Skip to main content

Concrete Systems

Concrete systems are what happens when a specialist harness composition meets a real task family. They are not abstract design patterns -- they are installable, reviewable, verifiable packages that compose skills, workflows, adapter bindings, eval packs, and a harness profile into a named automation surface.

How concrete systems work

Every concrete system follows the same structure:

  1. A task family (e.g. docs, research, code-review, release-validation).
  2. A harness profile that defines routing, tool projection, verification, and rollout policy for that family.
  3. A workflow that sequences steps -- search, analysis, synthesis, review, evidence capture.
  4. Skills that provide reusable instructions and contracts for each step.
  5. Adapter bindings that connect portable capability contracts to concrete execution (CLI, subprocess, API, frozen-tool).
  6. An eval pack that defines fixtures, graders, and no-regression checks.
  7. Verification and activation posture that gates live use behind evidence and review.

The result is a named, shareable, versioned package that turns "help me with documentation research" into a governed, verifiable workflow rather than a freeform chat prompt.

Current posture

Holon's concrete systems are currently defined at the spec level. The Specialist Harness Contract (repo-internal spec at specs/holon-specialist-harness-contract/spec.md) defines the packaging convention. The Plugin V2 Core Contract (specs/holon-plugin-v2-core/spec.md) defines the asset families that compositions draw from. Live marketplace distribution, activation, and UI-driven install are deferred.

The first concrete system prototype targets a low-risk local-draft docs-research composition. This validates the manifest, search, install, review, and verification model without opening unsafe execution surfaces (no 3D modeling, coding delegation, third-party UI, or live piloting).

Task families

Task families are stable labels drawn from the same vocabulary used by search-before-build and harness self-upgrade. Examples include:

Task familyWhat it coversTypical skills and assets
docsDocumentation research, authoring, review, drift detectionsource-trail, citation-check, synthesis, link-validation
researchBackground investigation, OSS analysis, comparison notesweb-search, paper-summarization, evidence-collection
code-reviewStructured code review against contracts and patternsdiff-analysis, spec-alignment, regression-check
release-validationPre-release checks, changelog verification, build gatesversion-check, changelog-alignment, artifact-inspection
planningGoal decomposition, roadmap drafting, prioritizationgoal-extraction, dependency-analysis, estimation

Each family is a search term, a profile scope, and a composition namespace. The same family labels are used in search-before-build queries to find existing specialist harnesses before building new ones.

Trust and activation

Concrete systems follow the same trust model as other v2-core packages:

  • trustChannel labels (core, official, verified-publisher, community, public, local-draft, private-org, personal-private) control install friction, activation friction, and update posture.
  • activationPosture (catalog-only, draft-installable, review-gated) controls whether the composition can be activated for live use.
  • Eval packs and evidence refs gate activation for non-local-draft packages.
  • The default posture is catalog-only -- discoverable and reviewable, but not activatable -- to prevent untrusted harnesses from reaching live execution.

Where to go next