Merge PR #7314: fix(agent): skip prewalk switch on read-only xd:// device calls (@roboomp)

packages/coding-agent/CHANGELOG.md

@@ -139,6 +139,10 @@
 
 - Fixed image paste failing on Wayland-only Linux sessions by reading PNG clipboard payloads through `wl-paste` before falling back to the native bridge ([#7316](https://github.com/can1357/oh-my-pi/issues/7316)).
 
+### Fixed
+
+- Fixed prewalk switching to the fast model during read-only investigation: `xd://` devices are dispatched through the `write` tool, so a read-only call such as an `lsp` navigation counted as the first edit/write and armed the one-way hand-off mid-planning. Device dispatches now carry the wrapped tool's approval tier and only trigger the switch at a `write`/`exec` tier — read-only `lsp`, `debug` inspection, and internal-URL `ast_edit` calls no longer downgrade the model ([#7312](https://github.com/can1357/oh-my-pi/issues/7312)).
+
 ## [17.2.3] - 2026-08-01
 
 ### Changed

packages/coding-agent/src/session/prewalk.ts

@@ -1,6 +1,6 @@
 import type { Agent, AgentMessage, AgentToolResult, AgentTurnEndContext } from "@oh-my-pi/pi-agent-core";
 import { invalidateMessageCache } from "@oh-my-pi/pi-agent-core/compaction";
-import type { Model } from "@oh-my-pi/pi-ai";
+import type { Model, ToolResultMessage } from "@oh-my-pi/pi-ai";
 import { prompt } from "@oh-my-pi/pi-utils";
 import type { LocalProtocolOptions } from "../internal-urls";
 import { resolveApprovedPlan } from "../plan-mode/approved-plan";
@@ -25,6 +25,28 @@ const PREWALK_ACTION_TOOLS: Record<string, true> = {
 };
 const PLAN_YOLO_HANDOFF_MESSAGE_TYPE = "plan-yolo-handoff";
 
+/**
+ * Whether a completed tool result is the first workspace-mutating action that
+ * arms the prewalk hand-off. A direct `edit`/`write` call always counts; a
+ * `write` that dispatched an `xd://` device (e.g. `lsp`, `ast_edit`, `debug`)
+ * counts only when the wrapped tool resolved to a `write`/`exec` approval tier.
+ * Read-only device calls — LSP navigation, `debug` inspection, `ast_edit` on
+ * internal URLs, help lookups — leave the tier `read` (or absent) and must not
+ * switch the model mid-investigation (issue #7312).
+ */
+function isPrewalkImplementationAction(result: ToolResultMessage): boolean {
+	if (!PREWALK_ACTION_TOOLS[result.toolName]) return false;
+	const details = result.details;
+	// A direct filesystem edit/write carries no `xd://` dispatch metadata.
+	if (!details || typeof details !== "object" || !("xdev" in details) || !details.xdev) return true;
+	const xdev = details.xdev;
+	// Device dispatch: switch only on a genuine mutation tier. An absent tier
+	// (help lookup, unresolved approval) declines the switch, matching the
+	// reporter's "stay on the large model a couple turns longer" preference.
+	if (typeof xdev !== "object" || !("tier" in xdev)) return false;
+	return xdev.tier === "write" || xdev.tier === "exec";
+}
+
 /** Capabilities the prewalk coordinator borrows from its owning session. */
 export interface PrewalkCoordinatorHost {
 	agent: Agent;
@@ -100,7 +122,7 @@ export class PrewalkCoordinator {
 
 		const todoGateOpen = this.#todoSeen || !this.#host.getActiveToolNames().includes("todo");
 		const action = todoGateOpen
-			? context.toolResults.find(result => PREWALK_ACTION_TOOLS[result.toolName])
+			? context.toolResults.find(result => isPrewalkImplementationAction(result))
 			: undefined;
 		if (!action) {
 			if (!this.#planInjected) {

packages/coding-agent/src/tools/xdev.ts

@@ -36,6 +36,7 @@ import type { RenderResultOptions } from "../extensibility/custom-tools/types";
 import { XD_URL_PREFIX } from "../internal-urls/xd-protocol";
 import type { Theme } from "../modes/theme/theme";
 import { truncateHeadBytes } from "../session/streaming-output";
+import { resolveToolTier, type ToolTier } from "./approval";
 import { renderDefaultToolExecution } from "./default-renderer";
 import type { Tool } from "./index";
 import { replaceTabs } from "./render-utils";
@@ -88,6 +89,13 @@ export interface XdevDispatch {
 	mode: "help" | "execute";
 	/** Validated inner args, kept for renderer delegation on result rebuilds. */
 	args?: Record<string, unknown>;
+	/**
+	 * Approval tier of the wrapped tool for {@link args} (`read` = no workspace
+	 * mutation). Absent for `help` dispatches and calls whose tier could not be
+	 * resolved. Consumed by the prewalk coordinator to skip read-only device
+	 * calls when deciding the model hand-off (issue #7312).
+	 */
+	tier?: ToolTier;
 	/** Details object returned by the wrapped tool, when executed. */
 	inner?: unknown;
 }
@@ -415,7 +423,18 @@ export async function dispatchXdevTool(
 		}
 
 		const validated = parseDeviceArgs(canonical as AiTool, content, toolCallId, () => renderDocs(canonical));
-		xdev = { ...xdev, args: validated };
+		// Record the wrapped tool's approval tier so the prewalk coordinator can
+		// tell a read-only device call (e.g. `lsp` navigation) from a real
+		// workspace mutation without re-decoding the payload. Best-effort: a
+		// throwing approval leaves the tier absent (prewalk then declines to
+		// switch), unlike the write gate which fails closed to `exec`.
+		let tier: ToolTier | undefined;
+		try {
+			tier = resolveToolTier(canonical, validated);
+		} catch {
+			tier = undefined;
+		}
+		xdev = { ...xdev, args: validated, tier };
 		const innerOnUpdate: AgentToolUpdateCallback | undefined = onUpdate
 			? partial =>
 					onUpdate({

packages/coding-agent/test/agent-session-prewalk.test.ts

@@ -400,6 +400,135 @@ describe("AgentSession prewalk", () => {
 		expect(session.model?.id).toBe(primary.id);
 	});
 
+	it("does not switch on a read-only xd:// device dispatched through write (issue #7312)", async () => {
+		const primary = modelOrThrow("claude-sonnet-4-5");
+		const target = modelOrThrow("claude-sonnet-4-6");
+		const modelRegistry = new ModelRegistry(authStorage, path.join(tempDir.path(), "models.yml"));
+
+		// A read-only lsp navigation is dispatched as `write xd://lsp`; the write
+		// result carries the wrapped tool's read tier. Like a bash step, it must
+		// not arm the hand-off — the model keeps reasoning about code shape on the
+		// strong model. Mirrors the bounded-continuation flow: one continuation,
+		// four turns, all primary, then a clean stop.
+		const readDeviceWrite: AgentTool<typeof writeToolSchema, { xdev: { tool: string; mode: string; tier: string } }> =
+			{
+				name: "write",
+				label: "Write",
+				description: "Dispatch a read-only device",
+				parameters: writeToolSchema,
+				async execute() {
+					return {
+						content: [{ type: "text", text: "references" }],
+						details: { xdev: { tool: "lsp", mode: "execute", tier: "read" } },
+					};
+				},
+			};
+		const mock = createMockModel({
+			responses: [
+				toolCall("t1", "record"),
+				toolCall("t2", "write"),
+				{ content: [{ type: "text", text: "Still planning." }], stopReason: "stop" },
+				{ content: [{ type: "text", text: "Done planning." }], stopReason: "stop" },
+			],
+		});
+		const requested: string[] = [];
+		const agent = new Agent({
+			getApiKey: () => "test-key",
+			initialState: {
+				model: primary,
+				systemPrompt: ["Test"],
+				tools: [recordTool as AgentTool, readDeviceWrite as AgentTool],
+				messages: [],
+				thinkingLevel: Effort.Medium,
+			},
+			convertToLlm,
+			streamFn: (model, context, options) => {
+				requested.push(`${model.provider}/${model.id}`);
+				return mock.stream(model, context, options);
+			},
+		});
+		session = new AgentSession({
+			agent,
+			sessionManager: SessionManager.inMemory(),
+			settings: Settings.isolated({ "compaction.enabled": false }),
+			modelRegistry,
+			toolRegistry: new Map([
+				[recordTool.name, recordTool as AgentTool],
+				[readDeviceWrite.name, readDeviceWrite as AgentTool],
+			]),
+			prewalk: { target },
+		});
+
+		await session.prompt("investigate the code shape");
+
+		expect(requested).toEqual(Array(4).fill(`${primary.provider}/${primary.id}`));
+		expect(session.model?.id).toBe(primary.id);
+	});
+
+	it("switches on a write-tier xd:// device dispatched through write (issue #7312)", async () => {
+		const primary = modelOrThrow("claude-sonnet-4-5");
+		const target = modelOrThrow("claude-sonnet-4-6");
+		const modelRegistry = new ModelRegistry(authStorage, path.join(tempDir.path(), "models.yml"));
+
+		// An lsp rename is a write-tier device call — it must arm the hand-off
+		// just like a direct edit/write: the write turn stays on the strong model,
+		// the next turn runs on the target.
+		const writeDeviceWrite: AgentTool<
+			typeof writeToolSchema,
+			{ xdev: { tool: string; mode: string; tier: string } }
+		> = {
+			name: "write",
+			label: "Write",
+			description: "Dispatch a write-tier device",
+			parameters: writeToolSchema,
+			async execute() {
+				return {
+					content: [{ type: "text", text: "renamed" }],
+					details: { xdev: { tool: "lsp", mode: "execute", tier: "write" } },
+				};
+			},
+		};
+		const mock = createMockModel({
+			responses: [toolCall("t1", "record"), toolCall("t2", "write"), { content: ["done"] }],
+		});
+		const requested: string[] = [];
+		const agent = new Agent({
+			getApiKey: () => "test-key",
+			initialState: {
+				model: primary,
+				systemPrompt: ["Test"],
+				tools: [recordTool as AgentTool, writeDeviceWrite as AgentTool],
+				messages: [],
+				thinkingLevel: Effort.Medium,
+			},
+			convertToLlm,
+			streamFn: (model, context, options) => {
+				requested.push(`${model.provider}/${model.id}`);
+				return mock.stream(model, context, options);
+			},
+		});
+		session = new AgentSession({
+			agent,
+			sessionManager: SessionManager.inMemory(),
+			settings: Settings.isolated({ "compaction.enabled": false }),
+			modelRegistry,
+			toolRegistry: new Map([
+				[recordTool.name, recordTool as AgentTool],
+				[writeDeviceWrite.name, writeDeviceWrite as AgentTool],
+			]),
+			prewalk: { target },
+		});
+
+		await session.prompt("rename the symbol");
+
+		expect(requested).toEqual([
+			`${primary.provider}/${primary.id}`,
+			`${primary.provider}/${primary.id}`,
+			`${target.provider}/${target.id}`,
+		]);
+		expect(session.model?.id).toBe(target.id);
+	});
+
 	it("re-arms continuation after tool progress between prose turns", async () => {
 		// Regression: a normal prewalk can split planning across several turns:
 		// prose plan, todo init, then prose before implementation. Each tool

packages/coding-agent/test/write-xdev-dispatch.test.ts

@@ -96,6 +96,9 @@ describe("read and write route xd:// device URLs", () => {
 			expect(previewResult.isError).toBeUndefined();
 			expect(previewResult.details?.xdev?.tool).toBe("ast_edit");
 			expect(previewResult.details?.xdev?.mode).toBe("execute");
+			// The dispatch records the wrapped tool's approval tier so prewalk can
+			// tell a mutation from a read-only device call (issue #7312).
+			expect(previewResult.details?.xdev?.tier).toBe("write");
 			const previewText = previewResult.content.find(entry => entry.type === "text")?.text ?? "";
 			expect(previewText).toContain("modernWrap");
 
@@ -109,6 +112,25 @@ describe("read and write route xd:// device URLs", () => {
 		}
 	});
 
+	it("records a read tier on the dispatch of a read-only device", async () => {
+		const readDevice: AgentTool = {
+			name: "peek",
+			label: "Peek",
+			description: "Read-only device",
+			parameters: type({ q: "string" }),
+			approval: () => "read",
+			async execute() {
+				return { content: [{ type: "text", text: "peeked" }] };
+			},
+		};
+		const xdev = createTestXdevState([readDevice]);
+		const write = new WriteTool(xdevSession(process.cwd(), { xdev }));
+
+		const result = await write.execute("write-xdev-read", { path: "xd://peek", content: JSON.stringify({ q: "x" }) });
+		expect(result.isError).toBeUndefined();
+		expect(result.details?.xdev).toMatchObject({ tool: "peek", mode: "execute", tier: "read" });
+	});
+
 	it("rejects near-miss xd addresses before filesystem fallback", async () => {
 		const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "write-xdev-near-miss-"));
 		try {