Agentic Harness Agents

Installable Agent Skills for Codex, plus compatible procedures and adapters for Claude Code, Cursor, GitHub Copilot, Gemini CLI, and other coding agents.

Current distribution version: 0.5.0-beta.1.

This repository owns procedure, not canonical project architecture. Canonical project truth lives in agentic-harness; deterministic composition, analysis, compilation, and audits live in agentic-harness-cli.

agentic-harness
project contract + catalog + policies + schemas + registries
        ↓
agentic-harness-agents
skills + prompts + adapters + behavior evals
        ↓
agentic-harness-cli
deterministic composition, analysis, migration, audit, validation

Quick start

Install one skill in Codex

Use Codex's built-in $skill-installer with a GitHub skill directory, then restart Codex so the skill is rediscovered:

$skill-installer install https://github.com/powerpuff-kitty/agentic-harness-agents/tree/main/skills/agentic-app

Replace agentic-app with any directory under skills/. For reproducible team installs, prefer a release tag instead of main.

Use skills project-locally

Codex discovers repository skills from real directories under .agents/skills/:

project/
├── AGENTS.md
├── .agentic/
└── .agents/
    └── skills/
        ├── agentic-app/
        │   └── SKILL.md
        └── security-review/
            └── SKILL.md

agentic-harness-cli installs selected skills into that location during composition.

Install the whole collection as a Codex plugin

This repository is a skills-only Codex plugin. Its native manifest is .codex-plugin/plugin.json, and its GitHub marketplace manifest is .agents/plugins/marketplace.json.

Workspace admins can import the repository as a GitHub plugin marketplace from Workspace settings → Plugins → Add → Import marketplace. Use the repository URL, leave Path empty, and select a branch, tag, or commit depending on whether you want automatic updates or an immutable version.

No external app or account authorization is required because this plugin contains skills only. Individual skills may use project-authorized tools/providers when a task explicitly requires them.

Canonical target model

project/
├── AGENTS.md
├── normal project files
├── .agentic/       project truth, governance, decisions, plans, tasks, docs, evals
└── .agents/skills/ installed reusable procedures

Skills read only the context relevant to the current task. Root AGENTS.md stays compact, .agentic/ remains canonical project truth, and vendor adapters remain thin.

Repository map

skills/                     installable Agent Skills
prompts/                    short task entrypoints
adapters/                   vendor-specific integration guidance
references/                 shared source, safety, context, and skill-contract rules
evals/                      behavior and routing acceptance criteria
.codex-plugin/plugin.json   native Codex plugin manifest
.agents/plugins/            GitHub marketplace manifest
manifest.json               version, compatibility, and skill inventory
.agentic/                   this repository's own project context

Skill contract

Every skills/<name>/SKILL.md follows the Agent Skills format:

• YAML frontmatter with only name and description;
• the description owns triggering and exclusions;
• concise procedural body using progressive disclosure;
• explicit Objective, Inputs, Context, Procedure, Output, and Completion sections;
• project truth remains external to the skill.

See references/skill-contract.md, references/context-engineering.md, and references/design-intelligence.md.

Core workflows

• agentic-app — initialize, migrate, upgrade, or audit an agent-native project.
• agentic-structure-audit — score Agentic Readiness without conflating it with ordinary code quality.
• model-fit — compare evidence-backed model profiles for a repository or task.
• agentic-improvement — improve instructions, task context, evidence reuse and handoff efficiency without requiring a CLI report.
• decision-intelligence — design bounded evidence-backed judgments, abstention and independently optional Jev integration.
• code-quality — review formatting, linting, type-safety and maintainability evidence using project-native tools and bounded fixes.
• seo-audit — audit public-web crawlability, indexability, rendering, canonical/localized URLs, structured data, programmatic pages and provider-specific AI discovery evidence.
• migration — migrate legacy repository layouts safely and idempotently.
• adr-management — create, supersede, and index durable decisions.
• adapter-sync — keep vendor adapters thin and canonical.
• specialist review/design/delivery skills cover security, accessibility, performance, API/data, product, design, releases, incidents, and research.

Skills-first efficiency and decisions

The two procedures above contain conditional skill-local guides, so their guidance path does not silently depend on a collection-wide file, CLI executable or TypeSafe account. They are ordinary skill directories, not newly published standalone release archives. Copy/install the whole selected skill directory, preserve customised local content and use the host's verified loading workflow. This change does not update existing device/project installations automatically.

decision-intelligence defaults to the current coding agent for guidance. It does not provide a local Jev model, make coding-agent inference free, enable hosted calls or authorise actions. Actual TypeSafe integration uses its independently installed official skill and current vendor docs with explicit provider/data permission. agentic-improvement preserves mandatory evidence/checks and keeps unknown usage unknown; neither procedure guarantees token savings or host enforcement.

Design Intelligence hierarchy

The design suite separates evidence, identity, UX, reusable UI systems, implementation sources, and verification instead of using one catch-all design skill:

design-intelligence            lifecycle orchestration: Analyze → Preserve → Compile → Verify
├── design-analysis            interpret measured evidence and imported analysis
├── design-research            visual/UI/UX pattern and flow research
├── identity-design             art direction, distinctiveness, brand-facing visual identity
├── product-design             user journeys, states, hierarchy, recovery, product UX
├── design-system              reusable tokens, components, states, layouts, patterns
├── component-resolution       project/internal/external implementation primitives
├── design-system-compliance   structural conformance and bypass detection
└── accessibility-audit        accessibility-specific verification

A Design Genome becomes design/identity authority only when the target project designates that artifact/version as accepted truth. Design Analysis is evidence, Design Task is structured implementation scope, and Design Analysis Diff is measurable drift evidence rather than a violation or subjective quality score. References, provider components, AI interpretations, and market prevalence may propose changes but do not silently become project requirements or accepted identity.

See manifest.json for the complete 33-skill inventory.

Validation

Local deterministic validation covers skill frontmatter, trigger overlap, routing, distribution, compatibility pins, and Design Intelligence safety/handoff fixtures:

python3 .github/scripts/validate_agents.py
python3 .github/scripts/validate_design_intelligence.py
python3 .github/scripts/package_plugin.py

evals/design-intelligence.json is a synthetic behavior contract fixture. It does not claim that any model has passed the scenarios; recorded model-behavior evaluation remains a separate future/run-specific artifact.

The new guidance-efficiency cases and grader exercise rubric integrity, missing/stale evidence, provider boundaries, verification claims and usage accounting through the same validation entrypoint. They are synthetic tests, not recorded model runs or a measured saving percentage.

CI runs the same validators and builds the versioned plugin ZIP/checksum under dist/.

License

Authored code and content are available under the MIT License. Third-party material retains its existing licenses and attribution requirements. Copied Harness templates and skills retain their MIT notice; independently written application code may use its own license.