Agent Harness

🗂️ Skills

Skills live in .agents/skills/ and are shared across every agent.

⚡ Quickstart

1. Install the skills with the skills.sh installer:

   npx skills@latest add codelytv/agent-harness
   

2. Pick the skills you want and the coding agents you want to install them on.

3. Run a skill in your agent, for example /codely-doc-create.

4. Done. Your agent now follows Codely's conventions.

Alternative: Claude Code plugin

1. Add this repository as a plugin marketplace and install the plugin:

   /plugin marketplace add CodelyTV/agent-harness
   /plugin install codely-skills@codely
   

2. Run a skill in Claude Code, for example /codely-doc-create.

3. Done. Every skill in this repo is now available in Claude Code.

📦 Available skills

Skills are grouped by category:

📃 Harness

Skill	What it does
codely-doc-create	Generates convention documentation from the current conversation, turning feedback and corrections into reusable docs.

Two ways to use codely-doc-create:

• After a conversation — run /codely-doc-create to turn the corrections the agent received during the session into a new doc.
• Before a conversation — run /codely-doc-create <description> to formalize a convention you want to document upfront.

🗺️ RPI

Skill	What it does
codely-plan-create	Breaks a task into reviewable phases (vertical slices), defines the public contracts to change, and saves an approved plan file.
codely-plan-create-subagents	Same planning flow, but delegates the codebase exploration to cheaper subagents running in parallel.
codely-plan-phase-implement	Executes a single phase of a plan at a time, updates its checkboxes, and stops for review without committing automatically.
codely-plan-create-github	Same planning flow, but stored as GitHub issues: a parent plan issue and one native sub-issue per phase.
codely-plan-phase-implement-github	Implements one phase issue at a time on its linked branch and opens a pull request that closes the issue on merge.
codely-plan-create-gitlab	Same planning flow, but stored as GitLab issues: a parent plan issue and one linked child issue per phase.
codely-plan-phase-implement-gitlab	Implements one phase issue at a time on its linked branch and opens a merge request that closes the issue on merge.

Typical flow:

• Plan the work with /codely-plan-create <task> (or /codely-plan-create-subagents <task> to delegate the exploration) to produce a phased plan file under .agents/plans/.
• Implement it phase by phase with /codely-plan-phase-implement <plan-file-path>, reviewing and committing after each one.

⚫ OpenSpec

Skill	What it does
codely-openspec-spec-create-from-code	Writes the OpenSpec specs of behaviour that already exists in the code, so the next change can modify or remove it instead of only adding to it.

Two ways to use codely-openspec-spec-create-from-code:

• Before a change — run /codely-openspec-spec-create-from-code <the change you are about to make> to specify the current behaviour it will touch, then continue with /openspec-explore.
• To catch up — run /codely-openspec-spec-create-from-code <capability> to specify an existing feature, or without arguments to list the features that still lack a spec.

Only intended and verified behaviour reaches the spec: bugs, accidental behaviour and dead code are reported instead.

🧩 codely-spec-plan schema

An OpenSpec custom schema that brings the codely-plan-create-subagents flow to OpenSpec. Each change produces only two things to review:

• Lean specs: one behaviour with SHALL per requirement and a single-line scenario.
• plan.md: goal, decisions and one vertical slice phase per part a user can complete a goal with, each with its contracts and a test strategy that proves it with few tests. Its checkboxes are the task list that /openspec-apply-change works through one phase per run, committing each phase.

Install it in your project and select it in openspec/config.yaml:

npx degit CodelyTV/agent-harness/openspec/schemas/codely-spec-plan openspec/schemas/codely-spec-plan

schema: codely-spec-plan

🔀 Git

Skill	What it does
codely-git-conventional-commit	Creates a Git commit following the team's Conventional Commits conventions and co-author trailer.

Run /codely-git-conventional-commit to stage and commit your changes with a conventional message.

🗣️ Output styles

Output styles change how the agent writes to you, not how it writes code.

Plugin	What it does
ste100	Answers in ASD-STE100 Simplified Technical English: short sentences, active voice, one word per meaning, no idioms.

/plugin marketplace add CodelyTV/agent-harness
/plugin install ste100@codely

Then run /config, look for Output Style, and select ASD-STE100.

🧠 Code intelligence

Plugin	What it does
ts7-lsp	TypeScript/JavaScript go-to-definition, references, and diagnostics through the TypeScript 7 native compiler LSP.

The native compiler (the Go rewrite formerly known as tsgo) indexes large monorepos roughly 10x faster and with far less CPU than the official typescript-lsp plugin, which spawns typescript-language-server + tsserver.

Requires a TypeScript 7+ tsc binary on PATH (e.g. npm install -g typescript).

/plugin marketplace add CodelyTV/agent-harness
/plugin install ts7-lsp@codely
/plugin disable typescript-lsp@claude-plugins-official

🔗 Unified rules and skills via .agents/

Each AI agent reads instructions from a different path. Maintaining them separately is error-prone, so this project centralizes everything and uses symlinks so each agent reads from its expected path while the content lives in a single place:

1. Rules are written once in AGENTS.md (one per directory if needed).
2. Skills live in .agents/skills/ and are shared across agents.
3. A make command generates the symlinks each agent expects:

Command	What it does
make claude-symlinks	Creates a CLAUDE.md → AGENTS.md symlink in every directory that has an AGENTS.md, and links .claude/skills → .agents/skills
make codex-symlinks	Links .codex/skills → .agents/skills
make copilot-symlinks	Links .github/skills → .agents/skills
make cursor-symlinks	Links .cursor/skills → .agents/skills
make junie-symlinks	Links .junie/skills → .agents/skills
make opencode-symlinks	Links .opencode/skills → .agents/skills

🛡️ export command blocked via hooks

The export command can leak environment variables (tokens, secrets) if an agent runs it. To prevent this, each agent has a pre-execution hook that blocks any shell command containing export:

Agent	Hook location	Mechanism
Claude Code	.claude/settings.json	Uses the if keyword with a glob pattern (Bash(*export*)) to match and block inline
Cursor	.cursor/hooks.json + hooks/block-export.sh	Runs a shell script that parses the command via jq and exits with code 2 on match
Copilot	.github/hooks/hooks.json + hooks/block-export.sh	Same approach as Cursor, adapted to Copilot's hook input format
OpenCode	.opencode/plugins/block-export.ts	A native TypeScript plugin (tool.execute.before) that inspects the command and throws to block