Weekly graph maintenance: backup → analyze → prune → report

[stevepridemore]

Summary

Adds a weekly scheduled task that runs every Sunday at 6 AM local time. The task does three things in sequence — backup, health analysis, and pruning — with safety gates between each.

This fills a real gap: the project has been running the nightly dream for weeks but no graph backups have actually been happening since 2026-05-06. A folder existed at `~/.claude/scheduled-tasks/weekly-graph-backup/` but was never registered as a routine in Claude Code's scheduled-tasks system, so it sat dormant. Same fate as `prompts/dream-weekly.md` (analysis prompt that's been in the repo since bootstrap, never scheduled).

This PR consolidates both into one weekly task with prune wrapped in.

What's in the PR

• `prompts/weekly-maintenance.md` (new) — source-controlled prompt with 11 steps:
  1. Backup (graph_export with `label: weekly-pre-maintenance`, retention 14)
  2. Graph health overview (graph_stats)
  3. Unresolved contradictions (graph_contradictions, surfaced for human review)
  4. Stale entities (graph_cypher; confidence < 0.2 + last_seen > 90 days)
  5. Orphaned nodes
  6. Weak clusters (max edge weight < 0.3)
  7. Potential duplicates (graph_merge_suggestions, surfaced for human review)
  8. Dream process health (manifest + audit log)
  9. Prune preview with sanity check (abort if >50 nodes flagged)
  10. Prune execute (only if preview is sane)
  11. Report — append summary to `weekly-maintenance.log` + write detailed markdown to `weekly-reports/YYYY-MM-DD.md`
• `scripts/sync-dream-skill.py` (extended) — refactored from one-task to N-tasks via a `TASKS` table. New `--task ` flag syncs only one; without flag, syncs all. Same CRLF→LF discipline preserved (SKILL.md parser breaks on CRLF).

Safety design

• Backup gate: if step 1 fails, every subsequent step is skipped. Prune does not run without a fresh backup as recovery option.
• Prune sanity check: if `graph_prune` preview shows more than 50 nodes to delete, the task aborts and logs a warning. A sudden large count usually signals a config change or bug, not legitimate cleanup.
• No auto-resolve, no auto-merge: contradictions and duplicate candidates are reported for human review. The task is decisive only about decay-driven pruning, which is the boring/safe case.

Schedule

Registered via the scheduled-tasks MCP (`create_scheduled_task`):

```
taskId: weekly-graph-maintenance
cronExpression: 0 6 * * 0 (Sunday 6:00 AM, local time)
nextRunAt: 2026-05-10T10:00:05Z (= Sunday 6:00 AM ET)
```

Schedule registration is local Claude Code state per machine — not in source control. The prompt body that runs IS in source control via this PR.

Cleanup

Deleted local stale `~/.claude/scheduled-tasks/weekly-graph-backup/` folder. The new task ID (`weekly-graph-maintenance`) replaces the never-registered `weekly-graph-backup` name and reflects the broader scope.

Test plan

• `python scripts/sync-dream-skill.py` syncs both nightly and weekly cleanly
• `mcp__scheduled-tasks__list_scheduled_tasks` confirms the new task is registered with the right cron
• First scheduled run on Sunday 6 AM — verify backup_file produced, prune executes (or correctly skips), and the report file lands at the expected path
• Verify `graph_prune` tool is pre-approved on the task so the run doesn't pause waiting for permission

🤖 Generated with Claude Code