doodooms/agentic-core
Canonical agents, skills, MCP integrations, and Expertise Pack authoring for Plugin Factory.
WHAT: Define technical solution structure, interfaces, and system boundaries. USE FOR: architecture briefs, API contracts, component topology, integrations, migrations, and material technical proposals. DO NOT USE FOR: product/domain semantics, delivery sequencing, implementation, runtime diagnosis, QA execution, or final acceptance.
WHAT: Retrieve, preserve, and measure only task-relevant agent context. USE FOR: evidence gaps in multi-agent work, durable context management, phase-boundary compaction, or local context/token projection reviews. DO NOT USE FOR: creating a competing specification/task store, hosted token counting, or reducing required safety and product behavior.
WHAT: Coordinate context-aware compaction, budget-aware work distribution, and bounded session communication. USE FOR: live context checks, Copilot/Codex budget snapshots, and authorized same- or cross-harness handoffs. DO NOT USE FOR: general delegation, implementation ownership, quota scraping, or silently resuming another agent.
WHAT: Safely configure, validate, install, package, deploy, and document software delivery/runtime surfaces. USE FOR: CI/CD, build and release pipelines, approved plugin installation, deployment/runtime configuration, observability, operational checks, or assigned documentation synchronization. DO NOT USE FOR: product feature implementation, architecture ownership, dynamic QA, or final review.
WHAT: Coordinate dependency-aware software delivery, Control Plane task state when needed, specialist handoffs, and repository lifecycle. USE FOR: non-trivial delivery, planning, cross-agent coordination, material changes, migrations, or authorized Git conflict handling. DO NOT USE FOR: architecture decisions, implementation, QA execution, final review, or operational mutations.
WHAT: Author, validate, and evolve packaged agent expertise and its plugin integrations. USE FOR: agent/skill/hook authoring, MCP server implementation guidance, Expertise Pack creation or updates, and skill-package evaluation. DO NOT USE FOR: product implementation, plugin installation/deployment, or general operational configuration.
WHAT: Design, assess, and independently falsify software behavior and its evidence. USE FOR: test design or review, adversarial QA, unknown runtime failures, browser journeys, code review, language-level review, or measurable agent evaluations. DO NOT USE FOR: production repair, architecture decisions, or final delivery acceptance.
WHAT: Gather and synthesize traceable repository or external technical evidence for a bounded decision. USE FOR: execution-path exploration, standards and version research, GitHub history, scholarly evidence, compatibility, or multi-source comparisons. DO NOT USE FOR: implementation, product decisions, planning ownership, QA verdicts, or final acceptance.
WHAT: Assess and test software trust boundaries, data exposure, database behavior, and security controls. USE FOR: static security design review, adversarial security testing, authentication/authorization, input handling, secrets, persistence, SQL, or migrations. DO NOT USE FOR: general QA, product repair, deployment operations, or final acceptance.
WHAT: Establish a sparse, evidence-grounded model of the problem space. USE FOR: material domain changes involving concepts, terminology, identity, relationships, lifecycle, invariants, contracts, assumptions, or unknowns. DO NOT USE FOR: product decisions, technical architecture, implementation, or mandatory modeling of trivial work.
WHAT: Select local implementation methods for behavior, data, algorithms, state, synchronization, and user interfaces. USE FOR: testable code changes, non-trivial local design, frontend behavior, measured performance problems, disposable prototypes, or behavior-preserving refactors. DO NOT USE FOR: system architecture, product semantics, independent QA, or operational deployment.