1 Step 1: starting agentic spec (n8n hosting, 4 agents calling Claude AI models)

The author writes the agentic part of the spec. Headline choices (in the highlight pill):

  • hosting: n8n — the team's existing workflow- automation platform; choice will be revisited in steps 5/6.
  • topology: hierarchical — a research coordinator decomposes queries and dispatches workers; consistent with the agent shape below.
Each agent declares:
  • archetype — orchestrator (coordinator), two extractors (web-searcher + document-analyzer), synthesis (combines findings with provenance).
  • models — both primary AND fallback candidates, per agent. Primary picks the model the agent should normally use (Opus for coordinator + synthesizer; Sonnet for extractors). Fallback degrades to a cheaper / faster tier when the primary fails or is unavailable (Sonnet fallback for the Opus agents; Haiku fallback for the Sonnet agents). The compiler has no opinion about who your fallback should be, so this is always a spec-author decision — declaring both roles up front is good practice.
  • Behavioural toggles: escalation.policy: never, mutating_actions: false, dynamic_ui: false, streaming: false, processing_mode: realtime, session_strategy: stateless. These shape what patterns can activate.
integrations.slack: true declares the only external system the agent talks to. Everything else (caching, async_messaging, NFRs, cost intent, operating_model) is left unspecified — the compiler fills it as assumptions.

spec.yaml (full)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: 'Multi-agent research system. A coordinator decomposes user queries,

    dispatches web-search and document-analysis workers in parallel,

    and a synthesizer combines findings with provenance. Notify on Slack.

    '
constraints:
  cloud: agnostic
  language: python
  platform: api
  features:
    ai_inference: true
    agentic_system:
      hosting: n8n
      topology: hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:
      - agent: research-coordinator
        archetype: orchestrator
        purpose: Decompose research queries, dispatch workers, evaluate synthesis,
          post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor
        purpose: Analyze internal documents for query-relevant content; return findings
          with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
2 Step 2: compile — REJECTED. The compiler shows exactly why.

The compiler rejects the spec based on its assumptions — every unspecified field gets a default value, the full pattern set is evaluated against the combined explicit+assumed view, and the rejection trips on a pattern that activated but couldn't be satisfied. Three things to read carefully in the verbose output:

  1. The annotated input spec body (first section): each line ending in # pattern-id shows which patterns activated on that spec value. Look for # agentic-jumpstarter--anthropic entries — those are the activation gates that brought the rejected pattern in (cloud, agentic_system, archetype, host_platform).
  2. The # ❌ rejection line: the gate the pattern failed — /constraints/features/key_value_store == True. The pattern needs a KV store for phase-state persistence between Stop-hook continuations; the spec didn't enable one.
  3. The 💡 Suggestions trailer: the full list of activation paths so the spec author knows which field to adjust to redirect away from this pattern.
The reader can also see what the compiler ASSUMED as defaults (the assumptions block): caching=false, async_messaging=false, low default throughput, 0-engineer ops team, tight default ceilings, etc. These assumptions feed back into pattern matching too — a different assumption set might activate different patterns.

compiler output (verbose -v)
# ─── ❌ COMPILER REJECTED THIS SPEC ───
❌ Constraints/NFRs trade-off requirements not met:
  [agentic-jumpstarter--anthropic] /constraints/features/key_value_store == True
  → Pattern requires a key-value store for phase state persistence between Stop hook continuations. Without features.key_value_store == true, the supporting infrastructure is missing.

💡 Suggestions — consider changing these activation fields:
  agentic-jumpstarter--anthropic activated by:
    /constraints/features/agentic_system != none
    /constraints/features/agentic_system/agents[*]/archetype contains-any [orchestrator]
    /constraints/features/agentic_system/agents[*]/models[*]/candidates[*]/host_platform contains-any [anthropic]
    /constraints/cloud in [agnostic | aws | azure | gcp | on-prem | nvidia | n/a]


# ─── input spec with pattern-activation annotations ───
# Each `# pattern-id` shows the patterns that activated on this spec value.
  cloud: agnostic  # agentic-jumpstarter--anthropic
    agentic_system:  # agentic-jumpstarter--anthropic
          archetype: orchestrator  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
          archetype: extractor  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
          archetype: extractor  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
          archetype: synthesis  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                - host_platform: anthropic  # agentic-jumpstarter--anthropic

# ─── what the compiler FILLED IN as assumptions ───
assumptions:
  constraints:
    tenantCount: 1
    features:
      caching: false
      async_messaging: false
      multi_tenancy: false
      batch_processing: false
      distributed_transactions: false
      real_time_streaming: false
      vector_search: false
      document_store: false
      key_value_store: false  # ❌ agentic-jumpstarter--anthropic
      graph_database: false
      time_series_db: false
      oltp_workload: true
      olap_workload: false
      cold_archive_tiering: false
      messaging_delivery_guarantee: null
  nfr:
    rpo_minutes: 60
    rto_minutes: 60
    latency:
      p95Milliseconds: 500
      p99Milliseconds: 1000
      jobStartP95Seconds: null
      jobStartP99Seconds: null
    throughput:
      peak_jobs_per_hour: null
      peak_query_per_second_read: 5
      peak_query_per_second_write: 1
    data:
      retention_days: 90
      pii: false
      compliance:
    consistency:
      needsReadYourWrites: false
    durability:
      strict: false
    security:
    # …  (more defaults below; expand the full output to see them)
see full compiled-spec.yaml (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: "Multi-agent research system. A coordinator decomposes user queries,
dispatches web-search and document-analysis workers in parallel,
and a synthesizer combines findings with provenance. Notify on Slack.
"
constraints:
  cloud: agnostic  # agentic-jumpstarter--anthropic
  language: python
  platform: api
  features:
    ai_inference: true
    agentic_system:  # agentic-jumpstarter--anthropic
      hosting: n8n
      topology: hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:
        - agent: research-coordinator
          archetype: orchestrator  # agentic-jumpstarter--anthropic
          purpose: Decompose research queries, dispatch workers, evaluate synthesis, post final report.
          escalation:
            policy: never
          mutating_actions: false
          dynamic_ui: false
          streaming: false
          processing_mode: realtime
          session_strategy: stateless
          models:
            - role: primary
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-opus
                  model_id: claude-opus-4-7
            - role: fallback
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-sonnet
                  model_id: claude-sonnet-4-6
        - agent: web-searcher
          archetype: extractor  # agentic-jumpstarter--anthropic
          purpose: Search the web for query subtopics, return findings with source attribution.
          escalation:
            policy: never
          mutating_actions: false
          dynamic_ui: false
          streaming: false
          processing_mode: realtime
          session_strategy: stateless
          models:
            - role: primary
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-sonnet
                  model_id: claude-sonnet-4-6
            - role: fallback
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-haiku
                  model_id: claude-haiku-4-5
        - agent: document-analyzer
          archetype: extractor  # agentic-jumpstarter--anthropic
          purpose: Analyze internal documents for query-relevant content; return findings with source attribution.
          escalation:
            policy: never
          mutating_actions: false
          dynamic_ui: false
          streaming: false
          processing_mode: realtime
          session_strategy: stateless
          models:
            - role: primary
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-sonnet
                  model_id: claude-sonnet-4-6
            - role: fallback
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-haiku
                  model_id: claude-haiku-4-5
        - agent: synthesizer
          archetype: synthesis  # agentic-jumpstarter--anthropic
          purpose: Combine findings into a coherent report preserving claim-level provenance.
          escalation:
            policy: never
          mutating_actions: false
          dynamic_ui: false
          streaming: false
          processing_mode: realtime
          session_strategy: stateless
          models:
            - role: primary
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-opus
                  model_id: claude-opus-4-7
            - role: fallback
              candidates:
                - host_platform: anthropic  # agentic-jumpstarter--anthropic
                  model_owner: anthropic
                  model_name: claude-sonnet
                  model_id: claude-sonnet-4-6
assumptions:
  constraints:
    saas-providers:
    disallowed-saas-providers:
    ai-inference-platforms:
    disallowed-ai-inference-platforms:
    model-vendors:
    disallowed-model-vendors:
    tenantCount: 1
    features:
      caching: false
      async_messaging: false
      multi_tenancy: false
      batch_processing: false
      distributed_transactions: false
      real_time_streaming: false
      vector_search: false
      document_store: false
      key_value_store: false  # ❌ agentic-jumpstarter--anthropic
      graph_database: false
      time_series_db: false
      oltp_workload: true
      olap_workload: false
      cold_archive_tiering: false
      messaging_delivery_guarantee: null
  nfr:
    availability:
      target: 0.95
    rpo_minutes: 60
    rto_minutes: 60
    latency:
      p95Milliseconds: 500
      p99Milliseconds: 1000
      jobStartP95Seconds: null
      jobStartP99Seconds: null
    throughput:
      peak_jobs_per_hour: null
      peak_query_per_second_read: 5
      peak_query_per_second_write: 1
    data:
      retention_days: 90
      pii: false
      compliance:
        gdpr: false
        gdpr_rtbf: false
        ccpa: false
        hipaa: false
        sox: false
    consistency:
      needsReadYourWrites: false
    durability:
      strict: false
    security:
      auth: oauth2_oidc
      tenant_isolation: n/a
      audit_logging: false
    agentic_system:
      target:
        prompt_injection_block_rate_min: 0.95
        output_safety_pass_rate_min: 0.99
        pii_leak_rate_max: 0.01
        jailbreak_detection_rate_min: 0.9
        hallucination_rate_max: 0.1
        toxic_output_rate_max: 0.005
  operating_model:
    on_call: false
    deploy_freq: weekly
    ops_team_size: 0
    single_resource_monthly_ops_usd: 10000
    amortization_months: 24
  cost:
    intent:
      priority: minimize-opex
    ceilings:
      monthly_operational_usd: 500
      one_time_setup_usd: 1000
    preferences:
      prefer_free_tier_if_possible: true
      prefer_saas_first: false
  patterns:
    meta:
      agentic-multi-agent-required:
      agentic-system-required:
    P0:
      arch-serverless:
        function_runtime_model: faas
        cold_start_optimization: minimal
        state_management: external
        event_sources: api_gateway
        observability_level: standard
      db-managed-postgres:
        provider: supabase
        instance_size: small
        storage_gb: 8
        backup_retention_days: 7
        connection_pooling: true
        high_availability: false
        ssl_mode: require
      arch-serverless-pay-per-use:
        compute_type: function
        memory_size_mb: 1024
        timeout_seconds: 30
        cold_start_optimization: standard
        concurrency_model: auto
      agentic-topology--hierarchical:
        message_envelope_shape: structured_json
        response_envelope_shape: structured_json
        failure_policy_retry_max: 2
        failure_policy_retry_backoff: exponential
        failure_policy_retry_initial_ms: 1000
        failure_policy_on_terminal_failure: surface_partial
        failure_policy_cancel_siblings_on_failure: false
        aggregation_strategy: concat_with_attribution
      agentic-jumpstarter--anthropic:
      agentic-session--stateless:
      hosting-agentic-workflow-automation--n8n:
      agentic-model-vendor--anthropic:
        enable_token_efficient_tool_use_beta: true
        prompt_caching_recommended_blocks:
          - system_prompt
          - tool_definitions
          - long_context_documents
      agentic-inference--anthropic:
        enable_token_efficient_tool_use_beta: true
      iac-terraform:
        state_backend: s3
        workspace_strategy: single
        module_structure: monorepo
        ci_cd_integration: github-actions
        drift_detection: false
    P1:
      agentic-output-safety-guardrail:
        detection_action: redact-and-continue
      obs-agentic-system:
        log_full_prompts: false
      agentic-input-safety-guardrail:
        fallback_on_guardrail_failure: fail-closed
        block_action: reject-with-explanation
      agentic-tool-scoped-credentials:
        fail_action_on_credential_error: fail-closed
      agentic-multi-pass-review:
        review_session_isolation: fully_independent
      api-rest-resource-oriented:
        pagination_style: offset
        max_page_size: 100
        versioning_strategy: uri
        filtering_style: query_params
        cache_strategy: etag
        id_format: uuid
        response_envelope: false
      agentic-validation-retry-loop:
      sync-request-reply-rest:
        timeout_seconds: 30
        retry_strategy: exponential_backoff
        max_retries: 3
        circuit_breaker_enabled: true
        rate_limiting_strategy: token_bucket
        idempotency_required: false
      agentic-tool-use-extraction:
        composes:
          layered_after:
            - agentic-output-safety-guardrail
      deploy-rolling:
        batch_size: 25_percent
        health_check_grace_period: 30
        rollback_on_failure: true
        deployment_strategy: replace_one_by_one
        max_unavailable: 0_percent
        composes:
          layered_after:
            - iac-terraform
      sec-auth-oauth2-oidc:
        oauth_flow: authorization_code
        token_storage: secure_storage
        pkce_enabled: true
        scope_strategy: minimal
        token_refresh: automatic
        id_token_validation: strict
        composes:
          wraps:
            - api-rest-resource-oriented
      crud-single-model:
        api_style: rest
        validation_strategy: server-side
        soft_delete: false
        audit_logging: false
        pagination_default_size: 20
      finops-cost-allocation-tags:
        tagging_strategy: hierarchical
        enforcement_level: required
        cost_allocation_model: showback
        tag_inheritance: true
        automated_tagging: true
      release-feature-flags:
        flag_storage: config_file
        evaluation_strategy: simple_boolean
        targeting_capability: none
        kill_switch_enabled: true
        audit_logging: false
      obs-telemetry-backend:
        metrics_retention_days: 15
        log_retention_days: 7
        backend_ha: false
      obs-open-telemetry-baseline:
        export_backend: otlp
        trace_sampling_strategy: parent-based
        trace_sampling_rate: 1.0
        metrics_export_interval: 60
        log_correlation: true
        resource_detection: true
        propagation_format: w3c-tracecontext
        composes:
          co_runs_with:
            - api-rest-resource-oriented
      secrets-env-vars:
        injection_method: runtime
        secret_rotation_support: false
        prefix_convention: APP_SECRET_
        validation_on_startup: true
      finops-budget-guardrails:
        budget_period: monthly
        alert_thresholds:
          - 50
          - 80
          - 100
        enforcement_action: alert
        tagging_strategy: mandatory
        policy_enforcement: soft
        cost_allocation_level: project
      ops-slo-error-budgets:
        slo_target_percentage: 99.9
        measurement_window_days: 30
        error_budget_policy: halt-deployments
        sli_type: availability
        alerting_threshold_percentage: 80
      agentic-provenance-tracking:
    P2:
      api-versioning-header:
        version_header_name: API-Version
        version_format: date-based
        fallback_behavior: latest-stable
        content_negotiation: false
        deprecation_policy: warning-header
      gov-system-manifest:
        manifest_path: docs/architecture/manifest.yaml
        manifest_format: yaml
        manifest_scope:
          - agent-tools
          - agent-skills
          - agent-models
          - agent-prompts
        pin_versions: true
        ci_validation: required
        drift_policy: fail-build
        composes:
          layered_after:
            - iac-terraform
          co_runs_with:
            - release-feature-flags
            - gov-adrs-mandatory
            - ops-runbooks
    P3:
      ops-runbooks:
        runbook_format: markdown
        incident_severity_levels: 4
        escalation_policy: tiered
        automation_integration: manual
        review_frequency: quarterly
      gov-adrs-mandatory:
        adr_format: madr
        storage_location: docs/adrs
        decision_threshold: significant
        review_requirement: peer-review

❌ Constraints/NFRs trade-off requirements not met:
  [agentic-jumpstarter--anthropic] /constraints/features/key_value_store == True
  → Pattern requires a key-value store for phase state persistence between Stop hook continuations. Without features.key_value_store == true, the supporting infrastructure is missing.

💡 Suggestions — consider changing these activation fields:
  agentic-jumpstarter--anthropic activated by:
    /constraints/features/agentic_system != none
    /constraints/features/agentic_system/agents[*]/archetype contains-any [orchestrator]
    /constraints/features/agentic_system/agents[*]/models[*]/candidates[*]/host_platform contains-any [anthropic]
    /constraints/cloud in [agnostic | aws | azure | gcp | on-prem | nvidia | n/a]
Rejection narrative: read top-down — the annotated body shows which spec fields the rejected pattern matched (activation gates), the # ❌ line shows the gate it failed, the 💡 Suggestions trailer tells you what to change. In this case the team doesn't run long-autonomous Stop-hook workflows, so the right move isn't to add a KV store — it's to disallow the pattern (step 3).
3 Step 3: disallow agentic-jumpstarter--anthropic

The team doesn't run long-autonomous Stop-hook workflows, so they explicitly tell the compiler to never consider that pattern by adding it to disallowed-patterns. This is the spec author's escape hatch when a pattern matches the shape but isn't operationally a fit. Expand the inline-spec disclosure below to see the spec at this stage.

spec.yaml (delta added in this step)
disallowed-patterns:
- agentic-jumpstarter--anthropic
view full spec.yaml at this step (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: 'Multi-agent research system. A coordinator decomposes user queries,

    dispatches web-search and document-analysis workers in parallel,

    and a synthesizer combines findings with provenance. Notify on Slack.

    '
constraints:
  cloud: agnostic
  language: python
  platform: api
  features:
    ai_inference: true
    agentic_system:
      hosting: n8n
      topology: hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:
      - agent: research-coordinator
        archetype: orchestrator
        purpose: Decompose research queries, dispatch workers, evaluate synthesis,
          post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor
        purpose: Analyze internal documents for query-relevant content; return findings
          with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
disallowed-patterns:
- agentic-jumpstarter--anthropic
4 Step 4: recompile — spec accepted. Two warnings surface.

No more rejection. The compiler accepts the spec and produces a complete pattern selection. Verbose mode adds two kinds of information beyond plain compile: (a) every spec field gets an inline # pattern-id, pattern-id, ... annotation listing the patterns that activated on it; (b) the rejected-patterns.yaml side file gets per-pattern reasoning for everything the compiler considered and dropped. Two issues still need attention:

  • warn_constraints from agentic-multi-agent-required — n8n is workflow-automation, not a multi-agent runtime.
  • [high] cost_opex_exceeds_ceiling — assumed ceilings don't fit this pattern set.
Expand the full compile output below to see every selected pattern and trace why it was picked.

compiler output (verbose -v)
# ─── what the compiler FILLED IN as assumptions ───
assumptions:
  constraints:
    tenantCount: 1
    features:
      caching: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (19 more)
      async_messaging: false # arch-serverless, agentic-output-safety-guardrail, obs-agentic-system, ... (20 more)
      multi_tenancy: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      batch_processing: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (12 more)
      distributed_transactions: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      real_time_streaming: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      vector_search: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      document_store: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      key_value_store: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      graph_database: false # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
      time_series_db: false # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
      oltp_workload: true # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      olap_workload: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      cold_archive_tiering: false
  nfr:
    rpo_minutes: 60  # secrets-env-vars, finops-budget-guardrails
    rto_minutes: 60 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
    latency:
      p95Milliseconds: 500  # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
      p99Milliseconds: 1000 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
    throughput:
      peak_query_per_second_read: 5  # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
      peak_query_per_second_write: 1 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (3 more)
    data:
      retention_days: 90
      pii: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      compliance:
    consistency:
      needsReadYourWrites: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (15 more)
    durability:
      strict: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
    security:
      auth: oauth2_oidc  # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (15 more)
      tenant_isolation: n/a # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      audit_logging: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
  operating_model:
    # …  (more defaults below; expand the full output to see them)

# ─── Matched Patterns based on input spec ───
# meta  = policy gates (always emitted when their feature flag is set)
# P0    = high priority — load-bearing architectural decisions
# P1    = mid priority — operational + observability + security baseline
# P2/P3 = lower priority — refinements + governance + docs
# Override priority by adding `patterns.<id>.recommended_priority: P0` to spec.
  patterns:
    meta:    # (2 patterns)
      - agentic-multi-agent-required  # Meta-policy pattern that activates when an agentic system has 2 or more agents and enforces topology specification AND surfaces multi-agent-specific risk warnings.
      - agentic-system-required  # Meta-policy pattern that activates when agentic_system is enabled and enforces that the spec's agentic-system fields are correctly populated and reference valid platforms.
    P0:    # (9 patterns)
      - arch-serverless  # Structure the system as stateless, event-driven function handlers backed by managed services. No persistent servers — each function activates on demand, executes, and terminates. The architectural commitment is to build around events and cloud-managed primitives rather than long-running processes.
      - db-managed-postgres  # Use low-ops managed Postgres DBaaS providers (e.g., Supabase and managed cloud Postgres offerings) to reduce DB operations overhead; validate quotas, compliance, and scale limits.
      - arch-serverless-pay-per-use  # Eliminate idle infrastructure cost by running compute on a pay-per-execution billing model rather than paying for always-on instances. Well-suited to bursty or unpredictable workloads where provisioned servers would sit idle most of the time. This is a cost optimisation pattern, not an architectural style — it can apply to any workload that tolerates cold starts.
      - agentic-topology--hierarchical  # Central coordinator/orchestrator agent dynamically decomposes tasks, delegates to specialized worker agents (extractors, analyzers, synthesizers), and combines results.
      - agentic-session--stateless  # Each request is independent; the agent does not persist any conversation context between invocations.
      - hosting-agentic-workflow-automation--n8n  # n8n workflow orchestration platform with AI Agent node and tool nodes.
      - agentic-model-vendor--anthropic  # Architectural pattern for systems whose models are built by Anthropic (the Claude family: Opus, Sonnet, Haiku).
      - agentic-inference--anthropic  # LLM inference via Anthropic's direct API (api.anthropic.com).
      - iac-terraform  # Manage infra declaratively with Terraform modules, state, and pipelines.
    P1:    # (20 patterns)
      - agentic-output-safety-guardrail  # Post-LLM scan of agent outputs before delivery to users or external systems.
      - obs-agentic-system  # Production observability tailored for agentic systems. Adds agent-specific tracing (LLM call inputs/outputs, tool dispatch, step graph), cost attribution per archetype/tenant/task, token budget accounting, eval-metric streams, and agent lifecycle metrics. Composes on top of obs-open-telemetry-baseline. Selected unconditionally when agentic_system is declared.
      - agentic-input-safety-guardrail  # Pre-flight inspection of user inputs and tool-returned content before they reach the agent's LLM.
      - agentic-tool-scoped-credentials  # Each tool call uses a credential narrowly scoped to that tool's operation, not a shared agent-wide credential.
      - agentic-multi-pass-review  # Architecture where one agent session generates output and a separate, independent session reviews it.
      - api-rest-resource-oriented  # REST API designed around resources (nouns) manipulated via standard HTTP verbs (GET, POST, PUT, DELETE, PATCH). Resources are identified by stable URLs, responses are cacheable by default, and pagination/filtering are expressed as query parameters. Simpler tooling and stronger HTTP cache semantics than GraphQL; well-suited to public APIs and CRUD-heavy domains.
      - agentic-validation-retry-loop  # Layered pattern that wraps any agent output with deterministic validation and retries with specific error feedback.
      - sync-request-reply-rest  # Synchronous HTTP APIs; simple integration; needs timeouts/retries/backpressure.
      - agentic-tool-use-extraction  # Single-shot structured data extraction using the tool_use mechanism with a forced tool_choice.
          composes:
            layered_after: ['agentic-output-safety-guardrail']
      - deploy-rolling  # Gradually replace instances; simplest for stateless services; requires backward-compatible changes.
          composes:
            layered_after: ['iac-terraform']
      - sec-auth-oauth2-oidc  # Use OAuth2 flows with OIDC identity tokens; standardized claims; delegated auth support.
          composes:
            wraps: ['api-rest-resource-oriented']
      - crud-single-model  # Simple CRUD on one canonical model; lowest complexity; best for straightforward domains.
      - finops-cost-allocation-tags  # Tagging/labeling strategy for per-tenant/product cost allocation and chargeback/showback.
      - release-feature-flags  # Decouple deploy from release; safer experiments; needs kill switches and governance.
      - obs-telemetry-backend  # Self-hosted observability backend combining Prometheus (metrics storage and alerting), Grafana (dashboards), and Loki (structured log aggregation). Provides the storage and query layer for traces, metrics, and logs; consumed by instrumentation and monitoring strategy patterns. Preferred when data residency, cost control, or customization requirements preclude a managed SaaS observability backend.
      - obs-open-telemetry-baseline  # Standardize traces/metrics/log correlation via OpenTelemetry; export to vendor or OSS backends.
          composes:
            co_runs_with: ['api-rest-resource-oriented']
      - secrets-env-vars  # Inject secrets as env vars; simple but risk of exposure; requires strict process controls.
      - finops-budget-guardrails  # Implement budgets, alerts, tagging, and policy-as-code guardrails to enforce cost ceilings and prevent runaway spend.
      - ops-slo-error-budgets  # Define SLOs and error budgets to balance reliability and velocity.
      - agentic-provenance-tracking  # Architectural pattern for synthesis agents that combine findings from multiple sources into a unified output.
    P2:    # (2 patterns)
      - api-versioning-header  # Version via headers/media types; keeps URLs stable; harder to debug and cache.
      - gov-system-manifest  # Pin and govern the inventory of components (agent-tools, agent-skills, agent-models, agent-prompts, services, data sources, external dependencies) the system depends on at a declared manifest path; CI validates on every PR and drift between manifest and built system fails the build.
          composes:
            layered_after: ['iac-terraform']
            co_runs_with: ['release-feature-flags', 'gov-adrs-mandatory', 'ops-runbooks']
    P3:    # (2 patterns)
      - ops-runbooks  # Standard runbooks for incidents and routine ops; reduces MTTR and on-call stress.
      - gov-adrs-mandatory  # Record architecture decisions and tradeoffs; improves continuity; keep lightweight.

# ─── warns and cost feasibility ───
# ============================================================
# Cost Feasibility Analysis (Summary)
# ============================================================
#
# Intent: minimize-opex
# Amortization: 24 months
# Total Patterns Selected: 35
#
# COST BREAKDOWN:
# ────────────────────────────────────────────────────────────
# Total CapEx (one-time):     $      80,450
# Pattern OpEx (monthly):     $         610
# Ops Team Cost (monthly):    $           0
# Total OpEx (monthly):       $         610
# Total TCO (24mo):         $      95,090
#
# COST CEILINGS:
# ────────────────────────────────────────────────────────────
# CapEx Ceiling:              $       1,000 ✗ FAIL
# OpEx Ceiling (monthly):     $         500 ✗ FAIL
#
# ⚠️  WARNINGS:
# ────────────────────────────────────────────────────────────
# [high] cost_opex_exceeds_ceiling:
#   Total monthly operational cost ($610) exceeds ceiling ($500) by $110 (intent: minimize-opex)
#
#   Suggestions:
#   - Increase monthly ceiling to $610 or more
#   - Remove high-cost patterns: ops-slo-error-budgets, agentic-topology--hierarchical, obs-agentic-system
#   - Reduce ops_team_size from 0
#
# ============================================================

# ============================================================
# ⚠️  Pattern Advisory Warnings
# (Patterns are still SELECTED — review these before finalizing)
# ============================================================
#
# [warning] warn_constraints:
#   agentic-multi-agent-required: Spec hosts a multi-agent system on a workflow-automation platform. These platforms model linear/branching node graphs, not multi-agent orchestration — they lack structured inter-agent envelopes, parallel agent execution with aggregation, hierarchical orchestration, and shared agent state. Sequential 2-agent flows can be expressed as LLM-node chains; hierarchical/parallel/routing/hybrid topologies typically cannot. Consider hosting-agentic-container (any cloud) or a managed agent platform (hosting-agentic-managed-ai-platform--aws-bedrock-agentcore / vertex-agent-builder / azure-ai-foundry-agents).
#
#   Suggestions:
#   - agentic-multi-agent-required: Spec hosts a multi-agent system on a workflow-automation platform. These platforms model linear/branching node graphs, not multi-agent orchestration — they lack structured inter-agent envelopes, parallel agent execution with aggregation, hierarchical orchestration, and shared agent state. Sequential 2-agent flows can be expressed as LLM-node chains; hierarchical/parallel/routing/hybrid topologies typically cannot. Consider hosting-agentic-container (any cloud) or a managed agent platform (hosting-agentic-managed-ai-platform--aws-bedrock-agentcore / vertex-agent-builder / azure-ai-foundry-agents).
#
# ============================================================
see full compiled-spec.yaml (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: "Multi-agent research system. A coordinator decomposes user queries,\ndispatches web-search and document-analysis workers in parallel,\nand a synthesizer combines findings with provenance. Notify on Slack.\n"
constraints:
  cloud: agnostic  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
  language: python # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
  platform: api # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
  features:
    ai_inference: true  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (11 more)
    agentic_system: # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (12 more)
      hosting: n8n  # hosting-agentic-workflow-automation--n8n
      topology: hierarchical # agentic-topology--hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:  # agentic-topology--hierarchical, agentic-multi-agent-required
      - agent: research-coordinator
        archetype: orchestrator  # agentic-multi-pass-review
        purpose: Decompose research queries, dispatch workers, evaluate synthesis, post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor  # agentic-multi-pass-review, agentic-validation-retry-loop, agentic-tool-use-extraction
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor  # agentic-multi-pass-review, agentic-validation-retry-loop, agentic-tool-use-extraction
        purpose: Analyze internal documents for query-relevant content; return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis  # agentic-provenance-tracking
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
disallowed-patterns:
- agentic-jumpstarter--anthropic
assumptions:
  constraints:
    saas-providers: []
    disallowed-saas-providers: []
    ai-inference-platforms: []
    disallowed-ai-inference-platforms: []
    model-vendors: []
    disallowed-model-vendors: []
    tenantCount: 1
    features:
      caching: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (19 more)
      async_messaging: false # arch-serverless, agentic-output-safety-guardrail, obs-agentic-system, ... (20 more)
      multi_tenancy: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      batch_processing: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (12 more)
      distributed_transactions: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      real_time_streaming: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      vector_search: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      document_store: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      key_value_store: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      graph_database: false # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
      time_series_db: false # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
      oltp_workload: true # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      olap_workload: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      cold_archive_tiering: false
  nfr:
    availability:
      target: 0.95  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (12 more)
    rpo_minutes: 60  # secrets-env-vars, finops-budget-guardrails
    rto_minutes: 60 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
    latency:
      p95Milliseconds: 500  # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
      p99Milliseconds: 1000 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
    throughput:
      peak_query_per_second_read: 5  # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
      peak_query_per_second_write: 1 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (3 more)
    data:
      retention_days: 90
      pii: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      compliance:
        gdpr: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
        gdpr_rtbf: false # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (1 more)
        ccpa: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
        hipaa: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
        sox: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
    consistency:
      needsReadYourWrites: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (15 more)
    durability:
      strict: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
    security:
      auth: oauth2_oidc  # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (15 more)
      tenant_isolation: n/a # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      audit_logging: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
    agentic_system:
      target:
        prompt_injection_block_rate_min: 0.95  # agentic-input-safety-guardrail
        output_safety_pass_rate_min: 0.99 # agentic-output-safety-guardrail, obs-agentic-system
        pii_leak_rate_max: 0.01 # agentic-output-safety-guardrail
        jailbreak_detection_rate_min: 0.9 # agentic-input-safety-guardrail
        hallucination_rate_max: 0.1 # obs-agentic-system, agentic-multi-pass-review
        toxic_output_rate_max: 0.005 # agentic-output-safety-guardrail
  operating_model:
    on_call: false
    deploy_freq: weekly
    ops_team_size: 0
    single_resource_monthly_ops_usd: 10000
    amortization_months: 24
  cost:
    intent:
      priority: minimize-opex
    ceilings:
      monthly_operational_usd: 500
      one_time_setup_usd: 1000
    preferences:
      prefer_free_tier_if_possible: true  # db-managed-postgres, agentic-output-safety-guardrail, obs-agentic-system, ... (12 more)
      prefer_saas_first: false
  patterns:
    meta:
      agentic-multi-agent-required: {}  # Meta-policy pattern that activates when an agentic system has 2 or more agents and enforces topology specification AND surfaces multi-agent-specific risk warnings.
      agentic-system-required: {} # Meta-policy pattern that activates when agentic_system is enabled and enforces that the spec's agentic-system fields are correctly populated and reference valid platforms.
    P0:
      arch-serverless:  # Structure the system as stateless, event-driven function handlers backed by managed services. No persistent servers — each function activates on demand, executes, and terminates. The architectural commitment is to build around events and cloud-managed primitives rather than long-running processes.
        function_runtime_model: faas  # Options: faas, container_functions, hybrid
        cold_start_optimization: minimal # Options: minimal, provisioned_concurrency, keep_warm
        state_management: external # Options: external, durable_functions, step_functions
        event_sources: api_gateway # Options: api_gateway, event_bus, stream_processing, multi_trigger
        observability_level: standard # Options: standard, enhanced, distributed_tracing
      db-managed-postgres: # Use low-ops managed Postgres DBaaS providers (e.g., Supabase and managed cloud Postgres offerings) to reduce DB operations overhead; validate quotas, compliance, and scale limits.
        provider: supabase  # Options: supabase, neon, render, railway, digitalocean-app-platform
        instance_size: small # Options: micro, small, medium, large
        storage_gb: 8 # Range: 1-500
        backup_retention_days: 7 # Range: 1-30
        connection_pooling: true # Boolean
        high_availability: false # Boolean
        ssl_mode: require # Options: disable, allow, prefer, require, verify-ca, verify-full
      arch-serverless-pay-per-use: # Eliminate idle infrastructure cost by running compute on a pay-per-execution billing model rather than paying for always-on instances. Well-suited to bursty or unpredictable workloads where provisioned servers would sit idle most of the time. This is a cost optimisation pattern, not an architectural style — it can apply to any workload that tolerates cold starts.
        compute_type: function  # Options: function, container
        memory_size_mb: 1024 # Options: 128, 256, 512, 1024, 2048, 4096
        timeout_seconds: 30 # Range: 3-900
        cold_start_optimization: standard # Options: standard, provisioned, warm-pool
        concurrency_model: auto # Options: auto, reserved, limited
      agentic-topology--hierarchical: # Central coordinator/orchestrator agent dynamically decomposes tasks, delegates to specialized worker agents (extractors, analyzers, synthesizers), and combines results.
        message_envelope_shape: structured_json  # Options: structured_json, freeform_string, hybrid
        response_envelope_shape: structured_json # Options: structured_json, freeform_string, hybrid
        failure_policy_retry_max: 2 # Min: 0
        failure_policy_retry_backoff: exponential # Options: exponential, linear, immediate, none
        failure_policy_retry_initial_ms: 1000 # Min: 0
        failure_policy_on_terminal_failure: surface_partial # Options: surface_partial, hard_fail, escalate_to_human
        failure_policy_cancel_siblings_on_failure: false # Boolean
        aggregation_strategy: concat_with_attribution # Options: sequential_pipe, concat_with_attribution, merge, last_wins, voting_quorum, judge_arbitration, none
      agentic-session--stateless: {} # Each request is independent; the agent does not persist any conversation context between invocations.
      hosting-agentic-workflow-automation--n8n: {} # n8n workflow orchestration platform with AI Agent node and tool nodes.
      agentic-model-vendor--anthropic: # Architectural pattern for systems whose models are built by Anthropic (the Claude family: Opus, Sonnet, Haiku).
        enable_token_efficient_tool_use_beta: true  # Boolean
        prompt_caching_recommended_blocks: # Options: system_prompt, tool_definitions, long_context_documents, few_shot_examples
        - system_prompt
        - tool_definitions
        - long_context_documents
      agentic-inference--anthropic: # LLM inference via Anthropic's direct API (api.anthropic.com).
        enable_token_efficient_tool_use_beta: true  # Boolean
      iac-terraform: # Manage infra declaratively with Terraform modules, state, and pipelines.
        state_backend: s3  # Options: s3, terraform-cloud, azurerm, gcs, local
        workspace_strategy: single # Options: single, per-environment, per-service
        module_structure: monorepo # Options: monorepo, separate-repos, registry
        ci_cd_integration: github-actions # Options: github-actions, gitlab-ci, terraform-cloud, atlantis, manual
        drift_detection: false # Boolean
    P1:
      agentic-output-safety-guardrail:  # Post-LLM scan of agent outputs before delivery to users or external systems.
        detection_action: redact-and-continue  # Options: redact-and-continue, block-and-explain, block-silent, human-review, regenerate
      obs-agentic-system: # Production observability tailored for agentic systems. Adds agent-specific tracing (LLM call inputs/outputs, tool dispatch, step graph), cost attribution per archetype/tenant/task, token budget accounting, eval-metric streams, and agent lifecycle metrics. Composes on top of obs-open-telemetry-baseline. Selected unconditionally when agentic_system is declared.
        log_full_prompts: false  # Boolean
      agentic-input-safety-guardrail: # Pre-flight inspection of user inputs and tool-returned content before they reach the agent's LLM.
        fallback_on_guardrail_failure: fail-closed  # Options: fail-open, fail-closed
        block_action: reject-with-explanation # Options: reject-with-explanation, reject-silent, sanitize-and-continue, human-review
      agentic-tool-scoped-credentials: # Each tool call uses a credential narrowly scoped to that tool's operation, not a shared agent-wide credential.
        fail_action_on_credential_error: fail-closed  # Options: fail-closed, fail-open, fallback-to-shared
      agentic-multi-pass-review: # Architecture where one agent session generates output and a separate, independent session reviews it.
        review_session_isolation: fully_independent  # Options: fully_independent, shared_context_acceptable
      api-rest-resource-oriented: # REST API designed around resources (nouns) manipulated via standard HTTP verbs (GET, POST, PUT, DELETE, PATCH). Resources are identified by stable URLs, responses are cacheable by default, and pagination/filtering are expressed as query parameters. Simpler tooling and stronger HTTP cache semantics than GraphQL; well-suited to public APIs and CRUD-heavy domains.
        pagination_style: offset  # Options: offset, cursor, page_number
        max_page_size: 100 # Range: 10-1000
        versioning_strategy: uri # Options: uri, header, query_param, none
        filtering_style: query_params # Options: query_params, json_body, graphql_like
        cache_strategy: etag # Options: etag, last_modified, cache_control, none
        id_format: uuid # Options: uuid, integer, slug, composite
        response_envelope: false # Boolean
      agentic-validation-retry-loop: {} # Layered pattern that wraps any agent output with deterministic validation and retries with specific error feedback.
      sync-request-reply-rest: # Synchronous HTTP APIs; simple integration; needs timeouts/retries/backpressure.
        timeout_seconds: 30  # Range: 1-300
        retry_strategy: exponential_backoff # Options: none, fixed_delay, exponential_backoff, exponential_backoff_jitter
        max_retries: 3 # Range: 0-10
        circuit_breaker_enabled: true # Boolean
        rate_limiting_strategy: token_bucket # Options: none, token_bucket, leaky_bucket, fixed_window, sliding_window
        idempotency_required: false # Boolean
      agentic-tool-use-extraction: # Single-shot structured data extraction using the tool_use mechanism with a forced tool_choice.
        composes:
          layered_after:
          - agentic-output-safety-guardrail
      deploy-rolling: # Gradually replace instances; simplest for stateless services; requires backward-compatible changes.
        batch_size: 25_percent  # Options: 10_percent, 25_percent, 50_percent, one_at_a_time, all_at_once
        health_check_grace_period: 30 # Range: 10-300
        rollback_on_failure: true # Boolean
        deployment_strategy: replace_one_by_one # Options: replace_one_by_one, rolling_batch, canary_first
        max_unavailable: 0_percent # Options: 0_percent, 10_percent, 25_percent
        composes:
          layered_after:
          - iac-terraform
      sec-auth-oauth2-oidc: # Use OAuth2 flows with OIDC identity tokens; standardized claims; delegated auth support.
        oauth_flow: authorization_code  # Options: authorization_code, client_credentials, device_code, implicit
        token_storage: secure_storage # Options: secure_storage, memory_only, encrypted_storage, httponly_cookie
        pkce_enabled: true # Boolean
        scope_strategy: minimal # Options: minimal, role_based, resource_specific
        token_refresh: automatic # Options: automatic, manual, sliding_window
        id_token_validation: strict # Options: strict, standard, relaxed
        composes:
          wraps:
          - api-rest-resource-oriented
      crud-single-model: # Simple CRUD on one canonical model; lowest complexity; best for straightforward domains.
        api_style: rest  # Options: rest, graphql, rpc
        validation_strategy: server-side # Options: server-side, client-side, both
        soft_delete: false # Boolean
        audit_logging: false # Boolean
        pagination_default_size: 20 # Range: 10-100
      finops-cost-allocation-tags: # Tagging/labeling strategy for per-tenant/product cost allocation and chargeback/showback.
        tagging_strategy: hierarchical  # Options: hierarchical, flat, hybrid
        enforcement_level: required # Options: required, recommended, optional
        cost_allocation_model: showback # Options: chargeback, showback, hybrid
        tag_inheritance: true # Boolean
        automated_tagging: true # Boolean
      release-feature-flags: # Decouple deploy from release; safer experiments; needs kill switches and governance.
        flag_storage: config_file  # Options: config_file, database, feature_flag_service, environment_variables
        evaluation_strategy: simple_boolean # Options: simple_boolean, percentage_rollout, user_targeting, multi_variate
        targeting_capability: none # Options: none, user_attributes, context_based, advanced_segments
        kill_switch_enabled: true # Boolean
        audit_logging: false # Boolean
      obs-telemetry-backend: # Self-hosted observability backend combining Prometheus (metrics storage and alerting), Grafana (dashboards), and Loki (structured log aggregation). Provides the storage and query layer for traces, metrics, and logs; consumed by instrumentation and monitoring strategy patterns. Preferred when data residency, cost control, or customization requirements preclude a managed SaaS observability backend.
        metrics_retention_days: 15  # Options: 7, 15, 30, 90, 365
        log_retention_days: 7 # Options: 1, 3, 7, 14, 30
        backend_ha: false # Boolean
      obs-open-telemetry-baseline: # Standardize traces/metrics/log correlation via OpenTelemetry; export to vendor or OSS backends.
        export_backend: otlp  # Options: otlp, jaeger, zipkin, prometheus, datadog, newrelic, honeycomb
        trace_sampling_strategy: parent-based # Options: always-on, always-off, parent-based, trace-id-ratio
        trace_sampling_rate: 1.0 # Range: 0.0-1.0
        metrics_export_interval: 60 # Range: 10-300
        log_correlation: true # Boolean
        resource_detection: true # Boolean
        propagation_format: w3c-tracecontext # Options: w3c-tracecontext, b3, jaeger, multi
        composes:
          co_runs_with:
          - api-rest-resource-oriented
      secrets-env-vars: # Inject secrets as env vars; simple but risk of exposure; requires strict process controls.
        injection_method: runtime  # Options: runtime, build-time, container-startup
        secret_rotation_support: false # Boolean
        prefix_convention: APP_SECRET_ # Options: APP_SECRET_, SECRET_, SENSITIVE_, none
        validation_on_startup: true # Boolean
      finops-budget-guardrails: # Implement budgets, alerts, tagging, and policy-as-code guardrails to enforce cost ceilings and prevent runaway spend.
        budget_period: monthly  # Options: monthly, quarterly, annual
        alert_thresholds:
        - 50
        - 80
        - 100
        enforcement_action: alert # Options: alert, prevent, throttle
        tagging_strategy: mandatory # Options: mandatory, recommended, optional
        policy_enforcement: soft # Options: soft, hard, audit
        cost_allocation_level: project # Options: project, team, environment, service
      ops-slo-error-budgets: # Define SLOs and error budgets to balance reliability and velocity.
        slo_target_percentage: 99.9  # Range: 90-99.999
        measurement_window_days: 30 # Options: 7, 28, 30, 90
        error_budget_policy: halt-deployments # Options: halt-deployments, alert-only, slow-rollouts, require-approval
        sli_type: availability # Options: availability, latency, throughput, correctness, composite
        alerting_threshold_percentage: 80 # Range: 50-100
      agentic-provenance-tracking: {} # Architectural pattern for synthesis agents that combine findings from multiple sources into a unified output.
    P2:
      api-versioning-header:  # Version via headers/media types; keeps URLs stable; harder to debug and cache.
        version_header_name: API-Version  # Options: API-Version, X-API-Version, Accept-Version, Custom-Header
        version_format: date-based # Options: semantic, date-based, sequential
        fallback_behavior: latest-stable # Options: latest-stable, oldest-supported, reject-request
        content_negotiation: false # Boolean
        deprecation_policy: warning-header # Options: sunset-header, warning-header, both
      gov-system-manifest: # Pin and govern the inventory of components (agent-tools, agent-skills, agent-models, agent-prompts, services, data sources, external dependencies) the system depends on at a declared manifest path; CI validates on every PR and drift between manifest and built system fails the build.
        manifest_path: docs/architecture/manifest.yaml
        manifest_format: yaml  # Options: yaml, toml, json
        manifest_scope: # Options: agent-tools, agent-skills, agent-models, agent-prompts, data_sources, services, external_dependencies
        - agent-tools
        - agent-skills
        - agent-models
        - agent-prompts
        pin_versions: true # Boolean
        ci_validation: required # Options: required, optional, off
        drift_policy: fail-build # Options: fail-build, warn-only, off
        composes:
          layered_after:
          - iac-terraform
          co_runs_with:
          - release-feature-flags
          - gov-adrs-mandatory
          - ops-runbooks
    P3:
      ops-runbooks:  # Standard runbooks for incidents and routine ops; reduces MTTR and on-call stress.
        runbook_format: markdown  # Options: markdown, wiki, structured_yaml, ticketing_system
        incident_severity_levels: 4 # Options: 3, 4, 5
        escalation_policy: tiered # Options: tiered, follow_the_sun, flat, hybrid
        automation_integration: manual # Options: manual, semi_automated, fully_automated
        review_frequency: quarterly # Options: monthly, quarterly, biannual, post_incident
      gov-adrs-mandatory: # Record architecture decisions and tradeoffs; improves continuity; keep lightweight.
        adr_format: madr  # Options: madr, nygard, y-statements, custom
        storage_location: docs/adrs # Options: docs/adrs, docs/architecture/decisions, adr, wiki
        decision_threshold: significant # Options: all, significant, strategic-only
        review_requirement: peer-review # Options: peer-review, architect-approval, team-consensus, none

# ============================================================
# Cost Feasibility Analysis (Summary)
# ============================================================
#
# Intent: minimize-opex
# Amortization: 24 months
# Total Patterns Selected: 35
#
# COST BREAKDOWN:
# ────────────────────────────────────────────────────────────
# Total CapEx (one-time):     $      80,450
# Pattern OpEx (monthly):     $         610
# Ops Team Cost (monthly):    $           0
# Total OpEx (monthly):       $         610
# Total TCO (24mo):         $      95,090
#
# COST CEILINGS:
# ────────────────────────────────────────────────────────────
# CapEx Ceiling:              $       1,000 ✗ FAIL
# OpEx Ceiling (monthly):     $         500 ✗ FAIL
#
# ⚠️  WARNINGS:
# ────────────────────────────────────────────────────────────
# [high] cost_opex_exceeds_ceiling:
#   Total monthly operational cost ($610) exceeds ceiling ($500) by $110 (intent: minimize-opex)
#
#   Suggestions:
#   - Increase monthly ceiling to $610 or more
#   - Remove high-cost patterns: ops-slo-error-budgets, agentic-topology--hierarchical, obs-agentic-system
#   - Reduce ops_team_size from 0
#
# ============================================================

# ============================================================
# Cost Feasibility Analysis (Details)
# ============================================================
#
# Intent: minimize-opex
# Amortization: 24 months
#
# Ops team size: 0 engineers (no ops cost)
#
# Ops Team Cost Algorithm (for reference):
#   Formula: ops_team_size × single_resource_monthly_ops_usd × on_call_multiplier × deploy_freq_multiplier
#   Based on:
#     - Google SRE Handbook (2016): On-call burden = 25-50% FTE overhead
#     - DORA State of DevOps (2021): Deploy frequency impact on ops overhead
#
# Calculating costs for 35 selected patterns:
#
# PER-PATTERN COSTS:
# ────────────────────────────────────────────────────────────
#
#  1. arch-serverless (match score: 34.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  2. db-managed-postgres (match score: 32.00)
#     Adoption: $1,200.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  3. agentic-output-safety-guardrail (match score: 31.00)
#     Adoption: $1,800.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  4. obs-agentic-system (match score: 31.00)
#     Adoption: $4,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  5. agentic-input-safety-guardrail (match score: 30.00)
#     Adoption: $1,800.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  6. agentic-tool-scoped-credentials (match score: 28.00)
#     Adoption: $3,000.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  7. arch-serverless-pay-per-use (match score: 28.00)
#     Adoption: $1,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  8. agentic-multi-pass-review (match score: 26.00)
#     Adoption: $5,000.0
#     Monthly (min): $50.0
#     Monthly (expected): $50.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $50.0
#
#  9. agentic-topology--hierarchical (match score: 26.00)
#     Adoption: $8,000.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  10. api-rest-resource-oriented (match score: 25.00)
#     Adoption: $750.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  11. agentic-validation-retry-loop (match score: 25.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  12. sync-request-reply-rest (match score: 25.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  13. agentic-session--stateless (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $10.0
#     Monthly (expected): $10.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $10.0
#
#  14. hosting-agentic-workflow-automation--n8n (match score: 25.00)
#     Adoption: $4,000.0
#     Monthly (min): $20.0
#     Monthly (expected): $20.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $20.0
#
#  15. agentic-multi-agent-required (match score: 25.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  16. agentic-model-vendor--anthropic (match score: 25.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  17. agentic-inference--anthropic (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $5.0
#     Monthly (expected): $5.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $5.0
#
#  18. agentic-tool-use-extraction (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $5.0
#     Monthly (expected): $5.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $5.0
#
#  19. agentic-system-required (match score: 24.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  20. deploy-rolling (match score: 24.00)
#     Adoption: $800.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  21. sec-auth-oauth2-oidc (match score: 23.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  22. crud-single-model (match score: 22.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  23. finops-cost-allocation-tags (match score: 21.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  24. release-feature-flags (match score: 19.00)
#     Adoption: $2,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  25. api-versioning-header (match score: 16.00)
#     Adoption: $1,200.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  26. obs-telemetry-backend (match score: 14.00)
#     Adoption: $2,000.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  27. obs-open-telemetry-baseline (match score: 12.00)
#     Adoption: $3,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  28. secrets-env-vars (match score: 11.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  29. finops-budget-guardrails (match score: 10.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  30. ops-runbooks (match score: 8.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  31. ops-slo-error-budgets (match score: 8.00)
#     Adoption: $4,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  32. agentic-provenance-tracking (match score: 8.00)
#     Adoption: $3,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  33. gov-system-manifest (match score: 7.00)
#     Adoption: $4,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  34. iac-terraform (match score: 7.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  35. gov-adrs-mandatory (match score: 7.00)
#     Adoption: $2,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
# Total Monthly OpEx: $610.0
# Monthly operational ceiling: $500 ✗ FAIL
# ============================================================

# ============================================================
# ⚠️  Pattern Advisory Warnings
# (Patterns are still SELECTED — review these before finalizing)
# ============================================================
#
# [warning] warn_constraints:
#   agentic-multi-agent-required: Spec hosts a multi-agent system on a workflow-automation platform. These platforms model linear/branching node graphs, not multi-agent orchestration — they lack structured inter-agent envelopes, parallel agent execution with aggregation, hierarchical orchestration, and shared agent state. Sequential 2-agent flows can be expressed as LLM-node chains; hierarchical/parallel/routing/hybrid topologies typically cannot. Consider hosting-agentic-container (any cloud) or a managed agent platform (hosting-agentic-managed-ai-platform--aws-bedrock-agentcore / vertex-agent-builder / azure-ai-foundry-agents).
#
#   Suggestions:
#   - agentic-multi-agent-required: Spec hosts a multi-agent system on a workflow-automation platform. These platforms model linear/branching node graphs, not multi-agent orchestration — they lack structured inter-agent envelopes, parallel agent execution with aggregation, hierarchical orchestration, and shared agent state. Sequential 2-agent flows can be expressed as LLM-node chains; hierarchical/parallel/routing/hybrid topologies typically cannot. Consider hosting-agentic-container (any cloud) or a managed agent platform (hosting-agentic-managed-ai-platform--aws-bedrock-agentcore / vertex-agent-builder / azure-ai-foundry-agents).
#
# ============================================================
5 Step 5: fix the agentic hosting mismatch — switch hosting from n8n to Dify

Dify is the natural fit: open-source (Apache 2.0), dual-mode SaaS / self-host, native multi-agent via Agent Node + plug-in Agent Strategies. One-line change (highlighted). Expand the inline-spec disclosure below to see the spec at this stage.

spec.yaml (delta added in this step)
constraints:
  features:
    agentic_system:
      hosting: dify
view full spec.yaml at this step (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: 'Multi-agent research system. A coordinator decomposes user queries,

    dispatches web-search and document-analysis workers in parallel,

    and a synthesizer combines findings with provenance. Notify on Slack.

    '
constraints:
  cloud: agnostic
  language: python
  platform: api
  features:
    ai_inference: true
    agentic_system:
      hosting: dify
      topology: hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:
      - agent: research-coordinator
        archetype: orchestrator
        purpose: Decompose research queries, dispatch workers, evaluate synthesis,
          post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor
        purpose: Analyze internal documents for query-relevant content; return findings
          with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
disallowed-patterns:
- agentic-jumpstarter--anthropic
6 Step 6: recompile — diff vs step 4. Only cost warning remains.

The diff: the n8n hosting pattern drops out, the Dify hosting pattern comes in. The workflow-automation warn clears. The cost-ceiling warn is still firing — that's the only "STILL FIRING" entry, addressed in step 7.

diff vs step 3
# ─── pattern SELECTION diff ───
  - hosting-agentic-workflow-automation--n8n (no longer selected) — n8n workflow orchestration platform with AI Agent node and tool nodes.
  + hosting-agentic-managed-ai-platform--dify (newly selected) — Visual agent platform with native multi-agent orchestration via Agent Node + plug-in Agent Strategies; built-in RAG pipeline, 50+ tool integrations, model management, observability.

# ─── warn / cost diff ───
# CLEARED since previous compile:
  [warning] warn_constraints: agentic-multi-agent-required: Spec hosts a multi-agent system on a workflow-automation platform. These platforms model linear/branching node

# STILL FIRING:
  [high] cost_opex_exceeds_ceiling: Total monthly operational cost ($590) exceeds ceiling ($500) by $90 (intent: minimize-opex)
see full compiled-spec.yaml (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: "Multi-agent research system. A coordinator decomposes user queries,\ndispatches web-search and document-analysis workers in parallel,\nand a synthesizer combines findings with provenance. Notify on Slack.\n"
constraints:
  cloud: agnostic  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
  language: python # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
  platform: api # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
  features:
    ai_inference: true  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (11 more)
    agentic_system: # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (12 more)
      hosting: dify  # hosting-agentic-managed-ai-platform--dify
      topology: hierarchical # agentic-topology--hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:  # agentic-topology--hierarchical, agentic-multi-agent-required
      - agent: research-coordinator
        archetype: orchestrator  # agentic-multi-pass-review
        purpose: Decompose research queries, dispatch workers, evaluate synthesis, post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor  # agentic-multi-pass-review, agentic-validation-retry-loop, agentic-tool-use-extraction
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor  # agentic-multi-pass-review, agentic-validation-retry-loop, agentic-tool-use-extraction
        purpose: Analyze internal documents for query-relevant content; return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis  # agentic-provenance-tracking
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
disallowed-patterns:
- agentic-jumpstarter--anthropic
assumptions:
  constraints:
    saas-providers: []
    disallowed-saas-providers: []
    ai-inference-platforms: []
    disallowed-ai-inference-platforms: []
    model-vendors: []
    disallowed-model-vendors: []
    tenantCount: 1
    features:
      caching: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (18 more)
      async_messaging: false # arch-serverless, agentic-output-safety-guardrail, obs-agentic-system, ... (19 more)
      multi_tenancy: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      batch_processing: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (12 more)
      distributed_transactions: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      real_time_streaming: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      vector_search: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      document_store: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (19 more)
      key_value_store: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (19 more)
      graph_database: false # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
      time_series_db: false # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
      oltp_workload: true # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      olap_workload: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      cold_archive_tiering: false
  nfr:
    availability:
      target: 0.95  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (12 more)
    rpo_minutes: 60  # secrets-env-vars, finops-budget-guardrails
    rto_minutes: 60 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (6 more)
    latency:
      p95Milliseconds: 500  # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
      p99Milliseconds: 1000 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
    throughput:
      peak_query_per_second_read: 5  # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (2 more)
      peak_query_per_second_write: 1 # arch-serverless, db-managed-postgres, arch-serverless-pay-per-use, ... (3 more)
    data:
      retention_days: 90
      pii: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      compliance:
        gdpr: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
        gdpr_rtbf: false # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (1 more)
        ccpa: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
        hipaa: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
        sox: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
    consistency:
      needsReadYourWrites: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (15 more)
    durability:
      strict: false  # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
    security:
      auth: oauth2_oidc  # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (15 more)
      tenant_isolation: n/a # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      audit_logging: false # arch-serverless, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
    agentic_system:
      target:
        prompt_injection_block_rate_min: 0.95  # agentic-input-safety-guardrail
        output_safety_pass_rate_min: 0.99 # agentic-output-safety-guardrail, obs-agentic-system
        pii_leak_rate_max: 0.01 # agentic-output-safety-guardrail
        jailbreak_detection_rate_min: 0.9 # agentic-input-safety-guardrail
        hallucination_rate_max: 0.1 # obs-agentic-system, agentic-multi-pass-review
        toxic_output_rate_max: 0.005 # agentic-output-safety-guardrail
  operating_model:
    on_call: false
    deploy_freq: weekly
    ops_team_size: 0
    single_resource_monthly_ops_usd: 10000
    amortization_months: 24
  cost:
    intent:
      priority: minimize-opex
    ceilings:
      monthly_operational_usd: 500
      one_time_setup_usd: 1000
    preferences:
      prefer_free_tier_if_possible: true  # db-managed-postgres, agentic-output-safety-guardrail, obs-agentic-system, ... (12 more)
      prefer_saas_first: false
  patterns:
    meta:
      agentic-multi-agent-required: {}  # Meta-policy pattern that activates when an agentic system has 2 or more agents and enforces topology specification AND surfaces multi-agent-specific risk warnings.
      agentic-system-required: {} # Meta-policy pattern that activates when agentic_system is enabled and enforces that the spec's agentic-system fields are correctly populated and reference valid platforms.
    P0:
      arch-serverless:  # Structure the system as stateless, event-driven function handlers backed by managed services. No persistent servers — each function activates on demand, executes, and terminates. The architectural commitment is to build around events and cloud-managed primitives rather than long-running processes.
        function_runtime_model: faas  # Options: faas, container_functions, hybrid
        cold_start_optimization: minimal # Options: minimal, provisioned_concurrency, keep_warm
        state_management: external # Options: external, durable_functions, step_functions
        event_sources: api_gateway # Options: api_gateway, event_bus, stream_processing, multi_trigger
        observability_level: standard # Options: standard, enhanced, distributed_tracing
      db-managed-postgres: # Use low-ops managed Postgres DBaaS providers (e.g., Supabase and managed cloud Postgres offerings) to reduce DB operations overhead; validate quotas, compliance, and scale limits.
        provider: supabase  # Options: supabase, neon, render, railway, digitalocean-app-platform
        instance_size: small # Options: micro, small, medium, large
        storage_gb: 8 # Range: 1-500
        backup_retention_days: 7 # Range: 1-30
        connection_pooling: true # Boolean
        high_availability: false # Boolean
        ssl_mode: require # Options: disable, allow, prefer, require, verify-ca, verify-full
      arch-serverless-pay-per-use: # Eliminate idle infrastructure cost by running compute on a pay-per-execution billing model rather than paying for always-on instances. Well-suited to bursty or unpredictable workloads where provisioned servers would sit idle most of the time. This is a cost optimisation pattern, not an architectural style — it can apply to any workload that tolerates cold starts.
        compute_type: function  # Options: function, container
        memory_size_mb: 1024 # Options: 128, 256, 512, 1024, 2048, 4096
        timeout_seconds: 30 # Range: 3-900
        cold_start_optimization: standard # Options: standard, provisioned, warm-pool
        concurrency_model: auto # Options: auto, reserved, limited
      agentic-topology--hierarchical: # Central coordinator/orchestrator agent dynamically decomposes tasks, delegates to specialized worker agents (extractors, analyzers, synthesizers), and combines results.
        message_envelope_shape: structured_json  # Options: structured_json, freeform_string, hybrid
        response_envelope_shape: structured_json # Options: structured_json, freeform_string, hybrid
        failure_policy_retry_max: 2 # Min: 0
        failure_policy_retry_backoff: exponential # Options: exponential, linear, immediate, none
        failure_policy_retry_initial_ms: 1000 # Min: 0
        failure_policy_on_terminal_failure: surface_partial # Options: surface_partial, hard_fail, escalate_to_human
        failure_policy_cancel_siblings_on_failure: false # Boolean
        aggregation_strategy: concat_with_attribution # Options: sequential_pipe, concat_with_attribution, merge, last_wins, voting_quorum, judge_arbitration, none
      agentic-session--stateless: {} # Each request is independent; the agent does not persist any conversation context between invocations.
      agentic-model-vendor--anthropic: # Architectural pattern for systems whose models are built by Anthropic (the Claude family: Opus, Sonnet, Haiku).
        enable_token_efficient_tool_use_beta: true  # Boolean
        prompt_caching_recommended_blocks: # Options: system_prompt, tool_definitions, long_context_documents, few_shot_examples
        - system_prompt
        - tool_definitions
        - long_context_documents
      agentic-inference--anthropic: # LLM inference via Anthropic's direct API (api.anthropic.com).
        enable_token_efficient_tool_use_beta: true  # Boolean
      iac-terraform: # Manage infra declaratively with Terraform modules, state, and pipelines.
        state_backend: s3  # Options: s3, terraform-cloud, azurerm, gcs, local
        workspace_strategy: single # Options: single, per-environment, per-service
        module_structure: monorepo # Options: monorepo, separate-repos, registry
        ci_cd_integration: github-actions # Options: github-actions, gitlab-ci, terraform-cloud, atlantis, manual
        drift_detection: false # Boolean
    P1:
      agentic-output-safety-guardrail:  # Post-LLM scan of agent outputs before delivery to users or external systems.
        detection_action: redact-and-continue  # Options: redact-and-continue, block-and-explain, block-silent, human-review, regenerate
      obs-agentic-system: # Production observability tailored for agentic systems. Adds agent-specific tracing (LLM call inputs/outputs, tool dispatch, step graph), cost attribution per archetype/tenant/task, token budget accounting, eval-metric streams, and agent lifecycle metrics. Composes on top of obs-open-telemetry-baseline. Selected unconditionally when agentic_system is declared.
        log_full_prompts: false  # Boolean
      agentic-input-safety-guardrail: # Pre-flight inspection of user inputs and tool-returned content before they reach the agent's LLM.
        fallback_on_guardrail_failure: fail-closed  # Options: fail-open, fail-closed
        block_action: reject-with-explanation # Options: reject-with-explanation, reject-silent, sanitize-and-continue, human-review
      agentic-tool-scoped-credentials: # Each tool call uses a credential narrowly scoped to that tool's operation, not a shared agent-wide credential.
        fail_action_on_credential_error: fail-closed  # Options: fail-closed, fail-open, fallback-to-shared
      agentic-multi-pass-review: # Architecture where one agent session generates output and a separate, independent session reviews it.
        review_session_isolation: fully_independent  # Options: fully_independent, shared_context_acceptable
      api-rest-resource-oriented: # REST API designed around resources (nouns) manipulated via standard HTTP verbs (GET, POST, PUT, DELETE, PATCH). Resources are identified by stable URLs, responses are cacheable by default, and pagination/filtering are expressed as query parameters. Simpler tooling and stronger HTTP cache semantics than GraphQL; well-suited to public APIs and CRUD-heavy domains.
        pagination_style: offset  # Options: offset, cursor, page_number
        max_page_size: 100 # Range: 10-1000
        versioning_strategy: uri # Options: uri, header, query_param, none
        filtering_style: query_params # Options: query_params, json_body, graphql_like
        cache_strategy: etag # Options: etag, last_modified, cache_control, none
        id_format: uuid # Options: uuid, integer, slug, composite
        response_envelope: false # Boolean
      agentic-validation-retry-loop: {} # Layered pattern that wraps any agent output with deterministic validation and retries with specific error feedback.
      sync-request-reply-rest: # Synchronous HTTP APIs; simple integration; needs timeouts/retries/backpressure.
        timeout_seconds: 30  # Range: 1-300
        retry_strategy: exponential_backoff # Options: none, fixed_delay, exponential_backoff, exponential_backoff_jitter
        max_retries: 3 # Range: 0-10
        circuit_breaker_enabled: true # Boolean
        rate_limiting_strategy: token_bucket # Options: none, token_bucket, leaky_bucket, fixed_window, sliding_window
        idempotency_required: false # Boolean
      agentic-tool-use-extraction: # Single-shot structured data extraction using the tool_use mechanism with a forced tool_choice.
        composes:
          layered_after:
          - agentic-output-safety-guardrail
      deploy-rolling: # Gradually replace instances; simplest for stateless services; requires backward-compatible changes.
        batch_size: 25_percent  # Options: 10_percent, 25_percent, 50_percent, one_at_a_time, all_at_once
        health_check_grace_period: 30 # Range: 10-300
        rollback_on_failure: true # Boolean
        deployment_strategy: replace_one_by_one # Options: replace_one_by_one, rolling_batch, canary_first
        max_unavailable: 0_percent # Options: 0_percent, 10_percent, 25_percent
        composes:
          layered_after:
          - iac-terraform
      sec-auth-oauth2-oidc: # Use OAuth2 flows with OIDC identity tokens; standardized claims; delegated auth support.
        oauth_flow: authorization_code  # Options: authorization_code, client_credentials, device_code, implicit
        token_storage: secure_storage # Options: secure_storage, memory_only, encrypted_storage, httponly_cookie
        pkce_enabled: true # Boolean
        scope_strategy: minimal # Options: minimal, role_based, resource_specific
        token_refresh: automatic # Options: automatic, manual, sliding_window
        id_token_validation: strict # Options: strict, standard, relaxed
        composes:
          wraps:
          - api-rest-resource-oriented
      crud-single-model: # Simple CRUD on one canonical model; lowest complexity; best for straightforward domains.
        api_style: rest  # Options: rest, graphql, rpc
        validation_strategy: server-side # Options: server-side, client-side, both
        soft_delete: false # Boolean
        audit_logging: false # Boolean
        pagination_default_size: 20 # Range: 10-100
      finops-cost-allocation-tags: # Tagging/labeling strategy for per-tenant/product cost allocation and chargeback/showback.
        tagging_strategy: hierarchical  # Options: hierarchical, flat, hybrid
        enforcement_level: required # Options: required, recommended, optional
        cost_allocation_model: showback # Options: chargeback, showback, hybrid
        tag_inheritance: true # Boolean
        automated_tagging: true # Boolean
      release-feature-flags: # Decouple deploy from release; safer experiments; needs kill switches and governance.
        flag_storage: config_file  # Options: config_file, database, feature_flag_service, environment_variables
        evaluation_strategy: simple_boolean # Options: simple_boolean, percentage_rollout, user_targeting, multi_variate
        targeting_capability: none # Options: none, user_attributes, context_based, advanced_segments
        kill_switch_enabled: true # Boolean
        audit_logging: false # Boolean
      hosting-agentic-managed-ai-platform--dify: # Visual agent platform with native multi-agent orchestration via Agent Node + plug-in Agent Strategies; built-in RAG pipeline, 50+ tool integrations, model management, observability.
        deployment_mode: cloud_managed  # Options: cloud_managed, self_host_docker, self_host_kubernetes
      obs-telemetry-backend: # Self-hosted observability backend combining Prometheus (metrics storage and alerting), Grafana (dashboards), and Loki (structured log aggregation). Provides the storage and query layer for traces, metrics, and logs; consumed by instrumentation and monitoring strategy patterns. Preferred when data residency, cost control, or customization requirements preclude a managed SaaS observability backend.
        metrics_retention_days: 15  # Options: 7, 15, 30, 90, 365
        log_retention_days: 7 # Options: 1, 3, 7, 14, 30
        backend_ha: false # Boolean
      obs-open-telemetry-baseline: # Standardize traces/metrics/log correlation via OpenTelemetry; export to vendor or OSS backends.
        export_backend: otlp  # Options: otlp, jaeger, zipkin, prometheus, datadog, newrelic, honeycomb
        trace_sampling_strategy: parent-based # Options: always-on, always-off, parent-based, trace-id-ratio
        trace_sampling_rate: 1.0 # Range: 0.0-1.0
        metrics_export_interval: 60 # Range: 10-300
        log_correlation: true # Boolean
        resource_detection: true # Boolean
        propagation_format: w3c-tracecontext # Options: w3c-tracecontext, b3, jaeger, multi
        composes:
          co_runs_with:
          - api-rest-resource-oriented
      secrets-env-vars: # Inject secrets as env vars; simple but risk of exposure; requires strict process controls.
        injection_method: runtime  # Options: runtime, build-time, container-startup
        secret_rotation_support: false # Boolean
        prefix_convention: APP_SECRET_ # Options: APP_SECRET_, SECRET_, SENSITIVE_, none
        validation_on_startup: true # Boolean
      finops-budget-guardrails: # Implement budgets, alerts, tagging, and policy-as-code guardrails to enforce cost ceilings and prevent runaway spend.
        budget_period: monthly  # Options: monthly, quarterly, annual
        alert_thresholds:
        - 50
        - 80
        - 100
        enforcement_action: alert # Options: alert, prevent, throttle
        tagging_strategy: mandatory # Options: mandatory, recommended, optional
        policy_enforcement: soft # Options: soft, hard, audit
        cost_allocation_level: project # Options: project, team, environment, service
      ops-slo-error-budgets: # Define SLOs and error budgets to balance reliability and velocity.
        slo_target_percentage: 99.9  # Range: 90-99.999
        measurement_window_days: 30 # Options: 7, 28, 30, 90
        error_budget_policy: halt-deployments # Options: halt-deployments, alert-only, slow-rollouts, require-approval
        sli_type: availability # Options: availability, latency, throughput, correctness, composite
        alerting_threshold_percentage: 80 # Range: 50-100
      agentic-provenance-tracking: {} # Architectural pattern for synthesis agents that combine findings from multiple sources into a unified output.
    P2:
      api-versioning-header:  # Version via headers/media types; keeps URLs stable; harder to debug and cache.
        version_header_name: API-Version  # Options: API-Version, X-API-Version, Accept-Version, Custom-Header
        version_format: date-based # Options: semantic, date-based, sequential
        fallback_behavior: latest-stable # Options: latest-stable, oldest-supported, reject-request
        content_negotiation: false # Boolean
        deprecation_policy: warning-header # Options: sunset-header, warning-header, both
      gov-system-manifest: # Pin and govern the inventory of components (agent-tools, agent-skills, agent-models, agent-prompts, services, data sources, external dependencies) the system depends on at a declared manifest path; CI validates on every PR and drift between manifest and built system fails the build.
        manifest_path: docs/architecture/manifest.yaml
        manifest_format: yaml  # Options: yaml, toml, json
        manifest_scope: # Options: agent-tools, agent-skills, agent-models, agent-prompts, data_sources, services, external_dependencies
        - agent-tools
        - agent-skills
        - agent-models
        - agent-prompts
        pin_versions: true # Boolean
        ci_validation: required # Options: required, optional, off
        drift_policy: fail-build # Options: fail-build, warn-only, off
        composes:
          layered_after:
          - iac-terraform
          co_runs_with:
          - release-feature-flags
          - gov-adrs-mandatory
          - ops-runbooks
    P3:
      ops-runbooks:  # Standard runbooks for incidents and routine ops; reduces MTTR and on-call stress.
        runbook_format: markdown  # Options: markdown, wiki, structured_yaml, ticketing_system
        incident_severity_levels: 4 # Options: 3, 4, 5
        escalation_policy: tiered # Options: tiered, follow_the_sun, flat, hybrid
        automation_integration: manual # Options: manual, semi_automated, fully_automated
        review_frequency: quarterly # Options: monthly, quarterly, biannual, post_incident
      gov-adrs-mandatory: # Record architecture decisions and tradeoffs; improves continuity; keep lightweight.
        adr_format: madr  # Options: madr, nygard, y-statements, custom
        storage_location: docs/adrs # Options: docs/adrs, docs/architecture/decisions, adr, wiki
        decision_threshold: significant # Options: all, significant, strategic-only
        review_requirement: peer-review # Options: peer-review, architect-approval, team-consensus, none

# ============================================================
# Cost Feasibility Analysis (Summary)
# ============================================================
#
# Intent: minimize-opex
# Amortization: 24 months
# Total Patterns Selected: 35
#
# COST BREAKDOWN:
# ────────────────────────────────────────────────────────────
# Total CapEx (one-time):     $      80,450
# Pattern OpEx (monthly):     $         590
# Ops Team Cost (monthly):    $           0
# Total OpEx (monthly):       $         590
# Total TCO (24mo):         $      94,610
#
# COST CEILINGS:
# ────────────────────────────────────────────────────────────
# CapEx Ceiling:              $       1,000 ✗ FAIL
# OpEx Ceiling (monthly):     $         500 ✗ FAIL
#
# ⚠️  WARNINGS:
# ────────────────────────────────────────────────────────────
# [high] cost_opex_exceeds_ceiling:
#   Total monthly operational cost ($590) exceeds ceiling ($500) by $90 (intent: minimize-opex)
#
#   Suggestions:
#   - Increase monthly ceiling to $590 or more
#   - Remove high-cost patterns: ops-slo-error-budgets, agentic-topology--hierarchical, obs-agentic-system
#   - Reduce ops_team_size from 0
#
# ============================================================

# ============================================================
# Cost Feasibility Analysis (Details)
# ============================================================
#
# Intent: minimize-opex
# Amortization: 24 months
#
# Ops team size: 0 engineers (no ops cost)
#
# Ops Team Cost Algorithm (for reference):
#   Formula: ops_team_size × single_resource_monthly_ops_usd × on_call_multiplier × deploy_freq_multiplier
#   Based on:
#     - Google SRE Handbook (2016): On-call burden = 25-50% FTE overhead
#     - DORA State of DevOps (2021): Deploy frequency impact on ops overhead
#
# Calculating costs for 35 selected patterns:
#
# PER-PATTERN COSTS:
# ────────────────────────────────────────────────────────────
#
#  1. arch-serverless (match score: 34.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  2. db-managed-postgres (match score: 32.00)
#     Adoption: $1,200.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  3. agentic-output-safety-guardrail (match score: 31.00)
#     Adoption: $1,800.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  4. obs-agentic-system (match score: 31.00)
#     Adoption: $4,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  5. agentic-input-safety-guardrail (match score: 30.00)
#     Adoption: $1,800.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  6. agentic-tool-scoped-credentials (match score: 28.00)
#     Adoption: $3,000.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  7. arch-serverless-pay-per-use (match score: 28.00)
#     Adoption: $1,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  8. agentic-multi-pass-review (match score: 26.00)
#     Adoption: $5,000.0
#     Monthly (min): $50.0
#     Monthly (expected): $50.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $50.0
#
#  9. agentic-topology--hierarchical (match score: 26.00)
#     Adoption: $8,000.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  10. api-rest-resource-oriented (match score: 25.00)
#     Adoption: $750.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  11. agentic-validation-retry-loop (match score: 25.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  12. sync-request-reply-rest (match score: 25.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  13. agentic-session--stateless (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $10.0
#     Monthly (expected): $10.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $10.0
#
#  14. agentic-multi-agent-required (match score: 25.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  15. agentic-model-vendor--anthropic (match score: 25.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  16. agentic-inference--anthropic (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $5.0
#     Monthly (expected): $5.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $5.0
#
#  17. agentic-tool-use-extraction (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $5.0
#     Monthly (expected): $5.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $5.0
#
#  18. agentic-system-required (match score: 24.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  19. deploy-rolling (match score: 24.00)
#     Adoption: $800.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  20. sec-auth-oauth2-oidc (match score: 23.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  21. crud-single-model (match score: 22.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  22. finops-cost-allocation-tags (match score: 21.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  23. release-feature-flags (match score: 19.00)
#     Adoption: $2,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  24. api-versioning-header (match score: 16.00)
#     Adoption: $1,200.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  25. hosting-agentic-managed-ai-platform--dify (match score: 15.00)
#     Adoption: $4,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  26. obs-telemetry-backend (match score: 14.00)
#     Adoption: $2,000.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $30.0
#
#  27. obs-open-telemetry-baseline (match score: 12.00)
#     Adoption: $3,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  28. secrets-env-vars (match score: 11.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  29. finops-budget-guardrails (match score: 10.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  30. ops-runbooks (match score: 8.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  31. ops-slo-error-budgets (match score: 8.00)
#     Adoption: $4,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $100.0
#
#  32. agentic-provenance-tracking (match score: 8.00)
#     Adoption: $3,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  33. gov-system-manifest (match score: 7.00)
#     Adoption: $4,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  34. iac-terraform (match score: 7.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
#  35. gov-adrs-mandatory (match score: 7.00)
#     Adoption: $2,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     Monthly OpEx: $0.0
#
# Total Monthly OpEx: $590.0
# Monthly operational ceiling: $500 ✗ FAIL
# ============================================================
7 Step 7: commit to a real cost story

The author replaces the assumed minimize-opex with explicit optimize-tco, declares 1 dedicated ops engineer at $14k/mo loaded, and sets ceilings that fit the team's actual budget ($20k/mo OpEx, $175k one-time CapEx). Every field flipped from assumed → explicit is highlighted. Expand the inline-spec disclosure below to see the spec at this stage.

spec.yaml (delta added in this step)
cost:
  intent:
    priority: optimize-tco
  ceilings:
    monthly_operational_usd: 20000
    one_time_setup_usd: 175000
operating_model:
  ops_team_size: 1
  single_resource_monthly_ops_usd: 14000
  on_call: true
  deploy_freq: daily
  amortization_months: 24
view full spec.yaml at this step (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: 'Multi-agent research system. A coordinator decomposes user queries,

    dispatches web-search and document-analysis workers in parallel,

    and a synthesizer combines findings with provenance. Notify on Slack.

    '
constraints:
  cloud: agnostic
  language: python
  platform: api
  features:
    ai_inference: true
    agentic_system:
      hosting: dify
      topology: hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:
      - agent: research-coordinator
        archetype: orchestrator
        purpose: Decompose research queries, dispatch workers, evaluate synthesis,
          post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor
        purpose: Analyze internal documents for query-relevant content; return findings
          with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
disallowed-patterns:
- agentic-jumpstarter--anthropic
cost:
  intent:
    priority: optimize-tco
  ceilings:
    monthly_operational_usd: 20000
    one_time_setup_usd: 175000
operating_model:
  ops_team_size: 1
  single_resource_monthly_ops_usd: 14000
  on_call: true
  deploy_freq: daily
  amortization_months: 24
8 Step 8: compile — clean. Review per-pattern assumed configs.

All warns cleared. Both cost ceilings PASS. The verbose mode displays each pattern under its priority bucket (meta / P0 / P1 / P2 / P3) with the full per-pattern defaultConfig values AND the alternative options on each line (e.g., scaling_strategy: horizontal # Options: horizontal, vertical, hybrid). The author scans not just the assumed values but the bucket placements — for example, obs-agentic-system sits at P1 by registry default (a generic recommendation), but for production research-automation where prompt injection, tool-misuse, and silent agent drift are first-class risks, the team treats it as load-bearing. Step 9 promotes it from P1 to P0.

Many patterns also carry a composes entry — the compiler's inlined relationship graph. The diagram below visualises the full graph for this compile: which patterns layer after which (build/deploy order), which wrap others (request-time concerns), which co-run (runtime siblings), and which dispatch to others (handoff edges). The implementing-architecture skill walks this graph to sequence scaffolding and runtime wiring — so the same compiler output that approves the architecture also tells the next agent where to start.

compiler output (verbose -v)
# ─── what the compiler FILLED IN as assumptions ───
assumptions:
  constraints:
    tenantCount: 1
    features:
      caching: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (19 more)
      async_messaging: false # arch-monolith, agentic-output-safety-guardrail, obs-agentic-system, ... (20 more)
      multi_tenancy: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      batch_processing: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (13 more)
      distributed_transactions: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      real_time_streaming: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      vector_search: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      document_store: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      key_value_store: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      graph_database: false # arch-monolith, db-managed-postgres, platform-kubernetes, ... (7 more)
      time_series_db: false # arch-monolith, db-managed-postgres, platform-kubernetes, ... (7 more)
      oltp_workload: true # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      olap_workload: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      cold_archive_tiering: false
  nfr:
    rpo_minutes: 60  # arch-monolith, platform-no-mesh, secrets-env-vars, ... (1 more)
    rto_minutes: 60 # arch-monolith, db-managed-postgres, platform-kubernetes, ... (7 more)
    latency:
      p95Milliseconds: 500  # arch-monolith, db-managed-postgres, platform-kubernetes, ... (3 more)
      p99Milliseconds: 1000 # arch-monolith, db-managed-postgres, platform-kubernetes, ... (3 more)
    throughput:
      peak_query_per_second_read: 5  # arch-monolith, db-managed-postgres, platform-kubernetes, ... (3 more)
      peak_query_per_second_write: 1 # arch-monolith, db-managed-postgres, platform-kubernetes, ... (4 more)
    data:
      retention_days: 90
      pii: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      compliance:
    consistency:
      needsReadYourWrites: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (15 more)
    durability:
      strict: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
    security:
      auth: oauth2_oidc  # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (15 more)
      tenant_isolation: n/a # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      audit_logging: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
  cost:

# ─── pattern SELECTION with per-pattern config + alternatives ───
# Each line under a pattern shows the value the compiler ASSUMED.
# The `# Options: …` annotation lists alternatives you can override
# by setting `patterns.<pid>.<field>` in the spec (see next step).
  patterns:
    meta:
      agentic-multi-agent-required: {}  # Meta-policy pattern that activates when an agentic system has 2 or more agents and enforces topology specification AND surfaces multi-agent-specific risk warnings.
      agentic-system-required: {} # Meta-policy pattern that activates when agentic_system is enabled and enforces that the spec's agentic-system fields are correctly populated and reference valid platforms.
    P0:
      arch-monolith:  # Single deployable unit; simplest operational model; scale with replicas behind a load balancer.
        scaling_strategy: horizontal  # Options: horizontal, vertical, hybrid
        load_balancer_type: layer7 # Options: layer4, layer7, dns
        session_management: stateless # Options: stateless, sticky-sessions, shared-state
        deployment_strategy: rolling # Options: rolling, blue-green, canary, recreate
      db-managed-postgres: # Use low-ops managed Postgres DBaaS providers (e.g., Supabase and managed cloud Postgres offerings) to reduce DB operations overhead; validate quotas, compliance, and scale limits.
        provider: supabase  # Options: supabase, neon, render, railway, digitalocean-app-platform
        instance_size: small # Options: micro, small, medium, large
        storage_gb: 8 # Range: 1-500
        backup_retention_days: 7 # Range: 1-30
        connection_pooling: true # Boolean
        high_availability: false # Boolean
        ssl_mode: require # Options: disable, allow, prefer, require, verify-ca, verify-full
      platform-kubernetes: # Standardize on Kubernetes; namespaces/operators; standardized ingress/service patterns.
        namespace_strategy: per-environment  # Options: per-environment, per-team, per-application, monolithic
        ingress_controller: nginx # Options: nginx, traefik, istio, ambassador, contour
        service_mesh: none # Options: none, istio, linkerd, consul
        operator_framework: native # Options: native, operator-sdk, kubebuilder, helm-operator
        autoscaling: hpa # Options: hpa, vpa, keda, hpa-vpa-combined
        network_policy: disabled # Options: disabled, default-deny, namespace-isolation, zero-trust
      agentic-topology--hierarchical: # Central coordinator/orchestrator agent dynamically decomposes tasks, delegates to specialized worker agents (extractors, analyzers, synthesizers), and combines results.
        message_envelope_shape: structured_json  # Options: structured_json, freeform_string, hybrid
        response_envelope_shape: structured_json # Options: structured_json, freeform_string, hybrid
        failure_policy_retry_max: 2 # Min: 0
        failure_policy_retry_backoff: exponential # Options: exponential, linear, immediate, none
        failure_policy_retry_initial_ms: 1000 # Min: 0
        failure_policy_on_terminal_failure: surface_partial # Options: surface_partial, hard_fail, escalate_to_human
        failure_policy_cancel_siblings_on_failure: false # Boolean
        aggregation_strategy: concat_with_attribution # Options: sequential_pipe, concat_with_attribution, merge, last_wins, voting_quorum, judge_arbitration, none
      platform-no-mesh: # Use ingress/API gateway and client libraries; lower complexity; fewer advanced traffic controls.
        gateway_type: api_gateway  # Options: api_gateway, ingress_controller, load_balancer
        client_library_approach: rest_http # Options: rest_http, grpc, graphql
        load_balancing: gateway_level # Options: gateway_level, dns_round_robin, client_side
        observability_strategy: application_level # Options: application_level, gateway_only, hybrid
      agentic-session--stateless: {} # Each request is independent; the agent does not persist any conversation context between invocations.
      agentic-model-vendor--anthropic: # Architectural pattern for systems whose models are built by Anthropic (the Claude family: Opus, Sonnet, Haiku).
        enable_token_efficient_tool_use_beta: true  # Boolean
        prompt_caching_recommended_blocks: # Options: system_prompt, tool_definitions, long_context_documents, few_shot_examples
        - system_prompt
        - tool_definitions
        - long_context_documents
      agentic-inference--anthropic: # LLM inference via Anthropic's direct API (api.anthropic.com).
        enable_token_efficient_tool_use_beta: true  # Boolean
      iac-terraform: # Manage infra declaratively with Terraform modules, state, and pipelines.
        state_backend: s3  # Options: s3, terraform-cloud, azurerm, gcs, local
        workspace_strategy: single # Options: single, per-environment, per-service
        module_structure: monorepo # Options: monorepo, separate-repos, registry
        ci_cd_integration: github-actions # Options: github-actions, gitlab-ci, terraform-cloud, atlantis, manual
        drift_detection: false # Boolean
    P1:
      agentic-output-safety-guardrail:  # Post-LLM scan of agent outputs before delivery to users or external systems.
        detection_action: redact-and-continue  # Options: redact-and-continue, block-and-explain, block-silent, human-review, regenerate
      obs-agentic-system: # Production observability tailored for agentic systems. Adds agent-specific tracing (LLM call inputs/outputs, tool dispatch, step graph), cost attribution per archetype/tenant/task, token budget accounting, eval-metric streams, and agent lifecycle metrics. Composes on top of obs-open-telemetry-baseline. Selected unconditionally when agentic_system is declared.
        log_full_prompts: false  # Boolean
      agentic-input-safety-guardrail: # Pre-flight inspection of user inputs and tool-returned content before they reach the agent's LLM.
        fallback_on_guardrail_failure: fail-closed  # Options: fail-open, fail-closed
        block_action: reject-with-explanation # Options: reject-with-explanation, reject-silent, sanitize-and-continue, human-review
      agentic-tool-scoped-credentials: # Each tool call uses a credential narrowly scoped to that tool's operation, not a shared agent-wide credential.
        fail_action_on_credential_error: fail-closed  # Options: fail-closed, fail-open, fallback-to-shared
      agentic-multi-pass-review: # Architecture where one agent session generates output and a separate, independent session reviews it.
        review_session_isolation: fully_independent  # Options: fully_independent, shared_context_acceptable
      api-rest-resource-oriented: # REST API designed around resources (nouns) manipulated via standard HTTP verbs (GET, POST, PUT, DELETE, PATCH). Resources are identified by stable URLs, responses are cacheable by default, and pagination/filtering are expressed as query parameters. Simpler tooling and stronger HTTP cache semantics than GraphQL; well-suited to public APIs and CRUD-heavy domains.
        pagination_style: offset  # Options: offset, cursor, page_number
        max_page_size: 100 # Range: 10-1000
        versioning_strategy: uri # Options: uri, header, query_param, none
        filtering_style: query_params # Options: query_params, json_body, graphql_like
        cache_strategy: etag # Options: etag, last_modified, cache_control, none
        id_format: uuid # Options: uuid, integer, slug, composite
        response_envelope: false # Boolean
      agentic-validation-retry-loop: {} # Layered pattern that wraps any agent output with deterministic validation and retries with specific error feedback.
      sync-request-reply-rest: # Synchronous HTTP APIs; simple integration; needs timeouts/retries/backpressure.
        timeout_seconds: 30  # Range: 1-300
        retry_strategy: exponential_backoff # Options: none, fixed_delay, exponential_backoff, exponential_backoff_jitter
        max_retries: 3 # Range: 0-10
        circuit_breaker_enabled: true # Boolean
        rate_limiting_strategy: token_bucket # Options: none, token_bucket, leaky_bucket, fixed_window, sliding_window
        idempotency_required: false # Boolean
      agentic-tool-use-extraction: # Single-shot structured data extraction using the tool_use mechanism with a forced tool_choice.
        composes:
          layered_after:
          - agentic-output-safety-guardrail
      deploy-rolling: # Gradually replace instances; simplest for stateless services; requires backward-compatible changes.
        batch_size: 25_percent  # Options: 10_percent, 25_percent, 50_percent, one_at_a_time, all_at_once
        health_check_grace_period: 30 # Range: 10-300
        rollback_on_failure: true # Boolean
        deployment_strategy: replace_one_by_one # Options: replace_one_by_one, rolling_batch, canary_first
        max_unavailable: 0_percent # Options: 0_percent, 10_percent, 25_percent
        composes:
          layered_after:
          - iac-terraform
      sec-auth-oauth2-oidc: # Use OAuth2 flows with OIDC identity tokens; standardized claims; delegated auth support.
        oauth_flow: authorization_code  # Options: authorization_code, client_credentials, device_code, implicit
        token_storage: secure_storage # Options: secure_storage, memory_only, encrypted_storage, httponly_cookie
        pkce_enabled: true # Boolean
        scope_strategy: minimal # Options: minimal, role_based, resource_specific
        token_refresh: automatic # Options: automatic, manual, sliding_window
        id_token_validation: strict # Options: strict, standard, relaxed
        composes:
          wraps:
          - api-rest-resource-oriented
      crud-single-model: # Simple CRUD on one canonical model; lowest complexity; best for straightforward domains.
        api_style: rest  # Options: rest, graphql, rpc
        validation_strategy: server-side # Options: server-side, client-side, both
        soft_delete: false # Boolean
        audit_logging: false # Boolean
        pagination_default_size: 20 # Range: 10-100
      finops-cost-allocation-tags: # Tagging/labeling strategy for per-tenant/product cost allocation and chargeback/showback.
        tagging_strategy: hierarchical  # Options: hierarchical, flat, hybrid
        enforcement_level: required # Options: required, recommended, optional
        cost_allocation_model: showback # Options: chargeback, showback, hybrid
        tag_inheritance: true # Boolean
        automated_tagging: true # Boolean
      release-feature-flags: # Decouple deploy from release; safer experiments; needs kill switches and governance.
        flag_storage: config_file  # Options: config_file, database, feature_flag_service, environment_variables
        evaluation_strategy: simple_boolean # Options: simple_boolean, percentage_rollout, user_targeting, multi_variate
        targeting_capability: none # Options: none, user_attributes, context_based, advanced_segments
        kill_switch_enabled: true # Boolean
        audit_logging: false # Boolean
      hosting-agentic-managed-ai-platform--dify: # Visual agent platform with native multi-agent orchestration via Agent Node + plug-in Agent Strategies; built-in RAG pipeline, 50+ tool integrations, model management, observability.
        deployment_mode: cloud_managed  # Options: cloud_managed, self_host_docker, self_host_kubernetes
      obs-telemetry-backend: # Self-hosted observability backend combining Prometheus (metrics storage and alerting), Grafana (dashboards), and Loki (structured log aggregation). Provides the storage and query layer for traces, metrics, and logs; consumed by instrumentation and monitoring strategy patterns. Preferred when data residency, cost control, or customization requirements preclude a managed SaaS observability backend.
        metrics_retention_days: 15  # Options: 7, 15, 30, 90, 365
        log_retention_days: 7 # Options: 1, 3, 7, 14, 30
        backend_ha: false # Boolean
      obs-open-telemetry-baseline: # Standardize traces/metrics/log correlation via OpenTelemetry; export to vendor or OSS backends.
        export_backend: otlp  # Options: otlp, jaeger, zipkin, prometheus, datadog, newrelic, honeycomb
        trace_sampling_strategy: parent-based # Options: always-on, always-off, parent-based, trace-id-ratio
        trace_sampling_rate: 1.0 # Range: 0.0-1.0
        metrics_export_interval: 60 # Range: 10-300
        log_correlation: true # Boolean
        resource_detection: true # Boolean
        propagation_format: w3c-tracecontext # Options: w3c-tracecontext, b3, jaeger, multi
        composes:
          co_runs_with:
          - api-rest-resource-oriented
      secrets-env-vars: # Inject secrets as env vars; simple but risk of exposure; requires strict process controls.
        injection_method: runtime  # Options: runtime, build-time, container-startup
        secret_rotation_support: false # Boolean
        prefix_convention: APP_SECRET_ # Options: APP_SECRET_, SECRET_, SENSITIVE_, none
        validation_on_startup: true # Boolean
      finops-budget-guardrails: # Implement budgets, alerts, tagging, and policy-as-code guardrails to enforce cost ceilings and prevent runaway spend.
        budget_period: monthly  # Options: monthly, quarterly, annual
        alert_thresholds:
        - 50
        - 80
        - 100
        enforcement_action: alert # Options: alert, prevent, throttle
        tagging_strategy: mandatory # Options: mandatory, recommended, optional
        policy_enforcement: soft # Options: soft, hard, audit
        cost_allocation_level: project # Options: project, team, environment, service
      ops-slo-error-budgets: # Define SLOs and error budgets to balance reliability and velocity.
        slo_target_percentage: 99.9  # Range: 90-99.999
        measurement_window_days: 30 # Options: 7, 28, 30, 90
        error_budget_policy: halt-deployments # Options: halt-deployments, alert-only, slow-rollouts, require-approval
        sli_type: availability # Options: availability, latency, throughput, correctness, composite
        alerting_threshold_percentage: 80 # Range: 50-100
      agentic-provenance-tracking: {} # Architectural pattern for synthesis agents that combine findings from multiple sources into a unified output.
    P2:
      api-versioning-header:  # Version via headers/media types; keeps URLs stable; harder to debug and cache.
        version_header_name: API-Version  # Options: API-Version, X-API-Version, Accept-Version, Custom-Header
        version_format: date-based # Options: semantic, date-based, sequential
        fallback_behavior: latest-stable # Options: latest-stable, oldest-supported, reject-request
        content_negotiation: false # Boolean
        deprecation_policy: warning-header # Options: sunset-header, warning-header, both
      gov-system-manifest: # Pin and govern the inventory of components (agent-tools, agent-skills, agent-models, agent-prompts, services, data sources, external dependencies) the system depends on at a declared manifest path; CI validates on every PR and drift between manifest and built system fails the build.
        manifest_path: docs/architecture/manifest.yaml
        manifest_format: yaml  # Options: yaml, toml, json
        manifest_scope: # Options: agent-tools, agent-skills, agent-models, agent-prompts, data_sources, services, external_dependencies
        - agent-tools
        - agent-skills
        - agent-models
        - agent-prompts
        pin_versions: true # Boolean
        ci_validation: required # Options: required, optional, off
        drift_policy: fail-build # Options: fail-build, warn-only, off
        composes:
          layered_after:
          - iac-terraform
          co_runs_with:
          - release-feature-flags
          - gov-adrs-mandatory
          - ops-runbooks
    P3:
      ops-runbooks:  # Standard runbooks for incidents and routine ops; reduces MTTR and on-call stress.
        runbook_format: markdown  # Options: markdown, wiki, structured_yaml, ticketing_system
        incident_severity_levels: 4 # Options: 3, 4, 5
        escalation_policy: tiered # Options: tiered, follow_the_sun, flat, hybrid
        automation_integration: manual # Options: manual, semi_automated, fully_automated
        review_frequency: quarterly # Options: monthly, quarterly, biannual, post_incident
      gov-adrs-mandatory: # Record architecture decisions and tradeoffs; improves continuity; keep lightweight.
        adr_format: madr  # Options: madr, nygard, y-statements, custom
        storage_location: docs/adrs # Options: docs/adrs, docs/architecture/decisions, adr, wiki
        decision_threshold: significant # Options: all, significant, strategic-only
        review_requirement: peer-review # Options: peer-review, architect-approval, team-consensus, none

# ─── warns and cost feasibility ───
# ============================================================
# Cost Feasibility Analysis (Summary)
# ============================================================
#
# Intent: optimize-tco
# Amortization: 24 months
# Total Patterns Selected: 36
#
# COST BREAKDOWN:
# ────────────────────────────────────────────────────────────
# Total CapEx (one-time):     $      85,450
# Pattern OpEx (monthly):     $       1,040
# Ops Team Cost (monthly):    $      12,600  (1 × $14,000)
# Total OpEx (monthly):       $      13,640
# Total TCO (24mo):         $     412,810
#
# COST CEILINGS:
# ────────────────────────────────────────────────────────────
# CapEx Ceiling:              $     175,000 ✓ PASS
# OpEx Ceiling (monthly):     $      20,000 ✓ PASS
#
# ============================================================
see full compiled-spec.yaml (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: "Multi-agent research system. A coordinator decomposes user queries,\ndispatches web-search and document-analysis workers in parallel,\nand a synthesizer combines findings with provenance. Notify on Slack.\n"
constraints:
  cloud: agnostic  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
  language: python # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
  platform: api # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
  features:
    ai_inference: true  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (12 more)
    agentic_system: # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (12 more)
      hosting: dify  # hosting-agentic-managed-ai-platform--dify
      topology: hierarchical # agentic-topology--hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:  # agentic-topology--hierarchical, agentic-multi-agent-required
      - agent: research-coordinator
        archetype: orchestrator  # agentic-multi-pass-review
        purpose: Decompose research queries, dispatch workers, evaluate synthesis, post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor  # agentic-multi-pass-review, agentic-validation-retry-loop, agentic-tool-use-extraction
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor  # agentic-multi-pass-review, agentic-validation-retry-loop, agentic-tool-use-extraction
        purpose: Analyze internal documents for query-relevant content; return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis  # agentic-provenance-tracking
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless # agentic-session--stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic  # agentic-inference--anthropic
            model_owner: anthropic # agentic-model-vendor--anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
disallowed-patterns:
- agentic-jumpstarter--anthropic
cost:
  intent:
    priority: optimize-tco
  ceilings:
    monthly_operational_usd: 20000
    one_time_setup_usd: 175000
operating_model:
  ops_team_size: 1
  single_resource_monthly_ops_usd: 14000
  on_call: true
  deploy_freq: daily
  amortization_months: 24
assumptions:
  constraints:
    saas-providers: []
    disallowed-saas-providers: []
    ai-inference-platforms: []
    disallowed-ai-inference-platforms: []
    model-vendors: []
    disallowed-model-vendors: []
    tenantCount: 1
    features:
      caching: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (19 more)
      async_messaging: false # arch-monolith, agentic-output-safety-guardrail, obs-agentic-system, ... (20 more)
      multi_tenancy: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      batch_processing: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (13 more)
      distributed_transactions: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      real_time_streaming: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      vector_search: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (22 more)
      document_store: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      key_value_store: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (20 more)
      graph_database: false # arch-monolith, db-managed-postgres, platform-kubernetes, ... (7 more)
      time_series_db: false # arch-monolith, db-managed-postgres, platform-kubernetes, ... (7 more)
      oltp_workload: true # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      olap_workload: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
      cold_archive_tiering: false
  nfr:
    availability:
      target: 0.95  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (13 more)
    rpo_minutes: 60  # arch-monolith, platform-no-mesh, secrets-env-vars, ... (1 more)
    rto_minutes: 60 # arch-monolith, db-managed-postgres, platform-kubernetes, ... (7 more)
    latency:
      p95Milliseconds: 500  # arch-monolith, db-managed-postgres, platform-kubernetes, ... (3 more)
      p99Milliseconds: 1000 # arch-monolith, db-managed-postgres, platform-kubernetes, ... (3 more)
    throughput:
      peak_query_per_second_read: 5  # arch-monolith, db-managed-postgres, platform-kubernetes, ... (3 more)
      peak_query_per_second_write: 1 # arch-monolith, db-managed-postgres, platform-kubernetes, ... (4 more)
    data:
      retention_days: 90
      pii: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      compliance:
        gdpr: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (32 more)
        gdpr_rtbf: false # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (1 more)
        ccpa: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (31 more)
        hipaa: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (33 more)
        sox: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (33 more)
    consistency:
      needsReadYourWrites: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (15 more)
    durability:
      strict: false  # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
    security:
      auth: oauth2_oidc  # agentic-output-safety-guardrail, obs-agentic-system, agentic-input-safety-guardrail, ... (15 more)
      tenant_isolation: n/a # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (14 more)
      audit_logging: false # arch-monolith, db-managed-postgres, agentic-output-safety-guardrail, ... (21 more)
    agentic_system:
      target:
        prompt_injection_block_rate_min: 0.95  # agentic-input-safety-guardrail
        output_safety_pass_rate_min: 0.99 # agentic-output-safety-guardrail, obs-agentic-system
        pii_leak_rate_max: 0.01 # agentic-output-safety-guardrail
        jailbreak_detection_rate_min: 0.9 # agentic-input-safety-guardrail
        hallucination_rate_max: 0.1 # obs-agentic-system, agentic-multi-pass-review
        toxic_output_rate_max: 0.005 # agentic-output-safety-guardrail
  cost:
    preferences:
      prefer_free_tier_if_possible: true  # db-managed-postgres, agentic-output-safety-guardrail, obs-agentic-system, ... (11 more)
      prefer_saas_first: false
  patterns:
    meta:
      agentic-multi-agent-required: {}  # Meta-policy pattern that activates when an agentic system has 2 or more agents and enforces topology specification AND surfaces multi-agent-specific risk warnings.
      agentic-system-required: {} # Meta-policy pattern that activates when agentic_system is enabled and enforces that the spec's agentic-system fields are correctly populated and reference valid platforms.
    P0:
      arch-monolith:  # Single deployable unit; simplest operational model; scale with replicas behind a load balancer.
        scaling_strategy: horizontal  # Options: horizontal, vertical, hybrid
        load_balancer_type: layer7 # Options: layer4, layer7, dns
        session_management: stateless # Options: stateless, sticky-sessions, shared-state
        deployment_strategy: rolling # Options: rolling, blue-green, canary, recreate
      db-managed-postgres: # Use low-ops managed Postgres DBaaS providers (e.g., Supabase and managed cloud Postgres offerings) to reduce DB operations overhead; validate quotas, compliance, and scale limits.
        provider: supabase  # Options: supabase, neon, render, railway, digitalocean-app-platform
        instance_size: small # Options: micro, small, medium, large
        storage_gb: 8 # Range: 1-500
        backup_retention_days: 7 # Range: 1-30
        connection_pooling: true # Boolean
        high_availability: false # Boolean
        ssl_mode: require # Options: disable, allow, prefer, require, verify-ca, verify-full
      platform-kubernetes: # Standardize on Kubernetes; namespaces/operators; standardized ingress/service patterns.
        namespace_strategy: per-environment  # Options: per-environment, per-team, per-application, monolithic
        ingress_controller: nginx # Options: nginx, traefik, istio, ambassador, contour
        service_mesh: none # Options: none, istio, linkerd, consul
        operator_framework: native # Options: native, operator-sdk, kubebuilder, helm-operator
        autoscaling: hpa # Options: hpa, vpa, keda, hpa-vpa-combined
        network_policy: disabled # Options: disabled, default-deny, namespace-isolation, zero-trust
      agentic-topology--hierarchical: # Central coordinator/orchestrator agent dynamically decomposes tasks, delegates to specialized worker agents (extractors, analyzers, synthesizers), and combines results.
        message_envelope_shape: structured_json  # Options: structured_json, freeform_string, hybrid
        response_envelope_shape: structured_json # Options: structured_json, freeform_string, hybrid
        failure_policy_retry_max: 2 # Min: 0
        failure_policy_retry_backoff: exponential # Options: exponential, linear, immediate, none
        failure_policy_retry_initial_ms: 1000 # Min: 0
        failure_policy_on_terminal_failure: surface_partial # Options: surface_partial, hard_fail, escalate_to_human
        failure_policy_cancel_siblings_on_failure: false # Boolean
        aggregation_strategy: concat_with_attribution # Options: sequential_pipe, concat_with_attribution, merge, last_wins, voting_quorum, judge_arbitration, none
      platform-no-mesh: # Use ingress/API gateway and client libraries; lower complexity; fewer advanced traffic controls.
        gateway_type: api_gateway  # Options: api_gateway, ingress_controller, load_balancer
        client_library_approach: rest_http # Options: rest_http, grpc, graphql
        load_balancing: gateway_level # Options: gateway_level, dns_round_robin, client_side
        observability_strategy: application_level # Options: application_level, gateway_only, hybrid
      agentic-session--stateless: {} # Each request is independent; the agent does not persist any conversation context between invocations.
      agentic-model-vendor--anthropic: # Architectural pattern for systems whose models are built by Anthropic (the Claude family: Opus, Sonnet, Haiku).
        enable_token_efficient_tool_use_beta: true  # Boolean
        prompt_caching_recommended_blocks: # Options: system_prompt, tool_definitions, long_context_documents, few_shot_examples
        - system_prompt
        - tool_definitions
        - long_context_documents
      agentic-inference--anthropic: # LLM inference via Anthropic's direct API (api.anthropic.com).
        enable_token_efficient_tool_use_beta: true  # Boolean
      iac-terraform: # Manage infra declaratively with Terraform modules, state, and pipelines.
        state_backend: s3  # Options: s3, terraform-cloud, azurerm, gcs, local
        workspace_strategy: single # Options: single, per-environment, per-service
        module_structure: monorepo # Options: monorepo, separate-repos, registry
        ci_cd_integration: github-actions # Options: github-actions, gitlab-ci, terraform-cloud, atlantis, manual
        drift_detection: false # Boolean
    P1:
      agentic-output-safety-guardrail:  # Post-LLM scan of agent outputs before delivery to users or external systems.
        detection_action: redact-and-continue  # Options: redact-and-continue, block-and-explain, block-silent, human-review, regenerate
      obs-agentic-system: # Production observability tailored for agentic systems. Adds agent-specific tracing (LLM call inputs/outputs, tool dispatch, step graph), cost attribution per archetype/tenant/task, token budget accounting, eval-metric streams, and agent lifecycle metrics. Composes on top of obs-open-telemetry-baseline. Selected unconditionally when agentic_system is declared.
        log_full_prompts: false  # Boolean
      agentic-input-safety-guardrail: # Pre-flight inspection of user inputs and tool-returned content before they reach the agent's LLM.
        fallback_on_guardrail_failure: fail-closed  # Options: fail-open, fail-closed
        block_action: reject-with-explanation # Options: reject-with-explanation, reject-silent, sanitize-and-continue, human-review
      agentic-tool-scoped-credentials: # Each tool call uses a credential narrowly scoped to that tool's operation, not a shared agent-wide credential.
        fail_action_on_credential_error: fail-closed  # Options: fail-closed, fail-open, fallback-to-shared
      agentic-multi-pass-review: # Architecture where one agent session generates output and a separate, independent session reviews it.
        review_session_isolation: fully_independent  # Options: fully_independent, shared_context_acceptable
      api-rest-resource-oriented: # REST API designed around resources (nouns) manipulated via standard HTTP verbs (GET, POST, PUT, DELETE, PATCH). Resources are identified by stable URLs, responses are cacheable by default, and pagination/filtering are expressed as query parameters. Simpler tooling and stronger HTTP cache semantics than GraphQL; well-suited to public APIs and CRUD-heavy domains.
        pagination_style: offset  # Options: offset, cursor, page_number
        max_page_size: 100 # Range: 10-1000
        versioning_strategy: uri # Options: uri, header, query_param, none
        filtering_style: query_params # Options: query_params, json_body, graphql_like
        cache_strategy: etag # Options: etag, last_modified, cache_control, none
        id_format: uuid # Options: uuid, integer, slug, composite
        response_envelope: false # Boolean
      agentic-validation-retry-loop: {} # Layered pattern that wraps any agent output with deterministic validation and retries with specific error feedback.
      sync-request-reply-rest: # Synchronous HTTP APIs; simple integration; needs timeouts/retries/backpressure.
        timeout_seconds: 30  # Range: 1-300
        retry_strategy: exponential_backoff # Options: none, fixed_delay, exponential_backoff, exponential_backoff_jitter
        max_retries: 3 # Range: 0-10
        circuit_breaker_enabled: true # Boolean
        rate_limiting_strategy: token_bucket # Options: none, token_bucket, leaky_bucket, fixed_window, sliding_window
        idempotency_required: false # Boolean
      agentic-tool-use-extraction: # Single-shot structured data extraction using the tool_use mechanism with a forced tool_choice.
        composes:
          layered_after:
          - agentic-output-safety-guardrail
      deploy-rolling: # Gradually replace instances; simplest for stateless services; requires backward-compatible changes.
        batch_size: 25_percent  # Options: 10_percent, 25_percent, 50_percent, one_at_a_time, all_at_once
        health_check_grace_period: 30 # Range: 10-300
        rollback_on_failure: true # Boolean
        deployment_strategy: replace_one_by_one # Options: replace_one_by_one, rolling_batch, canary_first
        max_unavailable: 0_percent # Options: 0_percent, 10_percent, 25_percent
        composes:
          layered_after:
          - iac-terraform
      sec-auth-oauth2-oidc: # Use OAuth2 flows with OIDC identity tokens; standardized claims; delegated auth support.
        oauth_flow: authorization_code  # Options: authorization_code, client_credentials, device_code, implicit
        token_storage: secure_storage # Options: secure_storage, memory_only, encrypted_storage, httponly_cookie
        pkce_enabled: true # Boolean
        scope_strategy: minimal # Options: minimal, role_based, resource_specific
        token_refresh: automatic # Options: automatic, manual, sliding_window
        id_token_validation: strict # Options: strict, standard, relaxed
        composes:
          wraps:
          - api-rest-resource-oriented
      crud-single-model: # Simple CRUD on one canonical model; lowest complexity; best for straightforward domains.
        api_style: rest  # Options: rest, graphql, rpc
        validation_strategy: server-side # Options: server-side, client-side, both
        soft_delete: false # Boolean
        audit_logging: false # Boolean
        pagination_default_size: 20 # Range: 10-100
      finops-cost-allocation-tags: # Tagging/labeling strategy for per-tenant/product cost allocation and chargeback/showback.
        tagging_strategy: hierarchical  # Options: hierarchical, flat, hybrid
        enforcement_level: required # Options: required, recommended, optional
        cost_allocation_model: showback # Options: chargeback, showback, hybrid
        tag_inheritance: true # Boolean
        automated_tagging: true # Boolean
      release-feature-flags: # Decouple deploy from release; safer experiments; needs kill switches and governance.
        flag_storage: config_file  # Options: config_file, database, feature_flag_service, environment_variables
        evaluation_strategy: simple_boolean # Options: simple_boolean, percentage_rollout, user_targeting, multi_variate
        targeting_capability: none # Options: none, user_attributes, context_based, advanced_segments
        kill_switch_enabled: true # Boolean
        audit_logging: false # Boolean
      hosting-agentic-managed-ai-platform--dify: # Visual agent platform with native multi-agent orchestration via Agent Node + plug-in Agent Strategies; built-in RAG pipeline, 50+ tool integrations, model management, observability.
        deployment_mode: cloud_managed  # Options: cloud_managed, self_host_docker, self_host_kubernetes
      obs-telemetry-backend: # Self-hosted observability backend combining Prometheus (metrics storage and alerting), Grafana (dashboards), and Loki (structured log aggregation). Provides the storage and query layer for traces, metrics, and logs; consumed by instrumentation and monitoring strategy patterns. Preferred when data residency, cost control, or customization requirements preclude a managed SaaS observability backend.
        metrics_retention_days: 15  # Options: 7, 15, 30, 90, 365
        log_retention_days: 7 # Options: 1, 3, 7, 14, 30
        backend_ha: false # Boolean
      obs-open-telemetry-baseline: # Standardize traces/metrics/log correlation via OpenTelemetry; export to vendor or OSS backends.
        export_backend: otlp  # Options: otlp, jaeger, zipkin, prometheus, datadog, newrelic, honeycomb
        trace_sampling_strategy: parent-based # Options: always-on, always-off, parent-based, trace-id-ratio
        trace_sampling_rate: 1.0 # Range: 0.0-1.0
        metrics_export_interval: 60 # Range: 10-300
        log_correlation: true # Boolean
        resource_detection: true # Boolean
        propagation_format: w3c-tracecontext # Options: w3c-tracecontext, b3, jaeger, multi
        composes:
          co_runs_with:
          - api-rest-resource-oriented
      secrets-env-vars: # Inject secrets as env vars; simple but risk of exposure; requires strict process controls.
        injection_method: runtime  # Options: runtime, build-time, container-startup
        secret_rotation_support: false # Boolean
        prefix_convention: APP_SECRET_ # Options: APP_SECRET_, SECRET_, SENSITIVE_, none
        validation_on_startup: true # Boolean
      finops-budget-guardrails: # Implement budgets, alerts, tagging, and policy-as-code guardrails to enforce cost ceilings and prevent runaway spend.
        budget_period: monthly  # Options: monthly, quarterly, annual
        alert_thresholds:
        - 50
        - 80
        - 100
        enforcement_action: alert # Options: alert, prevent, throttle
        tagging_strategy: mandatory # Options: mandatory, recommended, optional
        policy_enforcement: soft # Options: soft, hard, audit
        cost_allocation_level: project # Options: project, team, environment, service
      ops-slo-error-budgets: # Define SLOs and error budgets to balance reliability and velocity.
        slo_target_percentage: 99.9  # Range: 90-99.999
        measurement_window_days: 30 # Options: 7, 28, 30, 90
        error_budget_policy: halt-deployments # Options: halt-deployments, alert-only, slow-rollouts, require-approval
        sli_type: availability # Options: availability, latency, throughput, correctness, composite
        alerting_threshold_percentage: 80 # Range: 50-100
      agentic-provenance-tracking: {} # Architectural pattern for synthesis agents that combine findings from multiple sources into a unified output.
    P2:
      api-versioning-header:  # Version via headers/media types; keeps URLs stable; harder to debug and cache.
        version_header_name: API-Version  # Options: API-Version, X-API-Version, Accept-Version, Custom-Header
        version_format: date-based # Options: semantic, date-based, sequential
        fallback_behavior: latest-stable # Options: latest-stable, oldest-supported, reject-request
        content_negotiation: false # Boolean
        deprecation_policy: warning-header # Options: sunset-header, warning-header, both
      gov-system-manifest: # Pin and govern the inventory of components (agent-tools, agent-skills, agent-models, agent-prompts, services, data sources, external dependencies) the system depends on at a declared manifest path; CI validates on every PR and drift between manifest and built system fails the build.
        manifest_path: docs/architecture/manifest.yaml
        manifest_format: yaml  # Options: yaml, toml, json
        manifest_scope: # Options: agent-tools, agent-skills, agent-models, agent-prompts, data_sources, services, external_dependencies
        - agent-tools
        - agent-skills
        - agent-models
        - agent-prompts
        pin_versions: true # Boolean
        ci_validation: required # Options: required, optional, off
        drift_policy: fail-build # Options: fail-build, warn-only, off
        composes:
          layered_after:
          - iac-terraform
          co_runs_with:
          - release-feature-flags
          - gov-adrs-mandatory
          - ops-runbooks
    P3:
      ops-runbooks:  # Standard runbooks for incidents and routine ops; reduces MTTR and on-call stress.
        runbook_format: markdown  # Options: markdown, wiki, structured_yaml, ticketing_system
        incident_severity_levels: 4 # Options: 3, 4, 5
        escalation_policy: tiered # Options: tiered, follow_the_sun, flat, hybrid
        automation_integration: manual # Options: manual, semi_automated, fully_automated
        review_frequency: quarterly # Options: monthly, quarterly, biannual, post_incident
      gov-adrs-mandatory: # Record architecture decisions and tradeoffs; improves continuity; keep lightweight.
        adr_format: madr  # Options: madr, nygard, y-statements, custom
        storage_location: docs/adrs # Options: docs/adrs, docs/architecture/decisions, adr, wiki
        decision_threshold: significant # Options: all, significant, strategic-only
        review_requirement: peer-review # Options: peer-review, architect-approval, team-consensus, none

# ============================================================
# Cost Feasibility Analysis (Summary)
# ============================================================
#
# Intent: optimize-tco
# Amortization: 24 months
# Total Patterns Selected: 36
#
# COST BREAKDOWN:
# ────────────────────────────────────────────────────────────
# Total CapEx (one-time):     $      85,450
# Pattern OpEx (monthly):     $       1,040
# Ops Team Cost (monthly):    $      12,600  (1 × $14,000)
# Total OpEx (monthly):       $      13,640
# Total TCO (24mo):         $     412,810
#
# COST CEILINGS:
# ────────────────────────────────────────────────────────────
# CapEx Ceiling:              $     175,000 ✓ PASS
# OpEx Ceiling (monthly):     $      20,000 ✓ PASS
#
# ============================================================

# ============================================================
# Cost Feasibility Analysis (Details)
# ============================================================
#
# Intent: optimize-tco
# Amortization: 24 months
#
# Ops Team Cost Breakdown:
#   Base: 1 engineers × $14,000/month = $14,000
#   On-call multiplier: 1.5x (on-call burden)
#   Deploy frequency multiplier: 0.6x (deploy_freq: daily, high automation)
#   Adjusted ops cost: $14,000 × 1.5 × 0.6 = $12,600/month
#
#   Deploy Frequency Options (DORA State of DevOps):
#     on-demand: 0.5x  (very high automation)
#     daily:     0.6x  (high automation)
#     weekly:    0.8x  (moderate automation)
#     biweekly:  0.9x  (manual processes)
#     monthly:   1.0x  (very manual)
#     quarterly: 1.1x  (extremely manual)
#
#
# Ops Team Cost Algorithm (for reference):
#   Formula: ops_team_size × single_resource_monthly_ops_usd × on_call_multiplier × deploy_freq_multiplier
#   Based on:
#     - Google SRE Handbook (2016): On-call burden = 25-50% FTE overhead
#     - DORA State of DevOps (2021): Deploy frequency impact on ops overhead
#
# Calculating costs for 36 selected patterns:
#
# PER-PATTERN COSTS:
# ────────────────────────────────────────────────────────────
#
#  1. arch-monolith (match score: 34.00)
#     Adoption: $500.0
#     Monthly (min): $50.0
#     Monthly (expected): $50.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $500.0 + ($50.0 × 24) = $1,700.0
#
#  2. db-managed-postgres (match score: 32.00)
#     Adoption: $1,200.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,200.0 + ($0.0 × 24) = $1,200.0
#
#  3. agentic-output-safety-guardrail (match score: 31.00)
#     Adoption: $1,800.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,800.0 + ($30.0 × 24) = $2,520.0
#
#  4. obs-agentic-system (match score: 31.00)
#     Adoption: $4,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $4,500.0 + ($100.0 × 24) = $6,900.0
#
#  5. agentic-input-safety-guardrail (match score: 30.00)
#     Adoption: $1,800.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,800.0 + ($30.0 × 24) = $2,520.0
#
#  6. platform-kubernetes (match score: 30.00)
#     Adoption: $8,000.0
#     Monthly (min): $200.0
#     Monthly (expected): $200.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $8,000.0 + ($200.0 × 24) = $12,800.0
#
#  7. agentic-tool-scoped-credentials (match score: 28.00)
#     Adoption: $3,000.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $3,000.0 + ($30.0 × 24) = $3,720.0
#
#  8. agentic-multi-pass-review (match score: 26.00)
#     Adoption: $5,000.0
#     Monthly (min): $50.0
#     Monthly (expected): $50.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $5,000.0 + ($50.0 × 24) = $6,200.0
#
#  9. agentic-topology--hierarchical (match score: 26.00)
#     Adoption: $8,000.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $8,000.0 + ($100.0 × 24) = $10,400.0
#
#  10. platform-no-mesh (match score: 26.00)
#     Adoption: $1,500.0
#     Monthly (min): $200.0
#     Monthly (expected): $200.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,500.0 + ($200.0 × 24) = $6,300.0
#
#  11. api-rest-resource-oriented (match score: 25.00)
#     Adoption: $750.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $750.0 + ($0.0 × 24) = $750.0
#
#  12. agentic-validation-retry-loop (match score: 25.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $2,500.0 + ($0.0 × 24) = $2,500.0
#
#  13. sync-request-reply-rest (match score: 25.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $300.0 + ($0.0 × 24) = $300.0
#
#  14. agentic-session--stateless (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $10.0
#     Monthly (expected): $10.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,500.0 + ($10.0 × 24) = $1,740.0
#
#  15. agentic-multi-agent-required (match score: 25.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $0.0 + ($0.0 × 24) = $0.0
#
#  16. agentic-model-vendor--anthropic (match score: 25.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $0.0 + ($0.0 × 24) = $0.0
#
#  17. agentic-inference--anthropic (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $5.0
#     Monthly (expected): $5.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,500.0 + ($5.0 × 24) = $1,620.0
#
#  18. agentic-tool-use-extraction (match score: 25.00)
#     Adoption: $1,500.0
#     Monthly (min): $5.0
#     Monthly (expected): $5.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,500.0 + ($5.0 × 24) = $1,620.0
#
#  19. agentic-system-required (match score: 24.00)
#     Adoption: $0.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $0.0 + ($0.0 × 24) = $0.0
#
#  20. deploy-rolling (match score: 24.00)
#     Adoption: $800.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $800.0 + ($0.0 × 24) = $800.0
#
#  21. sec-auth-oauth2-oidc (match score: 23.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $3,500.0 + ($0.0 × 24) = $3,500.0
#
#  22. crud-single-model (match score: 22.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $300.0 + ($0.0 × 24) = $300.0
#
#  23. finops-cost-allocation-tags (match score: 21.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $2,500.0 + ($0.0 × 24) = $2,500.0
#
#  24. release-feature-flags (match score: 19.00)
#     Adoption: $2,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $2,000.0 + ($0.0 × 24) = $2,000.0
#
#  25. api-versioning-header (match score: 16.00)
#     Adoption: $1,200.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $1,200.0 + ($0.0 × 24) = $1,200.0
#
#  26. hosting-agentic-managed-ai-platform--dify (match score: 15.00)
#     Adoption: $4,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $4,000.0 + ($0.0 × 24) = $4,000.0
#
#  27. obs-telemetry-backend (match score: 14.00)
#     Adoption: $2,000.0
#     Monthly (min): $30.0
#     Monthly (expected): $30.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $2,000.0 + ($30.0 × 24) = $2,720.0
#
#  28. obs-open-telemetry-baseline (match score: 12.00)
#     Adoption: $3,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $3,500.0 + ($100.0 × 24) = $5,900.0
#
#  29. secrets-env-vars (match score: 11.00)
#     Adoption: $300.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $300.0 + ($0.0 × 24) = $300.0
#
#  30. finops-budget-guardrails (match score: 10.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $2,500.0 + ($0.0 × 24) = $2,500.0
#
#  31. ops-runbooks (match score: 8.00)
#     Adoption: $2,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $2,500.0 + ($0.0 × 24) = $2,500.0
#
#  32. ops-slo-error-budgets (match score: 8.00)
#     Adoption: $4,500.0
#     Monthly (min): $100.0
#     Monthly (expected): $100.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $4,500.0 + ($100.0 × 24) = $6,900.0
#
#  33. agentic-provenance-tracking (match score: 8.00)
#     Adoption: $3,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $3,000.0 + ($0.0 × 24) = $3,000.0
#
#  34. gov-system-manifest (match score: 7.00)
#     Adoption: $4,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $4,000.0 + ($0.0 × 24) = $4,000.0
#
#  35. iac-terraform (match score: 7.00)
#     Adoption: $3,500.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $3,500.0 + ($0.0 × 24) = $3,500.0
#
#  36. gov-adrs-mandatory (match score: 7.00)
#     Adoption: $2,000.0
#     Monthly (min): $0.0
#     Monthly (expected): $0.0
#     Ops cost: $0 (no infrastructure)
#     ──────────────────────────────────────
#     TCO (24mo): $2,000.0 + ($0.0 × 24) = $2,000.0
#
# Total TCO (24mo): $412,810.0
# Monthly operational ceiling: $20,000 ✓ PASS
# One-time setup ceiling: $175,000 ✓ PASS
# ============================================================
composes graph — pattern relationships (8 kept, 21 pruned)
graph LR
  n_gov_system_manifest["gov-system-manifest"] -->|co-runs with| n_gov_adrs_mandatory["gov-adrs-mandatory"]
  n_gov_system_manifest["gov-system-manifest"] -->|co-runs with| n_ops_runbooks["ops-runbooks"]
  n_gov_system_manifest["gov-system-manifest"] -->|co-runs with| n_release_feature_flags["release-feature-flags"]
  n_obs_open_telemetry_baseline["obs-open-telemetry-baseline"] -->|co-runs with| n_api_rest_resource_oriented["api-rest-resource-oriented"]
  n_agentic_tool_use_extraction["agentic-tool-use-extraction"] -->|layered after| n_agentic_output_safety_guardrail["agentic-output-safety-guardrail"]
  n_deploy_rolling["deploy-rolling"] -->|layered after| n_iac_terraform["iac-terraform"]
  n_gov_system_manifest["gov-system-manifest"] -->|layered after| n_iac_terraform["iac-terraform"]
  n_sec_auth_oauth2_oidc["sec-auth-oauth2-oidc"] -->|wraps| n_api_rest_resource_oriented["api-rest-resource-oriented"]
  n_agentic_provenance_tracking["agentic-provenance-tracking"] -.->|co-runs with| n_agentic_code_orchestration["agentic-code-orchestration"]
  n_agentic_provenance_tracking["agentic-provenance-tracking"] -.->|co-runs with| n_agentic_mcp_server["agentic-mcp-server"]
  n_agentic_provenance_tracking["agentic-provenance-tracking"] -.->|co-runs with| n_agentic_multi_tool_workflow["agentic-multi-tool-workflow"]
  n_obs_agentic_system["obs-agentic-system"] -.->|co-runs with| n_agentic_code_orchestration["agentic-code-orchestration"]
  n_obs_agentic_system["obs-agentic-system"] -.->|co-runs with| n_agentic_mcp_server["agentic-mcp-server"]
  n_obs_agentic_system["obs-agentic-system"] -.->|co-runs with| n_agentic_multi_tool_workflow["agentic-multi-tool-workflow"]
  n_obs_open_telemetry_baseline["obs-open-telemetry-baseline"] -.->|co-runs with| n_api_graphql_schema_first["api-graphql-schema-first"]
  n_deploy_rolling["deploy-rolling"] -.->|layered after| n_iac_bicep["iac-bicep"]
  n_deploy_rolling["deploy-rolling"] -.->|layered after| n_iac_cloudformation["iac-cloudformation"]
  n_gov_system_manifest["gov-system-manifest"] -.->|layered after| n_iac_cloudformation["iac-cloudformation"]
  n_agentic_input_safety_guardrail["agentic-input-safety-guardrail"] -.->|wraps| n_agentic_code_orchestration["agentic-code-orchestration"]
  n_agentic_input_safety_guardrail["agentic-input-safety-guardrail"] -.->|wraps| n_agentic_mcp_server["agentic-mcp-server"]
  n_agentic_input_safety_guardrail["agentic-input-safety-guardrail"] -.->|wraps| n_agentic_multi_tool_workflow["agentic-multi-tool-workflow"]
  n_agentic_output_safety_guardrail["agentic-output-safety-guardrail"] -.->|wraps| n_agentic_code_orchestration["agentic-code-orchestration"]
  n_agentic_output_safety_guardrail["agentic-output-safety-guardrail"] -.->|wraps| n_agentic_mcp_server["agentic-mcp-server"]
  n_agentic_output_safety_guardrail["agentic-output-safety-guardrail"] -.->|wraps| n_agentic_multi_tool_workflow["agentic-multi-tool-workflow"]
  n_agentic_tool_scoped_credentials["agentic-tool-scoped-credentials"] -.->|wraps| n_agentic_mcp_server["agentic-mcp-server"]
  n_agentic_tool_scoped_credentials["agentic-tool-scoped-credentials"] -.->|wraps| n_agentic_multi_tool_workflow["agentic-multi-tool-workflow"]
  n_agentic_validation_retry_loop["agentic-validation-retry-loop"] -.->|wraps| n_agentic_mcp_server["agentic-mcp-server"]
  n_agentic_validation_retry_loop["agentic-validation-retry-loop"] -.->|wraps| n_agentic_multi_tool_workflow["agentic-multi-tool-workflow"]
  n_sec_auth_oauth2_oidc["sec-auth-oauth2-oidc"] -.->|wraps| n_api_graphql_schema_first["api-graphql-schema-first"]
  classDef pruned stroke-dasharray:4,color:#aaa,fill:#222,stroke:#888
  class n_agentic_code_orchestration pruned
  class n_agentic_mcp_server pruned
  class n_agentic_multi_tool_workflow pruned
  class n_api_graphql_schema_first pruned
  class n_iac_bicep pruned
  class n_iac_cloudformation pruned
Solid arrows are kept edges (target pattern is in the selected set — the implementing agent will wire these). Dashed arrows + dimmed nodes are pruned edges (target pattern was not selected for this spec, so the compiler dropped the edge from the inlined graph — the pattern still ships, but without that relationship). Edge labels: layered after (build/deploy order), wraps (request-time concern), co-runs with (runtime siblings), dispatches to (handoff).
Clean compile. Every pattern's defaultConfig and bucket placement is visible — the spec author can scan and decide what to override in step 9.
9 Step 9: promote obs-agentic-system to P0 and flip log_full_prompts to true

Reviewing step 8 the team takes two decisions on obs-agentic-system:

  • Priority bump. Registry default is P1 — a reasonable generic recommendation. For this workload (production research-automation, multi-agent, customer-facing summaries) agentic observability is load-bearing; prompt injection, tool misuse, silent agent drift, and cost-runaway are all first-class risks. The team can't ship without per-agent tracing, token accounting, and eval streams in P0.
  • Config flip. The pattern's log_full_prompts field defaults to false (a PII-conservative default the registry enforces). The team has reviewed their data and confirmed PII is not in scope for this workload — so they turn it true to persist full prompts and tool-call inputs for replay and debugging.
Both overrides live side-by-side under the same patterns.obs-agentic-system entry — explicit choices win over registry defaults. Expand the inline spec below to see the spec at this stage.

per-pattern priority + config overrides — promotion from assumptions to explicit
# BEFORE — assumptions.patterns.P1.obs-agentic-system (registry default bucket = P1)
assumptions:
  patterns:
    P1:
      obs-agentic-system:    # registry defaultConfig applied implicitly
        log_full_prompts: False

# AFTER — patterns.obs-agentic-system (top-level priority + config overrides; bumps P1 → P0)
patterns:
  P0:    # was P1 (registry default)
    obs-agentic-system:
      log_full_prompts: True    # was False (registry defaultConfig)

# Promotion: on recompile, obs-agentic-system lands in the P0 bucket of
# assumptions.patterns.P0 (not P1). The approved
# architecture's bucket-grouped patterns block reflects the new placement.
view full spec.yaml at this step (expand inline)
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: 'Multi-agent research system. A coordinator decomposes user queries,

    dispatches web-search and document-analysis workers in parallel,

    and a synthesizer combines findings with provenance. Notify on Slack.

    '
constraints:
  cloud: agnostic
  language: python
  platform: api
  features:
    ai_inference: true
    agentic_system:
      hosting: dify
      topology: hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:
      - agent: research-coordinator
        archetype: orchestrator
        purpose: Decompose research queries, dispatch workers, evaluate synthesis,
          post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor
        purpose: Analyze internal documents for query-relevant content; return findings
          with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
disallowed-patterns:
- agentic-jumpstarter--anthropic
cost:
  intent:
    priority: optimize-tco
  ceilings:
    monthly_operational_usd: 20000
    one_time_setup_usd: 175000
operating_model:
  ops_team_size: 1
  single_resource_monthly_ops_usd: 14000
  on_call: true
  deploy_freq: daily
  amortization_months: 24
patterns:
  P0:
    obs-agentic-system:
      log_full_prompts: true
10 Step 10: promote ALL assumptions to explicit + prepend approval header

Final step: the compiler is run on the step-9 spec, then EVERY key under assumptions.* is lifted into the top-level spec body so the result has no assumptions section. This is the contract from skills/compiling-architecture/SKILL.md ("the approved architecture must have no assumptions block. Every field must be explicit"). The # STATUS: APPROVED comment block is prepended at the top — matching the format the skill documents. The footer below the panel shows the verification: re-compiling the promoted architecture must exit 0 (idempotent). The implementing-architecture skill verifies handoff readiness via grep -c "STATUS: APPROVED". Approver name and date are anonymised placeholders here; in practice the author fills them in at commit time.

architecture.yaml (approved + no assumptions block)
# STATUS: APPROVED
# Approved by: <architect-on-record>
# Approved at: <YYYY-MM-DD>
#
# This header is consumed by skills/implementing-architecture/SKILL.md
# to verify the architecture is handoff-ready. Recompilation of the
# underlying spec invalidates this approval — fresh review required.
project:
  name: research-agentic
  domain: knowledge-management
functional:
  summary: 'Multi-agent research system. A coordinator decomposes user queries,

    dispatches web-search and document-analysis workers in parallel,

    and a synthesizer combines findings with provenance. Notify on Slack.

    '
constraints:
  cloud: agnostic
  language: python
  platform: api
  features:
    ai_inference: true
    agentic_system:
      hosting: dify
      topology: hierarchical
      integrations:
        slack: true
        teams: false
        discord: false
        jira: false
        github: false
        gitlab: false
        linear: false
        email: false
        calendar: false
        notion: false
        confluence: false
        pagerduty: false
        webhook: false
      agents:
      - agent: research-coordinator
        archetype: orchestrator
        purpose: Decompose research queries, dispatch workers, evaluate synthesis,
          post final report.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
      - agent: web-searcher
        archetype: extractor
        purpose: Search the web for query subtopics, return findings with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: document-analyzer
        archetype: extractor
        purpose: Analyze internal documents for query-relevant content; return findings
          with source attribution.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-haiku
            model_id: claude-haiku-4-5
      - agent: synthesizer
        archetype: synthesis
        purpose: Combine findings into a coherent report preserving claim-level provenance.
        escalation:
          policy: never
        mutating_actions: false
        dynamic_ui: false
        streaming: false
        processing_mode: realtime
        session_strategy: stateless
        models:
        - role: primary
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-opus
            model_id: claude-opus-4-7
        - role: fallback
          candidates:
          - host_platform: anthropic
            model_owner: anthropic
            model_name: claude-sonnet
            model_id: claude-sonnet-4-6
    caching: false
    async_messaging: false
    multi_tenancy: false
    batch_processing: false
    distributed_transactions: false
    real_time_streaming: false
    vector_search: false
    document_store: false
    key_value_store: false
    graph_database: false
    time_series_db: false
    oltp_workload: true
    olap_workload: false
    cold_archive_tiering: false
  saas-providers: []
  disallowed-saas-providers: []
  ai-inference-platforms: []
  disallowed-ai-inference-platforms: []
  model-vendors: []
  disallowed-model-vendors: []
  tenantCount: 1
disallowed-patterns:
- agentic-jumpstarter--anthropic
cost:
  intent:
    priority: optimize-tco
  ceilings:
    monthly_operational_usd: 20000
    one_time_setup_usd: 175000
  preferences:
    prefer_free_tier_if_possible: true
    prefer_saas_first: false
operating_model:
  ops_team_size: 1
  single_resource_monthly_ops_usd: 14000
  on_call: true
  deploy_freq: daily
  amortization_months: 24
nfr:
  availability:
    target: 0.95
  rpo_minutes: 60
  rto_minutes: 60
  latency:
    p95Milliseconds: 500
    p99Milliseconds: 1000
  throughput:
    peak_query_per_second_read: 5
    peak_query_per_second_write: 1
  data:
    retention_days: 90
    pii: false
    compliance:
      gdpr: false
      gdpr_rtbf: false
      ccpa: false
      hipaa: false
      sox: false
  consistency:
    needsReadYourWrites: false
  durability:
    strict: false
  security:
    auth: oauth2_oidc
    tenant_isolation: n/a
    audit_logging: false
  agentic_system:
    target:
      prompt_injection_block_rate_min: 0.95
      output_safety_pass_rate_min: 0.99
      pii_leak_rate_max: 0.01
      jailbreak_detection_rate_min: 0.9
      hallucination_rate_max: 0.1
      toxic_output_rate_max: 0.005
patterns:
  meta:
    agentic-multi-agent-required: {}
    agentic-system-required: {}
  P0:
    arch-monolith:
      scaling_strategy: horizontal
      load_balancer_type: layer7
      session_management: stateless
      deployment_strategy: rolling
    db-managed-postgres:
      provider: supabase
      instance_size: small
      storage_gb: 8
      backup_retention_days: 7
      connection_pooling: true
      high_availability: false
      ssl_mode: require
    platform-kubernetes:
      namespace_strategy: per-environment
      ingress_controller: nginx
      service_mesh: none
      operator_framework: native
      autoscaling: hpa
      network_policy: disabled
    agentic-topology--hierarchical:
      message_envelope_shape: structured_json
      response_envelope_shape: structured_json
      failure_policy_retry_max: 2
      failure_policy_retry_backoff: exponential
      failure_policy_retry_initial_ms: 1000
      failure_policy_on_terminal_failure: surface_partial
      failure_policy_cancel_siblings_on_failure: false
      aggregation_strategy: concat_with_attribution
    platform-no-mesh:
      gateway_type: api_gateway
      client_library_approach: rest_http
      load_balancing: gateway_level
      observability_strategy: application_level
    agentic-session--stateless: {}
    agentic-model-vendor--anthropic:
      enable_token_efficient_tool_use_beta: true
      prompt_caching_recommended_blocks:
      - system_prompt
      - tool_definitions
      - long_context_documents
    agentic-inference--anthropic:
      enable_token_efficient_tool_use_beta: true
    iac-terraform:
      state_backend: s3
      workspace_strategy: single
      module_structure: monorepo
      ci_cd_integration: github-actions
      drift_detection: false
    obs-agentic-system:
      log_full_prompts: true
  P1:
    agentic-output-safety-guardrail:
      detection_action: redact-and-continue
    agentic-input-safety-guardrail:
      fallback_on_guardrail_failure: fail-closed
      block_action: reject-with-explanation
    agentic-tool-scoped-credentials:
      fail_action_on_credential_error: fail-closed
    agentic-multi-pass-review:
      review_session_isolation: fully_independent
    api-rest-resource-oriented:
      pagination_style: offset
      max_page_size: 100
      versioning_strategy: uri
      filtering_style: query_params
      cache_strategy: etag
      id_format: uuid
      response_envelope: false
    agentic-validation-retry-loop: {}
    sync-request-reply-rest:
      timeout_seconds: 30
      retry_strategy: exponential_backoff
      max_retries: 3
      circuit_breaker_enabled: true
      rate_limiting_strategy: token_bucket
      idempotency_required: false
    agentic-tool-use-extraction:
      composes:
        layered_after:
        - agentic-output-safety-guardrail
    deploy-rolling:
      batch_size: 25_percent
      health_check_grace_period: 30
      rollback_on_failure: true
      deployment_strategy: replace_one_by_one
      max_unavailable: 0_percent
      composes:
        layered_after:
        - iac-terraform
    sec-auth-oauth2-oidc:
      oauth_flow: authorization_code
      token_storage: secure_storage
      pkce_enabled: true
      scope_strategy: minimal
      token_refresh: automatic
      id_token_validation: strict
      composes:
        wraps:
        - api-rest-resource-oriented
    crud-single-model:
      api_style: rest
      validation_strategy: server-side
      soft_delete: false
      audit_logging: false
      pagination_default_size: 20
    finops-cost-allocation-tags:
      tagging_strategy: hierarchical
      enforcement_level: required
      cost_allocation_model: showback
      tag_inheritance: true
      automated_tagging: true
    release-feature-flags:
      flag_storage: config_file
      evaluation_strategy: simple_boolean
      targeting_capability: none
      kill_switch_enabled: true
      audit_logging: false
    hosting-agentic-managed-ai-platform--dify:
      deployment_mode: cloud_managed
    obs-telemetry-backend:
      metrics_retention_days: 15
      log_retention_days: 7
      backend_ha: false
    obs-open-telemetry-baseline:
      export_backend: otlp
      trace_sampling_strategy: parent-based
      trace_sampling_rate: 1.0
      metrics_export_interval: 60
      log_correlation: true
      resource_detection: true
      propagation_format: w3c-tracecontext
      composes:
        co_runs_with:
        - api-rest-resource-oriented
    secrets-env-vars:
      injection_method: runtime
      secret_rotation_support: false
      prefix_convention: APP_SECRET_
      validation_on_startup: true
    finops-budget-guardrails:
      budget_period: monthly
      alert_thresholds:
      - 50
      - 80
      - 100
      enforcement_action: alert
      tagging_strategy: mandatory
      policy_enforcement: soft
      cost_allocation_level: project
    ops-slo-error-budgets:
      slo_target_percentage: 99.9
      measurement_window_days: 30
      error_budget_policy: halt-deployments
      sli_type: availability
      alerting_threshold_percentage: 80
    agentic-provenance-tracking: {}
  P2:
    api-versioning-header:
      version_header_name: API-Version
      version_format: date-based
      fallback_behavior: latest-stable
      content_negotiation: false
      deprecation_policy: warning-header
    gov-system-manifest:
      manifest_path: docs/architecture/manifest.yaml
      manifest_format: yaml
      manifest_scope:
      - agent-tools
      - agent-skills
      - agent-models
      - agent-prompts
      pin_versions: true
      ci_validation: required
      drift_policy: fail-build
      composes:
        layered_after:
        - iac-terraform
        co_runs_with:
        - release-feature-flags
        - gov-adrs-mandatory
        - ops-runbooks
  P3:
    ops-runbooks:
      runbook_format: markdown
      incident_severity_levels: 4
      escalation_policy: tiered
      automation_integration: manual
      review_frequency: quarterly
    gov-adrs-mandatory:
      adr_format: madr
      storage_location: docs/adrs
      decision_threshold: significant
      review_requirement: peer-review
✓ compiler verified: re-compiling the promoted architecture exits 0 (clean)
Handoff-ready. Write this architecture.yaml to <app-repo>/docs/architecture/architecture.yaml, commit it. skills/implementing-architecture/SKILL.md reads from that path. Recompilation of the underlying spec invalidates the approval header — fresh review required.

What this demo did NOT cover

This walkthrough kept the loop tight — one signal per step. Several facilities of the compiler and the adjacent skills were intentionally out of scope:

The compiler's pattern registry is itself documented under patterns/*.json; every pattern carries reference_design_url + reference_developer_doc_url fields pointing at the canonical product / SDK docs for that technology.