GPT 5.4 Enhancement — Phase 4.2

Tracking: #66345
Priority: P2 — Power-user feature for multi-agent workflows
Depends on: #67514 (task-system parity), #67520 (plan mode)
Confidence: 70% (complex; Claude Code Tasks API does this but Hermes does not)

Concept

Multiple sessions / multiple agents can attach to a shared plan list and coordinate work. One agent kicks off a plan, another picks it up or works on different steps in parallel.

Modeled after Claude Code's Tasks API with CLAUDE_CODE_TASK_LIST_ID env var for cross-session visibility.

Design sketch

• src/agents/plan-store.ts extended with a shared plan namespace (env var or config key)
• Plan state persisted to ~/.openclaw/plans/<namespace>/plan.json
• File-level locking (or SQLite) for concurrent access
• New plan_changed event type so sessions can subscribe to plan updates from other sessions
• UI: Control UI shows shared plan with ownership annotations ("step 3: assigned to session-abc")

Why Phase 4

This is the highest-risk, narrowest-audience feature. It requires concurrency primitives, conflict resolution, and schema migration. Defer until plan mode ships and real multi-agent demand emerges.

Estimated size: ~500 LoC + significant test surface for concurrency

Open questions

1. File-based locking vs SQLite for concurrent plan access?
2. How do conflicting updates resolve? (last-write-wins vs merge)
3. Should cross-session plans survive process restart? (yes, via disk persistence)
4. Should there be an ownership concept per step? (suggested: optional)