V1 long-form archive. Pre-precision-pass version, preserved for engineers who prefer the dense narrative. New (V2) version: ../faq.html.

Why a Workbench per Epic Bundle, Not a Permanent Hub?

Bundles are disposable. An epic (or a small set of related epics) gets picked up, planned, shipped, then closed. A workbench matches that lifespan — start one, work in it for weeks or a quarter, archive it. A permanent team-wide hub drifts, accumulates stale context, and becomes a second Jira. This harness is explicitly throwaway.

Do I Have to Use Devin?

No. Every devkit command has a .cly variant that forces Claude: init.wb.cly, join.wb.cly, update.wb.cly. Inside a workbench, slash commands work the same under Claude or Devin because both read the same skills/ directory. .dev variants force Devin if you want to be explicit.

What Happens If Both Dev and QA Edit project.conf at the Same Time?

Git merge conflict. Resolve it normally and push. The file is a shell script — conflicts usually show up as duplicated array entries, which are easy to reconcile. Pull before starting a session and push when you pause; that is the simplest way to avoid the situation.

Can I Run Ralph Without Workspace-Mode?

Yes. scripts/ralph-plan.sh falls back to per-repo planning: it iterates each registered repo, runs ralph-plan inside, and stitches the per-repo results into ralph/workspace-plan.md. When the workspace-mode PR in the ai-ralph fork you use lands, the adapter switches to the native workspace command with no skill-level changes.

Where Do Secrets Go?

In .mcp.json, which is gitignored. Use ${ENV_VAR} refs for every credential — never literal tokens. .mcp.json.template is committed and shows the agreed server set; each collaborator copies it to .mcp.json locally and fills in their own env vars out-of-band. gh auth status provides GitHub credentials; no token lives in the repo.

How Do I Contribute a Skill Improvement Back to the Template?

Open a pull request against amit-t/ai-workbench. The workbench’s update.wb is strictly one-way (template → instance) by design, so upstream contributions go via PR, not via sync. If you hand-edit a template_owned path in a stamped workbench, update.wb will overwrite it on the next run — which is why the skill constitution (AGENTS.md) tells agents not to touch those paths.

Can I Rename the Workbench Repo?

Yes, but project.conf’s WORKBENCH_REPO needs to match afterwards. There is no automatic rename support — rename on GitHub, update project.conf locally, push. Every collaborator re-pulls.

Is the .workbench-state/ Dir Shared?

Yes — it is tracked in git. Both collaborators see the same published.json, approved.json, rejected.json. This is deliberate: the approval state is a shared artifact that ralph reads, so both sides need to agree on it.

Why Do I See “[wb] checking for updates…” Every Time I Run a wb Command?

You don’t, actually. The first meaningful wb.* command in any 12-hour window triggers a single GitHub-API call to compare your stamped wb’s template version against upstream ai-workbench. If a newer template version is available, a one-line banner is printed before your command runs. Otherwise the preamble is silent. The result is cached for 12 hours (configurable via check_ttl_hours in upstream version.json), so subsequent calls within the window do nothing. The check is fail-open: offline machines, rate-limit hits, and missing gh all skip silently. Trivial list aliases (wb.published, wb.approved, wb.rejected) never trigger the check at all. See Versioning + upgrades for the full picture.