fix(frontend): honest effort display in agent UI (gemini review)

Address two UI-accuracy issues found in review:
- AgentBuilder: the per-model effort select no longer disguises an unset
  value as 'High'. Adds an explicit 'Inherit' option; choosing it strips
  the effort key so the saved TOML omits it (and the call site falls back
  to per-tab → default), matching displayed intent to persisted state.
- ModelSelector: effort badges for models without an explicit override now
  reflect the actual effective effort (per-tab selector → default) instead
  of always showing the default constant, mirroring backend resolution.

2 files changed, 23 insertions, 8 deletions

diff --git a/packages/frontend/src/lib/components/AgentBuilder.svelte b/packages/frontend/src/lib/components/AgentBuilder.svelte
index bbdb83c..f1c9cdf 100644
--- a/packages/frontend/src/lib/components/AgentBuilder.svelte
+++ b/packages/frontend/src/lib/components/AgentBuilder.svelte
@@ -180,7 +180,17 @@ const modelCache = new Map();
 	}
 
 	function setEffortEntry(i: number, effort: string) {
-		formModels = formModels.map((m, idx) => (idx === i ? { ...m, effort } : m));
+		formModels = formModels.map((m, idx) => {
+			if (idx !== i) return m;
+			// Empty string = "inherit" (no per-model override). Strip the key so
+			// the saved TOML omits `effort` and the call site falls back to the
+			// per-tab selector / default.
+			if (!effort) {
+				const { effort: _dropped, ...rest } = m;
+				return rest;
+			}
+			return { ...m, effort };
+		});
 	}
 
 	async function openKeyModal(i: number) {
@@ -557,11 +567,12 @@ const modelCache = new Map();
 									{entry.model_id || "Select Model"}
 								</button>
 								<select
-									class="select select-bordered select-sm shrink-0 w-28"
-									title="Reasoning effort for this model"
-									value={entry.effort ?? DEFAULT_REASONING_EFFORT}
+									class="select select-bordered select-sm shrink-0 w-36"
+									title="Reasoning effort for this model. 'Inherit' uses the per-tab selector / default."
+									value={entry.effort ?? ""}
 									onchange={(e) => setEffortEntry(i, e.currentTarget.value)}
 								>
+									<option value="">Inherit ({REASONING_EFFORT_LABELS[DEFAULT_REASONING_EFFORT]})</option>
 									{#each REASONING_EFFORTS as effort}
 										<option value={effort}>{REASONING_EFFORT_LABELS[effort]}</option>
 									{/each}
diff --git a/packages/frontend/src/lib/components/ModelSelector.svelte b/packages/frontend/src/lib/components/ModelSelector.svelte
index c328511..8601795 100644
--- a/packages/frontend/src/lib/components/ModelSelector.svelte
+++ b/packages/frontend/src/lib/components/ModelSelector.svelte
@@ -37,12 +37,16 @@ const modelCache = new Map<string, string[]>();
 	}
 
 	/**
-	 * Human-readable effort label for a (possibly-unset / arbitrary) effort
-	 * string. Falls back to the default level's label when unset/invalid so the
-	 * displayed badge always reflects what will actually be used.
+	 * Human-readable effort label for a (possibly-unset) per-model effort. When
+	 * the model has no explicit override, the badge reflects what will ACTUALLY
+	 * run: the per-tab selector if valid, else the system default. This mirrors
+	 * the backend resolution order (per-model → per-tab → default) so the UI
+	 * never misrepresents the effective effort.
 	 */
 	function effortLabel(effort: string | undefined): string {
-		return REASONING_EFFORT_LABELS[isReasoningEffort(effort) ? effort : DEFAULT_REASONING_EFFORT];
+		if (isReasoningEffort(effort)) return REASONING_EFFORT_LABELS[effort];
+		const tab = isReasoningEffort(reasoningEffort) ? reasoningEffort : DEFAULT_REASONING_EFFORT;
+		return REASONING_EFFORT_LABELS[tab];
 	}
 
 	const {