Skill: init

The init-buddy-agent-harness skill is the primary way to adopt canonical agent configuration. It is an agent skill, not a script: it reads what you already have, proposes a plan, and only then writes. The CLI performs the mechanical linking step at the end.

Run it

In Claude Code, invoke the skill directly:

/buddy-agent-harness:init-buddy-agent-harness

Any agent that reads .agents/skills/ can also be asked in prose, from the repository root:

Initialize this repository's agent configuration.

The skill locates the Git repository root first. Canonical configuration always lives there, including in a monorepo.

The five phases

1. Survey

The skill inventories two surfaces and writes nothing.

The canonical surface: AGENTS.md, .agents/AGENTS.md, .agents/skills/.

The existing harness surface: instruction files, skill directories, commands, subagents, rules, MCP servers, and hooks across every supported harness. See Migrating Existing Configuration for the full detection table.

For each finding it records which of three states it is in: a real user-authored file, a symlink already resolving into .agents/ (a previous run), or something occupying a projection target that is neither (a conflict).

2. Classify

Every finding goes into exactly one bucket:

Bucket	Meaning	Action
already canonical	lives in .agents/ or root AGENTS.md	left alone
already linked	a symlink resolving into .agents/	skipped, which is what makes re-runs idempotent
consolidatable	harness instruction files whose content belongs in AGENTS.md	merged, with approval
portable	skill and command directories	moved into .agents/skills/
canonical-only	MCP servers, rules, subagents, hooks, output styles	reported and left in place

The canonical-only bucket is the important one. The skill reports those formats rather than converting them: subagents, rules, hooks, and output styles have no cross-harness format to convert into, and MCP servers have one that cannot be applied without inventing values you did not write. What stays canonical has the reasoning.

3. Confirm

Nothing you authored is touched before you see the plan. The skill presents:

• what will be created,
• which content moves into AGENTS.md,
• which harness files would become pointers,
• any frontmatter to be added, showing the derived name and description verbatim,
• which harnesses will be enabled,
• what is being left alone, and why.

Explicit approval is required before any step that deletes, replaces, or rewrites a user-authored file. Creating a missing directory or a missing AGENTS.md does not need approval.

4. Apply

1. Scaffold the baseline: .agents/, .agents/skills/, and a root AGENTS.md if absent. An existing AGENTS.md is never rewritten.
2. Move approved skills and commands to .agents/skills/<name>/SKILL.md, preferring git mv so history follows, and fix frontmatter per Portable Skills.
3. Merge approved instruction content into AGENTS.md, preserving your wording. Content is appended, not restructured.
4. Run the init command to create the projections and report the enabled harnesses.
5. Resolve any reported conflict and retry. --force replaces one named projection; --copy is only for environments without symlinks.
6. Settle the instruction files the CLI does not touch. Gemini CLI needs a .gemini/settings.json edit, and it is the only bridge left to write. No CLAUDE.md is written, at the root or beside a nested AGENTS.md: Claude Code reads AGENTS.md itself, and a generated copy would be read instead of it.

5. Verify and report

Each projection is checked to resolve into .agents/skills, and every migrated SKILL.md is checked to parse with valid frontmatter. The report states what was created, consolidated, linked, and left canonical-only.

The run ends by asking whether to run enhance, which offers guidance your AGENTS.md does not have. Decline and nothing more happens. That offer is the only opinion init carries, and even it is a question.

Rules the skill follows

These hold regardless of what you ask for mid-run:

• Never invent project policy, and never rewrite your AGENTS.md.
• Never convert tool settings between formats without a documented mapping. Unmapped settings stay canonical.
• A user-authored settings file is amended in place, never round-tripped through JSON.parse. Key order, indentation, and comments survive the edit, because harnesses disagree about whether comments are allowed.
• Symlinks belong in version control. They stay tracked and are not added to .gitignore.
• init writes AGENTS.md and .agents/skills/, and treats only those as standard. Other folders under .agents/, such as .agents/agents/ or .agents/commands/, are your project’s choice: init does not create them, leaves any you have, and does not call them standard.
• Local agent configuration only: no workflows, Actions, repository settings, security scanning, branch rules, or unrelated project files.

Agent worktree offers

Where Claude Code is in use, the skill also offers two settings, each on its own and written only on approval: ignoring .claude/worktrees/ in .gitignore and in each tree-scanning tool’s config, and setting worktree.baseRef to "head". Agent Worktrees explains why.

Which harnesses get enabled

The enabled set is the union of three sources: Claude Code and Cursor unconditionally, any harness whose directory already exists in the repository, and any you name explicitly with --harness codex,gemini-cli.

A detected harness directory means there is existing configuration to reconcile. It is a starting point for the conversation, not a standing instruction to maintain that harness. The run reports which harnesses it enabled and why.

There is no way to disable a harness, and enabling one is not the same as writing files for it. Most enabled harnesses read .agents/skills/ natively and receive nothing at all; only the projections are a real diff. See Harness Differences.

Re-running

Re-run after adding or changing a skill only if you need a new harness enabled. A directory-level symlink is live: a skill added to .agents/skills/ afterwards appears in every enabled harness with no further action. A --copy fallback is a snapshot instead, and does need a re-run.

A projection can also stop resolving later, most often on a clone that could not create the symlink. The doctor-buddy-agent-harness skill is what reports that.