Implement two-tier task manager escalation watchdog and optional manager

Problem

Running a permanent Claude “manager” in tmux is optional for most work: heartbeat already dispatches ready tasks and writes snapshots. Operators still want clarity on when a human or high-context agent should intervene. Today there is no first-class escalation path from dumb automation to manager.

Context

• Tier 1 (default): scripts/task-manager-heartbeat, task-manager-run, scripts/dispatch, tasks/logs/task_manager/*.snapshot.json, register JSON, dispatch logs under tasks/logs/dispatch-*.log.
• Tier 2 (escalation): Optional tmux manager (just manage / Jared) or a one-shot claude session with a bundled context pack, only when signals fire.
• Related exploration: upgrade-jared-task-manager-visibility-orchestration-and-operator-ux.

Possible Solutions

• A — Signal-only (docs + snapshot fields): Define escalation signals and surface them in snapshot.json / heartbeat text; operator spawns manager manually. Lowest code, less automation.
• B — Watchdog flags + notification: Tier 1 sets needs_escalation (or extends needs_attention) with reasons; optional desktop notification or log line; still manual manager.
• C — Automated manager spawn (opt-in): On critical signals, run a configured command (e.g. open tmux window or claude with prompt file). Recommended direction only if guarded (rate limit, dry-run mode) to avoid spawn storms.

Tag B as the pragmatic first pass unless product explicitly wants C.

Plan

• [x] Document canonical escalation signals (e.g. repeated dispatch non-zero for same slug, in_progress stale by task mtime / started_at, dispatch log idle, register/worktree mismatch).
• [x] Implement detection in task_manager_lib.py / task-manager-heartbeat (or small helper module); append structured reasons to snapshot payload and text summary.
• [x] Make optional tmux manager start path clearly non-default or triggered only when TASK_MANAGER_ESCALATION=1 / config flag (exact mechanism TBD).
• [x] Add tests under tests/scripts/test_task_manager_scripts.py for at least one escalation signal.
• [x] Update .codex/skills/task-manager/SKILL.md (or AGENTS) to describe two-tier operator model.

Implementation Progress

• Added structured escalation signaling in scripts/task_manager_lib.py:
• per-task escalation_reasons[] and needs_attention[] in snapshot tasks
• top-level escalation.required, escalation.signals[], counts.escalation_signals
• text heartbeat now includes an Escalation signals section when active
• Implemented minimal watchdog detections for first pass:
• dispatch_failed / dispatch_interrupted / dispatch_stalled
• stuck_in_progress (task is in_progress and stale by idle threshold)
• retained worker_idle / pickup_overdue as structured reason codes
• Added merge-conflict prevention for started_at churn:
• scripts/task-manager-heartbeat --dispatch-ready no longer edits task frontmatter by default
• optional opt-in remains via --mark-started-on-dispatch or TASK_MANAGER_MARK_STARTED_ON_DISPATCH=1
• Added/updated tests in tests/scripts/test_task_manager_scripts.py:
• structured escalation snapshot + text assertions
• default dispatch no longer mutates task status/frontmatter
• explicit opt-in path still marks task started
• Follow-on merge context:
• PR #705 conflict resolution kept this task's escalation schema and --mark-started-on-dispatch opt-in semantics while adding PR-CI probe + .worktrees defaults.
• Task-manager heartbeat now reflects both layers on main.

QA Exploration

• N/A (non-UI infra/scripts task)
• [x] QA exploration completed (or N/A for non-UI tasks)

Review Feedback

• Review rounds completed via just review; follow-up fixes applied before merge.
• [x] Review cleared