fix: route vision debate subagents correctly

src/resources/extensions/sf/prompts/guided-plan-milestone.md

@@ -1,6 +1,6 @@
 Plan milestone {{milestoneId}} ("{{milestoneTitle}}"). Read `.sf/DECISIONS.md` if it exists — respect existing decisions. Read `.sf/REQUIREMENTS.md` if it exists and treat Active requirements as the capability contract. If `REQUIREMENTS.md` is missing, continue in legacy compatibility mode but explicitly note missing requirement coverage. Use the **Roadmap** output template below to shape the milestone planning payload you send to `sf_plan_milestone`. Call `sf_plan_milestone` to persist the milestone planning fields and render `{{milestoneId}}-ROADMAP.md` from DB state. Do **not** write `{{milestoneId}}-ROADMAP.md`, `ROADMAP.md`, or other planning artifacts manually. If planning produces structural decisions, append them to `.sf/DECISIONS.md`. {{skillActivation}} Fill the Horizontal Checklist section with cross-cutting concerns considered during planning (requirements re-read, decisions re-evaluated, graceful shutdown, revenue paths, auth boundary, shared resources, reconnection). Omit for trivial milestones.
 
-Before calling `sf_plan_milestone`, run a bounded **Vision Alignment Meeting** for the milestone and roadmap as a real multi-agent review. Use the `subagent` tool in `mode: "debate"` with `rounds: 2` and a separate task for each participant lens below. Do **not** merely simulate every participant inside this planner response. If the `subagent` tool is unavailable or fails after one retry, record that explicitly in `trigger` and run the structured meeting inline as a degraded fallback. This is allowed to be broader and more nuanced than slice planning. Include at least these participant lenses:
+Before calling `sf_plan_milestone`, run a bounded **Vision Alignment Meeting** for the milestone and roadmap as a real multi-agent review. Use the `subagent` tool in `mode: "debate"` with `rounds: 2` and a separate task for each participant lens below. Do **not** merely simulate every participant inside this planner response. Use only supported agent names: `planner`, `reviewer`, `researcher`, and `scout`. Put the stakeholder role name inside the task text; do not invent agent names such as `combatant`, `delivery-lead`, `product-manager`, or `customer-panel`. If the `subagent` tool is unavailable or fails after one retry, record that explicitly in `trigger` and run the structured meeting inline as a degraded fallback. This is allowed to be broader and more nuanced than slice planning. Include at least these participant lenses:
 - Product Manager
 - User Advocate
 - Customer Panel

src/resources/extensions/sf/prompts/plan-milestone.md

@@ -43,7 +43,7 @@ If milestone research exists (inlined above), trust those findings and skip redu
 
 Narrate your decomposition reasoning — why you're grouping work this way, what risks are driving the order, what verification strategy you're choosing and why. Use complete sentences rather than planner shorthand or fragmentary notes.
 
-Before you persist the roadmap, run a bounded **Vision Alignment Meeting** as a real multi-agent review. Use the `subagent` tool in `mode: "debate"` with `rounds: 2` and a separate task for each participant lens below. Do **not** merely simulate every participant inside this planner response. If the `subagent` tool is unavailable or fails after one retry, record that explicitly in `trigger` and run the structured meeting inline as a degraded fallback. This is broader than slice planning and should feel allowed to be chatty and nuanced. Gather the strongest additions, cuts, and ordering changes from these participant lenses:
+Before you persist the roadmap, run a bounded **Vision Alignment Meeting** as a real multi-agent review. Use the `subagent` tool in `mode: "debate"` with `rounds: 2` and a separate task for each participant lens below. Do **not** merely simulate every participant inside this planner response. Use only supported agent names: `planner`, `reviewer`, `researcher`, and `scout`. Put the stakeholder role name inside the task text; do not invent agent names such as `combatant`, `delivery-lead`, `product-manager`, or `customer-panel`. If the `subagent` tool is unavailable or fails after one retry, record that explicitly in `trigger` and run the structured meeting inline as a degraded fallback. This is broader than slice planning and should feel allowed to be chatty and nuanced. Gather the strongest additions, cuts, and ordering changes from these participant lenses:
 - **Product Manager:** what is the real product move and what should the roadmap prove?
 - **User Advocate:** what must matter for the user experience and trust surface?
 - **Customer Panel:** multiple likely customer viewpoints, not a single flattened “user”.

src/resources/extensions/subagent/index.ts

@@ -50,6 +50,15 @@ const AGENT_ALIASES: Record<string, string> = {
 	default: "worker",
 	code: "reviewer",
 	coder: "typescript-pro",
+	"product-manager": "planner",
+	"user-advocate": "reviewer",
+	"customer-panel": "reviewer",
+	business: "planner",
+	"delivery-lead": "planner",
+	partner: "reviewer",
+	combatant: "reviewer",
+	architect: "planner",
+	moderator: "planner",
 	["g" + "sd-executor"]: "worker",
 	"sf-worker": "worker",
 	"sf-scout": "scout",
@@ -918,7 +927,13 @@ function buildSubagentProcessArgs(
 	tmpPromptPath: string | null,
 	modelOverride?: string,
 ): string[] {
-	const args: string[] = ["--mode", "json", "-p", "--no-session"];
+	const args: string[] = [
+		...getBundledExtensionCliArgs(),
+		"--mode",
+		"json",
+		"-p",
+		"--no-session",
+	];
 	const modelToUse = modelOverride ?? agent.model;
 	if (modelToUse) args.push("--model", modelToUse);
 	if (agent.tools && agent.tools.length > 0)
@@ -928,6 +943,14 @@ function buildSubagentProcessArgs(
 	return args;
 }
 
+function getBundledExtensionCliArgs(): string[] {
+	return (process.env.SF_BUNDLED_EXTENSION_PATHS ?? "")
+		.split(path.delimiter)
+		.map((s) => s.trim())
+		.filter(Boolean)
+		.flatMap((p) => ["--extension", p]);
+}
+
 interface SubagentLaunchSpec {
 	command: string;
 	args: string[];
@@ -1204,14 +1227,9 @@ async function runSingleAgent(
 		let wasAborted = false;
 
 		const exitCode = await new Promise<number>((resolve) => {
-			const bundledPaths = (process.env.SF_BUNDLED_EXTENSION_PATHS ?? "")
-				.split(path.delimiter)
-				.map((s) => s.trim())
-				.filter(Boolean);
-			const extensionArgs = bundledPaths.flatMap((p) => ["--extension", p]);
 			const proc = spawn(
 				launchSpec.command,
-				[...extensionArgs, ...launchSpec.args],
+				launchSpec.args,
 				{
 					cwd: cwd ?? defaultCwd,
 					env: launchSpec.env,
@@ -1388,15 +1406,9 @@ async function runSingleAgentInCmuxSplit(
 			);
 		}
 
-		const bundledPaths = (process.env.SF_BUNDLED_EXTENSION_PATHS ?? "")
-			.split(path.delimiter)
-			.map((s) => s.trim())
-			.filter(Boolean);
-		const extensionArgs = bundledPaths.flatMap((p) => ["--extension", p]);
-		const launchSpec = resolveSubagentLaunchSpec([
-			...extensionArgs,
-			...buildSubagentProcessArgs(agent, task, tmpPromptPath, modelOverride),
-		]);
+		const launchSpec = resolveSubagentLaunchSpec(
+			buildSubagentProcessArgs(agent, task, tmpPromptPath, modelOverride),
+		);
 		const launcherPath = writeNodeSubagentLauncher(
 			launchSpec,
 			cwd ?? defaultCwd,