feat(scheduler): sweepPhantomBusy — clear stuck active_tasks from crashed runs

[HongmingWang-Rabbit]

Summary

Extracts the sweepPhantomBusy 5-minute sweeper from closed PR #1664 to address monorepo issue #1372.

Root cause

When a workspace agent crashes mid-task (MiniMax timeout, OOM, panic in the finally block, SSRF kill, etc.), the active_tasks counter gets incremented on task start but never decremented. From then on the scheduler treats the workspace as perpetually busy and skips every cron fire with workspace busy — retry. The only fix was manual SQL against prod.

Fix

A periodic sweeper on the scheduler tick loop:

• Runs at most every 5 minutes (phantomSweepInterval), gated via lastSweepAt on the Scheduler struct — cheap no-op on non-sweep ticks.
• Finds workspaces where active_tasks > 0 AND status is not removed AND there has been no activity_logs row in the last 10 minutes (phantomStaleThreshold).
• UPDATE ... SET active_tasks = 0, current_task = '', updated_at = now() RETURNING id, name — logs each reset by name plus a summary count.

Safeguards

• Only sweeps when there is no activity log in the configurable stale window (default 10 min). A live agent logs at least a heartbeat/tool-call within that window, so genuinely busy workspaces are never touched.
• Excludes status = 'removed' workspaces so soft-deleted rows don't churn.
• Both thresholds are package-level constants (phantomSweepInterval, phantomStaleThreshold) — trivial to tune.
• Reuses the existing scheduler goroutine + mutex — no new goroutine, no new timer, no new panic surface.

Scope

Only workspace-server/internal/scheduler/scheduler.go changes. Extracted verbatim from the PR #1664 diff (the target section, not the cosmetic formatting noise).

Test plan

• go build ./... in workspace-server — clean
• go test ./internal/scheduler/... -count=1 — pass
• Deploy to staging and watch logs for Scheduler: phantom-busy sweep complete lines
• Verify that workspaces in the known "stuck" state clear to active_tasks = 0 within 5 min
• Verify that actively-working workspaces are not swept (their activity_logs are fresh)
• Confirm the existing 30-min manual SQL cron can be retired

Closes #1372.

🤖 Generated with Claude Code