Skills
The aiur-run Executor announces fallback to ~/.aiur/config when no local config exists. It can launch a new GitHub repository without per-repository init: startup ensures workflow/marker and complexity labels, while models are selected by configuration. It does not seed model/effort labels.
Shared credentials can live in ~/.aiur/.env outside Git. Local settings, permission requirements, pauses and human decisions retain precedence.
aiur-intro explains one-time home setup and per-repository readiness: GitHub access, base branch, CODEOWNERS or dispatch allowlist, labels and validation. The Executor tells the user what was configured and what still blocks startup or worker publication.
Aiur ships Agent Skills under .claude/skills/, with shared links for Codex under .codex/skills/ and Muse under .agents/skills/ in the source checkout. They split into two families by where they run:
- Agent-workspace skills are copied into every ticket workspace, so the agent working a ticket can load them on any repository.
- Executor skills stay in this repository and load in the Executor's own session, whether that Executor is a human or an agent driving Aiur.
Agent-workspace skills
These three skills, together with the complete pinned Compound Engineering set, are available in every ticket workspace under <workspace>/.claude/skills/, <workspace>/.codex/skills/, and <workspace>/.agents/skills/.
Muse loads workspace skills and rules only when agent.backend_configs.muse.trust_workspace is explicitly enabled.
| Skill | Loaded when | Covers |
|---|---|---|
| aiur-agent | Every ticket turn, via the shared per-turn prompt pointer | The agent:* label lifecycle, the brainstorm→plan→work→review flow, the Agent Workpad, milestone alerts, complexity routing, the dev loop, and Message Bus events, subscriptions, blockers, and attentions. |
| aiur-debug | Run, daemon, agent, or workspace failure | Correlated evidence and safe recovery order. |
| design-import | Large frontend design import | Disk-first import without overflowing agent context. |
Executor skills are deliberately excluded from this set: an issue worker has no reason to run Aiur itself.
The aiur-agent dev loop distinguishes a retryable GitHub budget hold from a credential failure. A budget-held agent requests a typed, expiring pause and is resumed automatically; it does not raise a credential attention for a healthy token.
Issue workers select focused tests from the target repository's instructions, package scripts, and CI configuration. The bundled aiur-agent skill's mix examples apply to Aiur's Elixir core; a TypeScript workspace uses its own runner.
The aiurdev --test / --test3 guard prevents destructive Aiur sandbox resets from issue workspaces, not ordinary repository tests.
Executor skills
These stay with the Executor and are not copied into ticket workspaces.
| Skill | Trigger | Covers |
|---|---|---|
| aiur-intro | First contact or installation | Operating modes, setup, and first run. |
| aiur-build | Plan a large feature | Requirements, ticket contracts, dependency graph, and Build Order handoff. |
| aiur-run | Operate a run | Accepted boundary through review and merge. |
| aiur-monitor | Inspect a live run | Status board and alert feed. |
| aiur-meta | Hourly audit | Operator surfaces, bottleneck, and durable findings. |
| aiur-handoff | Executor handoff | Boot document for the next Executor. |
| release | Release | Version, tag, and GitHub release. |
aiur-handoff, aiur-meta, and release are Claude-only.
Codex-native git workflow
| Skill | Covers |
|---|---|
| commit | Scope-aware commit creation. |
| push | Push and PR creation or update. |
| pull | Merge-based integration from origin/main. |
| land | Conflict, CI, review, and squash-merge handling. |
| debug | Log-tracing recovery. |
| linear | Raw Linear GraphQL operations. |
Compound Engineering
Aiur vendors the complete Compound Engineering 3.19.0 skill tree under .claude/skills/, so every dispatched workspace receives the same pinned set.
| Skill | Purpose |
|---|---|
ce-brainstorm | Clarify requirements and alternatives. |
ce-plan | Produce an implementation-ready plan. |
ce-work | Execute a concrete plan or request. |
ce-code-review | Review bugs, regressions, tests, and standards. |
ce-doc-review | Review documentation structure and accuracy. |
ce-debug | Diagnose failing behavior. |
The complexity router invokes those skills directly, and they conditionally invoke sibling CE skills; the complete tree keeps those branches reproducible.
| Pin record | Location |
|---|---|
| Version | compound-engineering.version |
| Managed skill names | compound-engineering.skills |
| Upstream MIT license | compound-engineering.LICENSE |
To update the vendored copy, clone the exact upstream release tag and run the guarded refresh script:
update_dir=$(mktemp -d)
git clone https://github.com/EveryInc/compound-engineering-plugin.git "$update_dir/compound-engineering"
git -C "$update_dir/compound-engineering" checkout <exact-release-ref>
scripts/update-compound-engineering-skills X.Y.Z "$update_dir/compound-engineering"Choose the release ref whose .claude-plugin/plugin.json reports X.Y.Z; the refresh script rejects a mismatch.
Review the upstream release notes and the resulting skill diff, then run the AgentSkills tests before committing. The script replaces only previously managed CE skill paths, refreshes the license, version, and manifest files, and recreates the Claude-to-Codex links.
The complexity routing rules that pick which CE skill to run live in complexity-routing.md.