Operations¶
This page covers the operational details maintainers should know before adopting Agent Workbench in a repository.
Project-local data¶
Agent Workbench creates .agent-workbench/ in the project.
Important paths:
| Path | Purpose |
|---|---|
.agent-workbench/ledger.sqlite |
Structured project ledger. |
.agent-workbench/designs/ |
Design Package drafts and imported sources. |
.agent-workbench/logs/ |
Validation and command logs. |
.agent-workbench/exports/ |
Human-readable exports. |
The ledger is operational state. Decide per project whether .agent-workbench/
should be kept local, archived, or committed.
Skill installation scope¶
gh skill install --scope user installs the skill for the user.
gh skill install --scope project installs it for the current project.
Project-scope installs create an agent-specific installed skill directory, such
as .agents/skills/. That directory is an installed environment artifact. Do
not treat it as the source skill package.
The source skill package in this repository is skills/agent-workbench/.
Public release asset¶
The installed skill wrapper fetches the pinned Linux x86_64 CLI from GitHub
Releases. Release binaries are built on ubuntu-latest and target a
glibc-compatible Linux runtime.
The release contains:
agent-workbench-<tag>-linux-x86_64.tar.gzagent-workbench-<tag>-checksums.txt
The wrapper verifies the archive before executing the cached CLI.
For project-scope development installs that live inside an Agent Workbench
source checkout, the wrapper first uses an already-built checkout binary from
target/debug/agent-workbench or target/release/agent-workbench. Normal
installed skills without a source checkout use the pinned release path.
Publishing a new version means creating a release that includes both the archive and checksum file. Published release assets are immutable; use a new tag instead of replacing an existing release.
Public documentation¶
The public documentation is built from docs/mkdocs.yml and deployed to GitHub
Pages by .github/workflows/docs.yml.
The docs workflow runs when public docs or the docs workflow change on main.
It builds with MkDocs strict mode, uploads the generated site/ artifact, and
deploys it through GitHub Pages.
Cache¶
The wrapper caches the CLI under the user's cache directory, using
XDG_CACHE_HOME when set.
To force a fresh download, remove the cached release directory.
Version pinning¶
Set AGENT_WORKBENCH_VERSION when the agent should use a specific release tag.
Set AGENT_WORKBENCH_REPO only when testing a fork or alternate release source.
Most users should not set either variable.
Privacy¶
The ledger can contain:
- task titles
- user corrections
- command names and logs
- design summaries
- Git commit and file references
- review findings
Treat .agent-workbench/ as project operational data and apply the same privacy
rules you use for logs and design documents.