/fusion-plan

name: fusion-plan
description: >-
  Produce a high-confidence, CONCISE implementation / architecture / strategy PLAN that follows the OMC
  plan system end to end. It first runs OMC's real requirements interview (auto-chains the `omc-plan`
  skill: one question at a time, explore-first, Analyst consult) to produce an initial plan in
  `.omc/plans/`, then DEEPENS that plan with a 3-round ITERATIVE fusion loop — each round an Opus 4.8
  panelist (Agent subagent) and a GPT-5.5 panelist (via codex) critique/improve it IN PARALLEL and BLIND,
  then Opus 4.8 (the orchestrator) judges and synthesizes one tighter plan that seeds the next round,
  stopping early on convergence. Writes the final dense plan back to the same `.omc/plans/<slug>.md`, then
  offers the OMC handoff (`/omc-plan --review` → `/team` or `/ralph`). Use for high-stakes planning,
  architecture decisions, strategy calls, and hard debugging — especially when the user runs /plan or asks
  how to approach a non-trivial build. Reuses the `fusion` skill's run_codex.sh + judge rubric. NOT for
  trivial tasks (single pass is enough).

Fusion-Plan — OMC interview, then a 3-round iterative panel

This skill follows the **OMC plan system**: requirements are gathered by OMC's real interview FIRST, then the fusion panel deepens the resulting plan, and the output flows back into OMC's review → execution pipeline. The fusion panel replaces only the *plan-thinking* step; storage, format, the quality gate, and execution all stay on OMC's rails.

**Hard rule (same as fusion): Opus 4.8 always judges and synthesizes. You are the judge — stay separate from the panelists.** The Opus panelist is always a *spawned subagent*, never you.

Step 0 — Requirements (pick the mode FIRST)

**Decide interactive vs non-interactive before doing anything — this determines whether you may interview.**

• **Non-interactive** when ANY of these is true: invoked with `--no-interview`; running autonomously while

executing a larger plan; running inside a spawned sub-agent; or you otherwise cannot reach a human who will answer. **This is the default during plan execution.**

• **Interactive** only when a human is clearly present and able to answer right now — typically the user

typed `/fusion-plan ...` directly in the main session at the start of a planning session.

Non-interactive mode (autonomous / sub-agent / `--no-interview`)

**HARD RULE: do NOT interview. Do NOT call `AskUserQuestion`. Do NOT auto-chain the `omc-plan` interview.** There is no one to answer — blocking on a question would hang the run. Requirements were already gathered upstream, so derive `SEED_PLAN` from the context that already exists, in this order:

1. The task/story doc if one exists (e.g. `docs/stories/STORY-<id>.md`). 2. The larger plan that flagged this task (e.g. the relevant `.omc/plans/*.md` section). 3. The task description / prompt args you were given.

If that context is genuinely insufficient to plan, do NOT guess and do NOT ask — stop and return a short note listing exactly what's missing, so the human (or the orchestrator) resolves it. Then go to Step 1.

Interactive mode (human present)

Run OMC's real requirements interview by invoking the OMC plan skill on the user's request **verbatim**:

Skill("oh-my-claudecode:omc-plan")   # pass the user's request; do NOT pass --consensus

• Let OMC auto-detect **interview vs direct** mode (one question at a time via `AskUserQuestion`,

explore-first, Analyst consult). It writes an initial plan to **`.omc/plans/<slug>.md`**.

• Do **NOT** pass `--consensus` (redundant with the fusion panel) and do **NOT** let OMC hand off to

execution yet — stop at the plan.

• **Capture the plan file path** as `SEED_PLAN` (newest `.omc/plans/*.md` if not reported). It carries the

gathered requirements — they must survive into the final plan.

• If `omc-plan` is unavailable, fall back to a minimal inline interview (one question at a time).

Step 1 — Preconditions

This skill targets the `opus4.8-gpt5.5` panel. Check codex is present:

command -v codex && codex --version

• If `codex` is installed → run the full Opus 4.8 + GPT-5.5 loop below.
• If `codex` is **missing** → tell the user, then fall back to two independent Opus 4.8 panelists per round

(`opus4.8-4.8`) rather than failing. Note the downgrade and how to enable GPT-5.5 (`npm i -g @openai/codex` + `codex login`).

Read the shared references once (they live in the sibling skill):

cat ~/.claude/skills/fusion/references/panel.md
cat ~/.claude/skills/fusion/references/judge_rubric.md

Honor `panel.md`: every panelist gets the same input **verbatim**, no assigned "lenses" or personas.

Step 2 — The loop, SEEDED from the OMC plan (round R = 1, 2, 3)

The panel does NOT draft blind — round 1 is already seeded by `SEED_PLAN`, so the interview's requirements anchor everything.

2a. Build each panelist's prompt

• **Every round (1, 2, 3) — independent critique-and-improve:** the original request **verbatim**, plus

the current plan, plus: > Here is the current plan (round 1: the requirements + initial plan from a stakeholder interview; > later rounds: the synthesized plan from the previous round): > ``` > <round 1: SEED_PLAN | round R>1: PLAN v(R-1)> > ``` > You are one of several independent expert planners; you will not see the others' work. Research with > web + bash as needed, then return an IMPROVED, COMPLETE plan. **Preserve every requirement and > constraint already captured** — do not drop them. Fix wrong assumptions, fill gaps, add missing > risks/edge cases, sharpen sequencing and decisions, and CUT padding — the plan must get tighter and > denser, not longer. If you genuinely believe no material improvement is possible, say so explicitly on > the first line (`NO_MATERIAL_CHANGE`) and return the plan unchanged.

Within a round, keep panelists **isolated** — never paste one panelist's round-R output into