design-upgrade

Analyze the team composition proposed in the preceding /design Step 2 output and produce a second opinion along two axes — model and role — recommending only reinforcing changes.

Read ${CLAUDE_SKILL_DIR}/../_common/team-upgrade-analysis.md for the axis-agnostic engine: the two-axis Scope, restraint principle, role-gap detection + HARD LIMIT, the per-OPERATION gates (restraint / split / remove / merge), OPERATION output format and integrity rules, the per-existing-role invariant, the degraded-axis handling, and the Cross-axis synthesis. Resolve every {PLACEHOLDER} in that core against the ## Bindings below. The Precondition, design-lite conflict guard, schema-drift sync-note, and 3-path Fallback that follow are this skill's injected prose.

Bindings

Placeholder	Value
{SCOPE_LABEL}	탐색 범위
{FIELD_ORDER}	역할 / 탐색 범위 / 모델
{SOURCE_STEP}	/design Step 2
{ROSTER_SOT}	design/SKILL.md Step 2 (English prose: role / exploration scope / model, plus the model-alias set)
{LABEL_RELATION}	conceptual translation (the Korean OPERATION labels conceptually correspond to Step 2's English-prose fields)
{ROLE_GAP_ANCHOR}	(empty — no skill-specific gap anchor)
{RE_EXPLORATION_ANCHOR}	the interview / task narrative
{LITE_GUARD_TARGET}	design-lite
{PRECONDITION}	the preceding /design Step 2 team composition is in context
{MODEL_AXIS_DEGRADES_UNDER}	Fallback Path 2
{SAVED_DOC_LOCATION}	docs/<slug>.md
{REFEED_TARGET}	/design Step 2 re-proposal (same target on every path)

This skill declares no ## Operations Layer — it uses the core's OPERATION set and forbidden-set unchanged, whatever that set currently contains. The set is deliberately not re-enumerated here: a copy of it would go stale the next time the core gains or loses an operation.

Precondition

This skill requires the preceding /design Step 2 team composition to be in context. It is a disable-model-invocation: true second-opinion skill and does NOT auto-chain /design.

design-lite conflict guard: if the most recent proposal in context was made by /design-lite — judged by the command that produced it, not by its model mix, because /design now proposes sonnet by default and its smallest roster (one domain seat plus the verification seat) has the same two-sonnet shape — both axes conflict with the lite contract — UPGRADE→opus violates the sonnet pin, and every role-axis OPERATION mutates a roster the lite skill fixes by design (in either direction: the fixed size is the contract, so shrinking it breaks the contract exactly as growing it does). Emit the caveat and proceed only on explicit user confirmation. (Consistent with the cross-reference in design-lite/SKILL.md.)

Schema-drift sync note: the source of truth is design/SKILL.md Step 2's team-composition field set (role / exploration scope / model, plus the model-alias set — English prose). If Step 2 changes that field set, manually sync this skill's OPERATION Korean labels and model aliases (the Korean labels are this skill's emit surface, conceptually corresponding to Step 2's English fields). Do NOT touch design/SKILL.md. The residual risk of a missed sync is accepted without a lint.

Fallback

When the precondition is not met, follow a 3-path fallback:

• Path 1 — session roster paste: the user pastes the Step 2 composition → full two-axis analysis.
• Path 2 — reverse-engineer from a saved design doc (degraded — role axis only): docs/<slug>.md captures architecture/scope but does NOT save the Step 2 roster (Step 4 saves only architecture / decisions / issues / order). Role-gap detection is possible from the doc's scope, but the model axis is not (no current model assignments). State this limitation when taking this path. Because the model axis is impossible at the source, the core's degraded-axis handling fires here ({MODEL_AXIS_DEGRADES_UNDER} = Fallback Path 2): the split gate's cross-axis reconciliation and the Cross-axis synthesis treat the UPGRADE comparison term as N/A and proceed on the role axis alone (skip the SPLIT-vs-UPGRADE tradeoff weighing). Also, since no interview is in session, the lightweight re-exploration anchor is the saved doc's architecture/scope narrative instead of the interview — limit to surfaces the doc cites/names and their adjacent siblings under the same HARD LIMIT. REMOVE and MERGE are suppressed on this path for the same reason the model axis is: both take an existing roster role as their lookup key, and the saved doc holds no roster at all, so there is nothing to name. A gap-driven ADD needs no such key and survives; when an ADD on this path would exceed the team-size ceiling, the pairing it owes cannot be constructed here — say so and route the user back to a live Step 2 rather than emitting an unpaired over-budget ADD.
• Path 3 — start a new /design: when neither is available, two-axis analysis is impossible — emit a Korean notice (e.g., "분석할 팀 구성이 없습니다. 먼저 /cc-cmds:design을 실행해 팀 구성을 받은 뒤 다시 호출하세요.") and stop.