GPT 5.4 Enhancement — Phase 4.1

Tracking: #66345
Priority: P2 — Power-user differentiator (not parity-critical)
Depends on: #67514 (task-system parity), #67520 (plan mode)
Confidence: 75% (novel feature — not in Hermes or Claude Code)

Concept

A skill (.openclaw/skills/<name>/SKILL.md) can declare a plan-template block in its YAML frontmatter. When the agent invokes that skill, the runtime instantiates the template into update_plan automatically, giving the agent a starting checklist instead of deriving one from scratch.

Use case

Users with repeatable workflows (release cuts, deployment runs, incident response, code review flows) want the agent to start with a known plan structure every time. Today GPT-5 has to re-derive the plan from scratch each invocation, which is token-wasteful and inconsistent.

Example

# .openclaw/skills/deploy/SKILL.md frontmatter
---
name: deploy
plan-template:
  - step: "Run test suite"
    activeForm: "Running test suite"
  - step: "Build production artifacts"
    activeForm: "Building production artifacts"
  - step: "Deploy to staging"
    activeForm: "Deploying to staging"
  - step: "Run smoke tests on staging"
    activeForm: "Running smoke tests"
  - step: "Promote to production"
    activeForm: "Promoting to production"
---

When the user says "deploy the app" and the deploy skill activates, the runtime calls update_plan with the template steps (all pending) before the first agent turn, so the agent starts executing step 1 immediately.

Files to create

• new src/agents/skill-planner.ts — reads plan-template from skill frontmatter, instantiates into update_plan
• src/plugins/skill-state.ts — hook to call skill-planner on skill activation
• Tests

Estimated size: ~400 LoC

Why Phase 4

This is not in Hermes and not in Claude Code. It's a genuine differentiator but lower priority than the parity items. Ship after Phase 3 lands and we have baseline parity numbers.