Driver

The framework has zero guidance on plan mode. A grep across .claude/, docs/, workflows/, CLAUDE.md returned only one mention — a spike doc referencing Anthropic's opusplan model alias. No rule, no skill, no hook tells the agent when to enter plan mode.

This was surfaced in a real session as a concrete gap: plan mode would have helped on a multi-step /handover discovery (5+ tool calls, unclear path) and on a /setup --enable-lsp debug loop (figuring out why a marketplace add was silently failing). The work got done but was less efficient than it could have been with explicit planning up front.

ApexYard already has precedent for this kind of agent-discipline rule: .claude/rules/parallel-work.md documents when to offer /fan-out. A plan-mode rule is the same shape — primarily self-discipline, no hook (plan mode is harness-owned and not visible to PreToolUse).

Scope

• New file: .claude/rules/plan-mode.md. One short page, structured exactly like .claude/rules/parallel-work.md for consistency:
  • When to enter plan mode (heuristic — multi-step coordination with ≥4 dependent steps, unclear path, hard-to-reverse action, validating a fan-out split before spawning parallel agents)
  • When NOT to enter (single read, single edit, well-scoped patch with obvious next-action, conversation already established the plan)
  • Self-check question for the agent before answering: "did I just write a multi-step plan in prose without entering plan mode? That's a missed opportunity"
  • Brief mention of opusplan model alias (Opus during plan, Sonnet during execution) as prior art for tier-aware plan mode
• One-line import added to CLAUDE.md: @.claude/rules/plan-mode.md so the rule loads into agent context on every session.
• No hook (plan mode is harness-owned; hooks can't see it). The rule explicitly notes this so future maintainers don't try to add one.

Acceptance Criteria

• .claude/rules/plan-mode.md exists with the four sections above (when-to-enter, when-NOT-to, self-check, prior art)
• Imported into CLAUDE.md via @.claude/rules/plan-mode.md
• Rule structure mirrors .claude/rules/parallel-work.md (same headings, same self-check pattern, same length)
• Rule explicitly notes plan-mode is harness-owned — no hook enforcement is possible, this is self-discipline only
• Manual smoke test: a fresh session reading CLAUDE.md picks up the new rule and the agent can articulate the trigger heuristic when asked "when should you use plan mode"

Risks / Dependencies

• Low risk — doc-only change.
• Plan-mode mechanics could change in a future Claude Code release; same risk as any other rule doc. The rule references the Claude Code plan-mode docs URL so a stale rule is recoverable from the link alone.