Plan

Build or revise one living Epic roadmap. The Epic body owns conceptual depth; executable Tasks/Bugs remain direct children. End at the roadmap and /sy:spec handoffs. Never implement or spec leaves.

Plan against fresh origin/main unless the user names another base. Work code read-only. Tracker writes use the tracker skill (/sy:tracker).

Before anything else — before this turn spends any research — run the tracker preflight (${CLAUDE_PLUGIN_ROOT}/skills/shared/references/preflight.md). A failure stops here with its single ## Action needed block, not partway through the roadmap.

$ARGUMENTS

Invariants

• Work backwards: North Star → capabilities → dependency order → near executable leaves.
• Premise + prior-work check comes first: before shaping, validate the objective's premise and search for existing, shipped, duplicate, or sibling work (tracker find-issues plus a code/PR search); prior delivery or duplication reshapes or ends the roadmap rather than re-planning it.
• At most the resolved plan.max_active_tasks cap of leaves may be active /sy:spec-ready/in-spec/in-ship/in-review at once — resolve per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/config-values.md.
• One Task/Bug ≈ one coherent PR. Keep far work conceptual until evidence justifies decomposition.
• Not every issue surfaced mid-ship earns its own leaf: a small, adjacent, low-risk fix folds into the current PR as a recorded scope extension, and a follow-up must justify itself against that (see ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/scope-discipline.md).
• Objective is stable; path is provisional and should adapt to shipped evidence.
• Flat tracker execution, fractal conceptual map.
• Ask one question at a time, via AskUserQuestion, only when the answer changes seam, sequence, blocker, or outcome — see ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/user-interaction.md. At an approval point that authorizes tracker writes, name the mutations the go-ahead covers — create/edit the Epic, create/edit its children, and post the plan checkpoint — so auto-mode consent is informed rather than a bare "proceed".

Scope and delegation

Read small cohesive evidence directly. Use sy:sweep for large code/ticket/PR/docs surfaces and sy:seam only for one unresolved boundary that changes roadmap shape. At most the resolved limits.max_depth_agents cap in flight — resolve per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/config-values.md. Agent returns are compressed leads; verify decisive spans and own the cut.

Read durable cross-session memory early — memory_list (or memory_search on the tools/surfaces the objective touches) per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/memory.md; a lesson that bears on the objective enters the brief as a known anchor, and a lesson this objective's own research directly contradicts is refuted immediately (memory_refute, same reference) rather than carried forward or silently dropped.

Seed every agent prompt with known anchors — paths, symbols, entry points, keys — and name ground already covered; agents must not rediscover what the caller knows.

Machine-facing agent briefs stay pointer-dense. Human-facing Epic maps and decision logs remain clear prose. Both, and the roadmap and # SEAMS comments this skill posts, are written under ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/context-economy.md.

State router

Classify before loading detailed procedure:

NEW       objective, no Epic yet        → references/new-objective.md
REENTRY   existing Epic                 → references/reentry.md
SHAPING   evidence gathered             → references/roadmap-shaping.md
CAPTURE   roadmap decisions settled     → references/checkpoint-handoff.md

Load only the reference for the current state. Do not preload mutually exclusive procedures.

Completion bar

The Epic body must show North Star, conceptual horizon ladder, completed branches, current active set (≤ the resolved plan.max_active_tasks cap) with keys/kickoffs, queued conceptual work, critical path, blockers, and parallel-safe set. Every planning run that changes the tracker adds one decision-log delta ending in a Plan checkpoint footer. A roadmap entry or checkpoint that shipped evidence later overrules is corrected on its own surface, not left stale — the retroactive-honesty invariant in ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/write-integrity.md. When transcript.attach resolves true, render and attach this session's transcript to the Epic ($KIND=plan) per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/transcript-attach.md.

When every horizon is delivered and every child is done for delivered reasons, set the Epic done.

Every subagent dispatch resolves its model from config and passes it as the Agent invocation's actual model override, per ${CLAUDE_PLUGIN_ROOT}/skills/shared/references/model-dispatch.md; a nested dispatch inherits nothing and must resolve again.