fix(prewalk): stopped completion turn loop

Limited the hidden continuation safety net to the assistant turn immediately following the plan nudge, so later bash-only completion ends normally.

Added regression coverage for commit-style flows that never call edit or write.

Fixes #5551

packages/coding-agent/CHANGELOG.md

@@ -13,6 +13,7 @@
 ### Fixed
 
 - Fixed Bash internal URLs remaining unresolved when used as unquoted arguments inside command substitutions ([#5535](https://github.com/can1357/oh-my-pi/issues/5535)).
+- Fixed prewalk repeatedly continuing after a bash-only task such as `commit` had already completed ([#5551](https://github.com/can1357/oh-my-pi/issues/5551)).
 
 ## [16.5.2] - 2026-07-14
 

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

@@ -1748,6 +1748,8 @@ export class AgentSession {
 	#prewalk: Prewalk | undefined;
 	/** True once the plan nudge has been queued; scrubbed from context at the switch. */
 	#prewalkPlanInjected = false;
+	/** True until the first assistant turn after the plan nudge completes. */
+	#prewalkContinuePending = false;
 	/** True once any successful `todo` call landed — opens the prewalk
 	 *  trigger gate: the switch fires at the first edit/write AFTER the todo
 	 *  list exists (sessions without an ACTIVE todo tool skip the gate). */
@@ -2247,23 +2249,22 @@ export class AgentSession {
 		const prewalk = this.#prewalk;
 		if (!prewalk || context?.message.role !== "assistant") return;
 
-		// Structural safety net: every branch below assumes the agent loop will
-		// run another turn. It won't if THIS turn had no tool calls — the loop
-		// treats a text-only turn as "the agent is done" and ends the session
-		// with no further prompting. The plan nudge explicitly asks for a prose
-		// reply, which makes a text-only turn common right after it — observed
-		// silently killing production SWE-bench runs before any code was ever
-		// written. Force one more turn only in that specific, self-created
-		// hazard window.
-		if (this.#prewalkPlanInjected && context.toolResults.length === 0) {
-			this.agent.steer({
-				role: "custom",
-				customType: PREWALK_CONTINUE_MESSAGE_TYPE,
-				content: prewalkContinuePrompt,
-				attribution: "agent",
-				display: false,
-				timestamp: Date.now(),
-			});
+		// The plan nudge can produce a prose-only reply, which the agent loop
+		// treats as completion before any implementation starts. Keep the
+		// safety net open only for the turn immediately following that nudge:
+		// once the model calls any tool, later text-only completion is genuine.
+		if (this.#prewalkContinuePending) {
+			this.#prewalkContinuePending = false;
+			if (context.toolResults.length === 0) {
+				this.agent.steer({
+					role: "custom",
+					customType: PREWALK_CONTINUE_MESSAGE_TYPE,
+					content: prewalkContinuePrompt,
+					attribution: "agent",
+					display: false,
+					timestamp: Date.now(),
+				});
+			}
 		}
 
 		// Todo gate: the plan nudge instructs "finish the plan, then init the
@@ -2284,6 +2285,7 @@ export class AgentSession {
 		if (!action) {
 			if (!this.#prewalkPlanInjected) {
 				this.#prewalkPlanInjected = true;
+				this.#prewalkContinuePending = true;
 				this.agent.steer({
 					role: "custom",
 					customType: PREWALK_PLAN_MESSAGE_TYPE,
@@ -2344,6 +2346,7 @@ export class AgentSession {
 		}
 		this.#prewalk = { target, thinkingLevel };
 		this.#prewalkPlanInjected = true;
+		this.#prewalkContinuePending = true;
 		this.agent.steer({
 			role: "custom",
 			customType: PREWALK_PLAN_MESSAGE_TYPE,

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

@@ -291,6 +291,53 @@ describe("AgentSession prewalk", () => {
 		]);
 		expect(session.model?.id).toBe(target.id);
 	});
+
+	it("does not continue a completed bash-only task after the plan-nudge window closes", 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"));
+
+		const mock = createMockModel({
+			responses: [
+				toolCall("t1", "record"),
+				toolCall("t2", "bash"),
+				{ content: [{ type: "text", text: "Commit complete." }], stopReason: "stop" },
+			],
+		});
+		const requested: string[] = [];
+		const agent = new Agent({
+			getApiKey: () => "test-key",
+			initialState: {
+				model: primary,
+				systemPrompt: ["Test"],
+				tools: [recordTool as AgentTool, bashTool 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,
+			prewalk: { target },
+		});
+
+		await session.prompt("commit the current changes");
+
+		expect(requested).toEqual([
+			`${primary.provider}/${primary.id}`,
+			`${primary.provider}/${primary.id}`,
+			`${primary.provider}/${primary.id}`,
+		]);
+	});
+
 	it("skips the todo gate when todo is registered but not active (subagent-style restricted slates)", async () => {
 		// Regression: the gate used to key on the tool REGISTRY, so a session
 		// whose active-tool slate excluded `todo` (subagents strip it) while the