v3.4.0

Add offline moadim export / moadim import CLI commands to back up and restore the tracked config (routines, agents, notifications.toml, user_prompt.md) as a single portable JSON bundle, excluding gitignored runtime/local files. Import validates paths and routine.toml syntax before writing, writes atomically, skips existing files unless --force, and supports --dry-run.

Add moadim lock [--scope shared|local], moadim unlock [--scope shared|local|all] and moadim lock-status, so the CLI can pause and resume all routine scheduling the same way the REST API and MCP tools already can.

Avoid duplicate routine run-history entries when a live workbench already has a persisted cleanup record.

When a routine has several invalid env entries, the validation error now always reports the first invalid key in sorted order, instead of one that varied between runs.

Persist explicit scheduled-fire evidence before in-process scheduler launch attempts.

Fix flags being written to and read from a phantom title-slug directory for routines that live in a folder or have been renamed/moved, by keying all flag operations (create, list, resolve) and flag_count on the routine's actual on-disk relative directory instead of slugify(title). On first start, flags already stranded in such a title-slug directory are moved into the routine's real directory (a flag whose name already exists there is left in place), so none disappear.

Keep test routine storage under isolated temporary homes when a fixture omits its explicit home override.

Expose in-process routine-scheduler resync health, completion time, and rebuild failures after routine mutations.

Add a long-running-now panel and OVERRUNNING tile to the Reliability page: flags live runs whose elapsed time already exceeds 1.5× their routine's own p95 duration, worst overrun first with a link to the live run — catching a hung agent while it's still running.

Stop logging routine scheduler resync request failed: routine scheduler is not running on every daemon start. The routine scheduler now starts before the startup sync requests its first resync.

Retry the on_failure webhook on transient delivery failures. A failed curl POST is now tried up to 3 times with exponential backoff (100 ms, then 200 ms) before the failure is logged, so one dropped request no longer loses a routine's failure notification.