haowjy/meridian-work-coordination

Work Coordination

The orchestrator owns work state — subagents should not mutate it unless explicitly instructed, because concurrent mutations from multiple spawns create race conditions and inconsistent status.

If meaningful repo work is about to start, create or attach to a work item:

meridian work start "descriptive name"   # create new
meridian work switch descriptive-name    # attach to existing

Dashboard

meridian work                    # dashboard — what's in flight
meridian work list               # list active work items
meridian work list --done        # list done/archived items
meridian work show auth-refactor # drill into one work item

Status Management

Status values are free-form. Keep the current phase visible:

meridian work update auth-refactor --status designing
meridian work update auth-refactor --status implementing
meridian work done auth-refactor
meridian work reopen auth-refactor
meridian work delete stale-item          # remove empty work items
meridian work delete old-item --force    # remove even if it has artifacts

work done archives the work directory. work reopen restores it. work delete removes the work item entirely — requires --force if it has artifacts.

Artifact Placement

$MERIDIAN_WORK_DIR — scoped to the current work item. Archived when the work completes.

$MERIDIAN_FS_DIR — long-lived reference material. Persists across work items.

Rule of thumb: if it helps this work item, use $MERIDIAN_WORK_DIR. If it helps any task understand the project, use $MERIDIAN_FS_DIR.

Commit Work Artifacts

Agent sessions are ephemeral — compaction, crashes, and context limits erase conversation state. Work artifacts in $MERIDIAN_WORK_DIR and $MERIDIAN_FS_DIR are the only thing that survives. Commit them to git after creating or updating them so a future agent can resume from artifacts alone. Don't batch until the end — commit as you go so progress is never lost to a mid-session failure.