GPT 5.4 Enhancement v3 — Phase 3.C

Tracking: #66345
Priority: P1 — The marquee Claude Code-quality feature (opt-in only)
Hard depends on: #61845 merging (plan-mode runtime from nick0809cn-ux)
Also depends on: #67514 (task-system parity), #67519 (plan rendering)

Context

PR #61845 already builds the plan-mode runtime core:

• SessionRuntimeMode = "plan" | "normal" | "auto"
• enter_plan_mode / exit_plan_mode agent tools
• plan-mode-hook.ts mutation gate (blocks 11+ tool categories)
• todo_write / task_create / task_update plan-mode tools
• Session-metadata persistence of plan state

This issue covers the layer above #61845: user-facing entry points and the approval UX that makes plan mode trustworthy. Without this, plan mode is two phases with no human review gate.

Scope

Entry points (all opt-in, never auto-enabled)

1. Slash command: /plan <optional goal> — one-shot opt-in for the next turn
2. Config flag: agents.defaults.planMode.enabled: false — per-agent default
3. Model gate: agents.defaults.planMode.autoEnableFor: [] — regex list (empty by default to preserve Hermes parity)
4. Control UI toggle: button + keyboard shortcut mirroring Claude Code's Shift+Tab and Codex's /plan

Approval state machine

exit_plan_mode called
  → plan_approval_requested event emitted
  → UI renders plan checklist + Approve / Edit / Reject buttons
  → On Approve: inject approved plan as system message, enter Phase B (normal mode)
  → On Edit: re-enter Phase A with user's edited plan
  → On Reject: abort turn
  → On timeout (configurable, default 10min): auto-reject with message

Channel-native approval UX

Tool-name reconciliation

#61845 introduces todo_write / task_create / task_update alongside the existing update_plan. This PR ships an adapter so both tool surfaces produce the same AgentPlanEventData events. Both names resolve to the same handler. Deprecation of one name deferred to Phase 4.

Strict-agentic interaction

• Phase A: planning-only retry guard is disabled (the whole point is to plan)
• Phase B: strict-agentic operates normally

Files to create/modify

• new src/commands/plan.commands.ts
• src/config/types.agent-defaults.ts, src/config/zod-schema.agent-defaults.ts
• new src/agents/plan-approval.ts + test
• src/infra/agent-events.ts (new event types)
• src/channels/plugins/approvals.ts (extend for plan approval)
• Control UI approval modal
• Per-channel approval handlers
• src/agents/pi-embedded-runner/run/incomplete-turn.ts (bypass guard in Phase A)

Estimated size: ~600 LoC

Open questions (resolve at implementation)

1. Does "ok do it" during Phase A count as approval → Phase B?
2. Should exec read-only whitelist be user-editable in config?
3. Should plan mode block sessions_spawn during Phase A?

Verification

• gpt54-plan-mode-approve.scenario.yaml — plan → approve → execute
• gpt54-plan-mode-reject.scenario.yaml — plan → reject → abort
• gpt54-default-no-plan-mode.scenario.yaml — default GPT-5 does NOT enter plan mode