Repository manualdocs/CAPABILITY_ARCHITECTURE.md

Capability Architecture

Canonical documentation rendered from the repository docs/ folder.

referenceSource file

Capability Architecture

The suite uses control prompts, standalone investigations, standalone production or operations, independent gates, and genuine investigate→execute or brief→produce pairs. Pairing is a semantic decision, never a symmetry target.

Pair test

A pair is justified only when the first prompt remains non-mutating, produces a stable reviewable handoff, the second owns a distinct production or mutation responsibility, and both share objective acceptance criteria. Compare the pair against a single full-cycle prompt periodically; remove the pair when it no longer improves safety, reviewability, parallelism, or output quality.

Shape counts

  • Control: 5
  • Investigative: 93
  • Executive: 32
  • Operational: 66
  • Gates: 3
  • Genuine pairs: 32

Genuine pairs

Investigation or briefExecution or production
MD-25 Project Consistency, Uniformity, and Coherence — Investigation and PlanMD-26 Project Consistency, Uniformity, and Coherence — Authorized Execution and Verification
MD-27 Code Quality and Maintainability — Investigation and PlanMD-28 Code Quality and Maintainability — Authorized Execution and Verification
MD-29 Debugging, Root Cause, and Bug Resolution — Investigation and PlanMD-30 Debugging, Root Cause, and Bug Resolution — Authorized Execution and Verification
MD-31 Safe Cleanup, Dead Code, and Repository Hygiene — Investigation and PlanMD-32 Safe Cleanup, Dead Code, and Repository Hygiene — Authorized Execution and Verification
MD-33 Refactoring and Modernization — Investigation and PlanMD-34 Refactoring and Modernization — Authorized Execution and Verification
MD-35 Testing and Verification — Investigation and PlanMD-36 Testing and Verification — Authorized Execution and Verification
MD-37 Security Posture, Vulnerability, and Fixes — Investigation and PlanMD-38 Security Posture, Vulnerability, and Fixes — Authorized Execution and Verification
MD-39 Dependency and Software Supply Chain Security — Investigation and PlanMD-40 Dependency and Software Supply Chain Security — Authorized Execution and Verification
MD-41 Configuration, Environment, and Secrets Security — Investigation and PlanMD-42 Configuration, Environment, and Secrets Security — Authorized Execution and Verification
MD-43 Identity, Authentication, Authorization, and Session Security — Investigation and PlanMD-44 Identity, Authentication, Authorization, and Session Security — Authorized Execution and Verification
MD-45 API, CLI, and Public Contract Integrity — Investigation and PlanMD-46 API, CLI, and Public Contract Integrity — Authorized Execution and Verification
MD-47 Data Storage, Schema, and Migration — Investigation and PlanMD-48 Data Storage, Schema, and Migration — Authorized Execution and Verification
MD-49 Performance and Efficiency — Investigation and PlanMD-50 Performance and Efficiency — Authorized Execution and Verification
MD-51 Reliability, Resilience, and Recovery — Investigation and PlanMD-52 Reliability, Resilience, and Recovery — Authorized Execution and Verification
MD-53 CI, Build, Delivery, and Release Pipeline — Investigation and PlanMD-54 CI, Build, Delivery, and Release Pipeline — Authorized Execution and Verification
MD-55 Observability, Detection, and Operations — Investigation and PlanMD-56 Observability, Detection, and Operations — Authorized Execution and Verification
MD-57 Cloud, Container, and Infrastructure Security — Investigation and PlanMD-58 Cloud, Container, and Infrastructure Security — Authorized Execution and Verification
MD-59 Privacy and Data Protection — Investigation and PlanMD-60 Privacy and Data Protection — Authorized Execution and Verification
MD-61 Agent, Skill, Tool, and MCP Capability Lifecycle — Investigation and PlanMD-62 Agent, Skill, Tool, and MCP Capability Lifecycle — Authorized Execution and Verification
MD-63 Agentic Security, Permissions, and Trust Boundaries — Investigation and PlanMD-64 Agentic Security, Permissions, and Trust Boundaries — Authorized Execution and Verification
MD-65 Multi-Agent Architecture and Agent Definitions — Investigation and PlanMD-66 Multi-Agent Architecture and Agent Definitions — Authorized Execution and Verification
MD-67 Model and AI System Security — Investigation and PlanMD-68 Model and AI System Security — Authorized Execution and Verification
MD-69 Prompt Injection, Untrusted Input, and Tool-Output Security — Investigation and PlanMD-70 Prompt Injection, Untrusted Input, and Tool-Output Security — Authorized Execution and Verification
MD-71 User Experience, Accessibility, and Localization — Investigation and PlanMD-72 User Experience, Accessibility, and Localization — Authorized Execution and Verification
MD-73 Brand, Design System, and Content Coherence — Investigation and PlanMD-74 Brand, Design System, and Content Coherence — Authorized Execution and Verification
MD-75 Documentation, Knowledge, Taxonomy, and Troubleshooting — Investigation and PlanMD-76 Documentation, Knowledge, Taxonomy, and Troubleshooting — Authorized Execution and Verification
MD-77 Prompt Suite Consistency, Security, and Coherence — Investigation and PlanMD-78 Prompt Suite Consistency, Security, and Coherence — Authorized Execution and Verification
MD-83 Research-Backed Blog, Article, and Newsletter — Investigation and Editorial BriefMD-84 Research-Backed Blog, Article, and Newsletter — Drafting, Editing, and Publication Package
MD-85 Script, Video, Podcast, and Storyboard — Research and Production BriefMD-86 Script, Video, Podcast, and Storyboard — Writing and Production Package
MD-87 Academic Paper and Scholarly Manuscript — Research Design and Evidence PlanMD-88 Academic Paper and Scholarly Manuscript — Drafting, Citation, and Revision Package
MD-101 Presentation Narrative, Story Architecture, and Deck BlueprintMD-102 HTML, CSS, and JavaScript Slides and Presentation Production
MD-117 Learning Experience, Curriculum, Course, Lesson, and Assessment — Research and Design BriefMD-118 Learning Experience, Curriculum, Course, Lesson, and Assessment — Material Production and Validation

Standalone design rules

  • Use a standalone investigation when its decision or brief may feed many different downstream actions.
  • Use a standalone operational prompt when it owns a bounded end-to-end artifact or workflow with its own verification.
  • Use a gate when independence from the producer is material.
  • Do not split research, review, simulation, incident response, communication, reporting, or planning merely to create visual symmetry.

New organizational capabilities

The organizational expansion remains standalone by default. Feature delivery, experiments, localization, media post-production, academic review, department operations, OSINT, legal, finance, people, procurement, product, reporting, prompt engineering, and LLM governance compose with existing prompts through scenarios rather than artificial twins.

Permanent capability identity

Each prompt has a permanent capability_id, stable prompt_slug, current prompt_id, and presentation sequence. Integrations route by capability_id; sequence is never an integration key. Current prompt IDs are not reassigned after this release. Historical numeric IDs are namespaced and resolved through compatibility/.

Proof obligations by shape

  • Control: requires deterministic contract and policy tests.
  • Investigation: requires healthy, problematic, and adversarial evidence fixtures.
  • Executive: additionally requires contaminated-handoff refusal tests, dry-run checks, and rollback behavior.
  • Operational: requires artifact-shape, authorization, and verification fixtures.
  • Gate: requires independence, false-pass, and false-block evaluation.
  • Genuine pair: requires pair-vs-single evidence and periodic rejustification.

A structurally valid prompt remains behaviorally unproven until representative model executions are measured.

Reusable gates instead of prompt inflation

The gate registry provides independent evidence, publication, human-decision, production-change, and multi-artifact-coherence checks. These are policy-level reusable reviewers and do not increase the prompt count merely to create symmetric files.

Prompt-body role contract

Role selection changes the body contract, not only metadata:

  • Every role declares a canonical authorization boundary, tool policy, runtime markers, output contract, and task-specific completion criteria.
  • Investigators remain read-only and create evidence, findings, proposed actions, and acceptance criteria.
  • Executives contain domain-specific decision rules and reference the frozen acceptance criteria instead of copying verification lists.
  • Operational prompts own bounded full-cycle artifacts and use mode-aware tool policies.
  • Gates remain independent and do not repair the subject they judge.

The generated BODY_QUALITY_AUDIT.json and focused regression tests enforce these invariants.

Conditional auto-orchestration

MD-191 through MD-198 are standalone conditional capabilities. They are not a forced pair family and do not run as a universal preamble. Each has an explicit trigger and non-trigger. Generic skill execution and generic looping are separated from acquisition and exit adjudication so availability cannot silently become authority.

Conditional auto-orchestration shape

MD-191 through MD-198 form a conditional control layer, not a new twin family. Intent clarification, skill-fit analysis, discovery, acquisition, creation, execution, looping, and exit adjudication remain distinct because each has different authority and evidence. Any prompt, scenario, or exact installed skill may be the target of MD-197; eligibility depends on measurable value, not domain. Any exact installed skill may be bound through MD-196; availability alone never establishes genuine need.

Paired review and exact-twin control

True investigative/executive pairs include an explicit human review layer between handoff_frozen and execution preparation. The planner presents the plan, incorporates requested revisions, re-verifies and re-freezes the handoff, and obtains review again. Explicit execution consent then authorizes only the reciprocal paired_prompt_id; the runtime rejects substitute executors.

Prompt-ingestion mutation boundary

MD-199 owns agentic prompt-addition analysis. tools/add_prompt.py owns canonical mutation. The boundary separates semantic judgment from deterministic identity allocation and registry updates: the agent may refine and propose, but only the transaction tool may write the prompt ecosystem. Unknown routes fail closed, skill references never imply installation, and paired capabilities cannot be synthesized by the generic importer.