railyardMake a delivery
RailyardTroubleshooting

Troubleshooting

Recover from the first broken proof, not the loudest symptom. Keep the receipt, run log, and exact failing command; identify the owner of that boundary; restore its prerequisite or evidence; then rerun only the check the fix invalidated.

The practice is a short loop: observe the held or failed receipt, trace it to install, delivery, transport, trust, or fleet ownership, repair that boundary, and require a fresh receipt. The mechanisms below are evidence for that loop.

Install failures

  • Marketplace add fails: check authentication, network access, and the exact novotnyllc/marketplace slug. Retry the Claude Code or Codex marketplace command from Install.
  • The plugin is already installed: confirm the installed entry points at the expected marketplace, then update it through the same harness instead of creating a second copy.
  • The CLI is too old: update Claude Code or Codex until its plugin marketplace command is available. The Codex equivalent is codex plugin marketplace add.
  • A plugin is present but a skill is missing: run the harness plugin listing, restart the harness, and verify the plugin source before retrying the request.

First-delivery failures

  • The skill cannot be found: finish Install and confirm railyard is listed before asking for a delivery.
  • There is no GitHub remote: verify origin before requesting a pull request. A local delivery can still run checks and report the missing publish step.
  • The repository has no test suite: let the route choose the smallest existing verification command and record that no focused check is available when none exists.
  • The review gate is stuck: inspect delivery gates for unresolved review threads, settlement timing, or a missing post-merge proof.

Fleet and Roundhouse failures

  • The agent cannot reach a host: run SSH Doctor and repair the first failing transport layer before changing enrollment.
  • Enrollment expired or was rejected: inspect enrollment and TOFU, keep the old authority in place, and repeat the ceremony after its prerequisites are current.
  • The sponsor channel is unreachable: restore the already-trusted channel. Do not substitute an unverified first contact.
  • Key generation failed: generate the key on the target host, keep the private key there, and restart possession proof. The enrollment recovery guide names the recovery path.

Where to look

The run log records the route, dispatches, checks, review rounds, retries, and terminal result. The delivery receipt identifies the merge and post-merge proof. For diagnosis, start with Doctor and SSH Doctor; for a fleet-wide readiness question, use Fleet readiness.