Skills

A run can give its agent procedures as Agent Skills (a SKILL.md with a name and a description) and the subagents they use, so that a task names the procedure ("Use the implement skill") instead of spelling it out in every caller's prompt. The text lives once, in this repository or the configured agent configuration repository.

Skill or subagent	Source	What it does
implement	skills/implement/SKILL.md	Make the change, run the checks, then review the whole diff with diff-reviewer; only blockers trigger a second pass, on the fix delta
review	workflow/review.md, the review caller's standing task, with a header	Review a pull request read-only and return one verdict comment
diff-reviewer (subagent)	skills/agents/diff-reviewer.md	Argue against an uncommitted change and tag each finding blocker, important or nit

Turning it on

It is off by default. [agent] skills = true in the run's configuration, or the workflow input agent-skills: true, installs them. All three real harnesses get the embedded procedures above, and also the root skills/ tree from [agent] config-repo. fake refuses the option because it cannot use skills.

The runner owns the current harness's user skills directory: each run replaces that tree, including when the option is off or no configuration repository is named. Other harnesses' directories are left alone. Embedded procedures win over source files at the same paths.

To use agenthome's skills, set agent.config-repo (or the caller's AGENT_CONFIG_REPO repository variable) to the configuration repository that owns them, enable skills, and name the skill in the task, for example "Use the ast-explore skill to investigate this issue." No repository names are built into the installer. Sources and refs belong in repository-owned configuration, not dispatcher-controlled task inputs. With skills enabled, per-agent configuration files may be absent, allowing a skills-only source; the normal generated configuration and explicit model then apply.

Skills are always read from the source repository's root skills/, not config-path (which still selects per-agent configuration). The installer accepts at most 64 non-executable regular files, at most 1 MiB combined, with UTF-8 contents and bounded, simple relative paths under skill directories. It reads committed Git blobs rather than following checkout links. These checks prevent a source tree from redirecting file reads or destination paths; they do not make skill instructions trustworthy. Do not put secrets or arbitrary configuration in that tree. Tool installation is separate: this does not execute agenthome's install-tools step or skill scripts.

The dispatch caller turns it on for implement with claude, opencode or vibe, and its task then names the skill ("Use the implement skill") in place of the procedure. Other profiles and agents keep the prompt they had. The review caller does not use it yet: its standing task is still passed as the task.

Where each agent reads them

The binary embeds the files and run writes them, as the sandbox user, into that user's home:

Agent	Skills	Subagents	Source read
claude	~/.claude/skills/NAME/SKILL.md	~/.claude/agents/NAME.md	The skills and subagents documentation; @agentclientprotocol/claude-agent-acp@0.88.0 starts the SDK with the user setting source
opencode	~/.config/opencode/skills/NAME/SKILL.md	~/.config/opencode/agents/NAME.md	opencode v1.18.35's src/skill/index.ts, src/config/paths.ts and src/config/agent.ts
vibe	$VIBE_HOME/skills/NAME/SKILL.md (~/.config/agentic-job/vibe/skills/)	$VIBE_HOME/agents/NAME.toml and $VIBE_HOME/prompts/NAME.md	Vibe 2.26.1's vibe/core/config/harness_files/_paths.py, vibe/core/skills/manager.py, vibe/core/agents/ and vibe/core/prompts/__init__.py

Vibe has no Markdown subagents. Its subagent is an agent profile with agent_type = "subagent", which its task tool runs. The profile's system_prompt_id names the reviewer's text, without frontmatter, as its system prompt. The profile sets nothing else, so the reviewer runs with the main agent's tools and permissions. The launcher gives Vibe only its user source, so a target repository's .vibe/ cannot replace them.

The launcher turns off opencode's project configuration and external skills (~/.claude/skills, ~/.agents/skills), so its global configuration directory is the only place it finds them. The subagent's mode: subagent is what makes it one for opencode; Claude Code ignores the key.

For Vibe, the vibe CI job's cargo test --test vibe (AGENTIC_JOB_TEST_VIBE_PYTHON) loads the pinned Vibe's skill and agent managers, with the launcher's user-only source, against a home that run's files were written to. It checks that they find implement, review and the diff-reviewer subagent with its prompt there. It also checks that they ignore a target repository's .vibe/ in a folder Vibe trusts, which the project source would read. It makes no model request.

The opt-in test cargo test --test skills runs the pinned opencode (opencode debug skill and debug agent) and the Claude Agent SDK the pinned adapter resolves, with the setting sources the adapter passes (its init message), against a home that run's files were written to, and checks that each finds implement, review and diff-reviewer there and that opencode ignores a target repository's own. Neither makes a model request. The ACP adapter itself and a real session with a model are not part of it. It needs both packages installed outside the repository:

npm install --no-save --prefix "$TMP/agents" opencode-ai@1.18.35 @agentclientprotocol/claude-agent-acp@0.88.0
AGENTIC_JOB_TEST_OPENCODE="$TMP/agents/node_modules/.bin/opencode" \
AGENTIC_JOB_TEST_CLAUDE_ACP="$TMP/agents/node_modules/@agentclientprotocol/claude-agent-acp" \
  cargo test --locked --test skills -- --nocapture

CI does not run it yet: an npm install outside a lockfile is a zizmor finding the ingress check refuses, so CI needs a locked install, or a maintainer's exemption, first.

What it does not protect

A skill is instructions, not a control. The run's bounds, check and the review are what stand between the agent and a merge; the self-review is the agent's diligence, capped at two passes and stopped on no blockers. Important findings and nits go in the description as follow-ups; the review job remains the merge gate. A target repository's .claude/agents/diff-reviewer.md would replace Claude Code's user subagent of that name (project subagents win there, user skills win over project ones), which can make a self-review weaker but cannot change what the run may hand back. Subagent tasks count against limits.max-tasks (the workflows' default is 4), which also bounds the review loop.

The skills are built into the binary: skills/ and workflow/review.md are part of the source digest that decides whether a published release is this source's: changing them, the review caller's task included, needs a new release before a caller that does not build the binary can use that commit.

Not yet: skills for fix, triage and research, and a caller's configuration overriding embedded procedures (#542).