pi-pr-review 1.10.2 → 1.10.3

package/.release-please-manifest.json

@@ -1,3 +1,3 @@
 {
-  ".": "1.10.2"
+  ".": "1.10.3"
 }

package/CHANGELOG.md

@@ -1,5 +1,12 @@
 # Changelog
 
+## [1.10.3](https://github.com/10ego/pi-pr-review/compare/v1.10.2...v1.10.3) (2026-07-15)
+
+
+### Bug Fixes
+
+* **publish:** retry invalid review output once ([#27](https://github.com/10ego/pi-pr-review/issues/27)) ([9689505](https://github.com/10ego/pi-pr-review/commit/968950579c4eeec0ac264dccf3bda9b1833df160))
+
 ## [1.10.2](https://github.com/10ego/pi-pr-review/compare/v1.10.1...v1.10.2) (2026-07-15)
 
 

package/README.md

@@ -184,6 +184,8 @@ Publishing is off by default.
 
 The extension—not the model—owns publishing. It creates one formal review with the event hardcoded to `COMMENT`; it never submits `APPROVE` or `REQUEST_CHANGES`. Before writing, it verifies the current PR head, validates inline anchors, and checks for a review of the same head by the current GitHub identity.
 
+If the agent's final review is not valid exact-contract JSON, the extension logs the validation reason and automatically asks the same agent to correct its completed output once, with all tools disabled and without changing the captured posting authority. A valid correction is cached and publication is attempted under the original flags/config. If that single correction is also invalid or attempts to call a tool, publication stops and reports the error instead of looping. Overlapping input cancels the correction, remains unqueued, and must be retried after the agent settles.
+
 Closed or merged PRs use a body-only review. Open PRs attach eligible P0–P3 findings as inline comments and keep nits or off-diff findings in the review body. When multiple findings target the same diff anchor, the first is attached inline and later findings remain in the review body so publication neither fails nor drops review content.
 
 After a review completes, you can directly ask the agent to “post the inline review” or “post it as an inline review.” The extension handles that request directly before an agent turn, selecting the latest cached review for an unnumbered request or the named PR in “publish the review for PR #123.” Only fresh interactive/RPC input can trigger this path; extension-generated, queued, or steering input cannot. The extension publishes only validated cached content, never replacement model-authored text, and never starts or reruns review agents. A direct request permits stale publication.

package/extensions/review-table.ts

@@ -456,6 +456,8 @@ export default function registerReviewTable(
 				session?: CompletedReviewSessionIdentity;
 		  }
 		| undefined;
+	let outputRepairAttempted = false;
+	let outputRepairCancelled = false;
 	const telemetryTracker = new ReviewTelemetryTracker();
 	const persistTelemetry = (completion: ReviewPerformanceTelemetry["completion"]) => {
 		const telemetry = telemetryTracker.finish(completion);
@@ -554,6 +556,8 @@ export default function registerReviewTable(
 		loopCoordinator.clear();
 		selfReviewCoordinator.clear();
 		pendingCompletion = undefined;
+		outputRepairAttempted = false;
+		outputRepairCancelled = false;
 		telemetryTracker.clear();
 	};
 
@@ -573,12 +577,19 @@ export default function registerReviewTable(
 
 	pi.on("agent_settled", () => {
 		selfReviewCoordinator.clear();
+		if (!outputRepairCancelled) return;
+		loopCoordinator.clear();
+		outputRepairAttempted = false;
+		outputRepairCancelled = false;
+		persistTelemetry("cleared");
 	});
 
 	pi.on("session_tree", (event, ctx) => {
 		loopCoordinator.clear();
 		selfReviewCoordinator.clear();
 		pendingCompletion = undefined;
+		outputRepairAttempted = false;
+		outputRepairCancelled = false;
 		restoreCompletedReviews(ctx);
 		telemetryTracker.clear();
 		const session = sessionIdentity(ctx);
@@ -595,6 +606,13 @@ export default function registerReviewTable(
 		// Any new input revokes the prior top-level task generation before it can
 		// authorize a replay or a queued/steering continuation.
 		selfReviewCoordinator.clear();
+		if (outputRepairAttempted) {
+			outputRepairCancelled = true;
+			ctx.abort();
+			ctx.ui.notify("PR review output correction was cancelled; overlapping input was not queued, so retry it when the agent settles", "warning");
+			return { action: "handled" as const };
+		}
+		outputRepairAttempted = false;
 		const source = event.source as ReviewLoopInputSource;
 		const directPublish = parseDirectPublishRequest(event.text);
 		if (
@@ -715,7 +733,20 @@ export default function registerReviewTable(
 	pi.on("message_end", async (event, ctx) => {
 		if (event.message.role !== "assistant") return;
 		const completion = classifyAssistantCompletion(event.message.stopReason, hasToolCall(event.message));
+		if (outputRepairCancelled) {
+			if (completion === "continue_tools") ctx.abort();
+			return;
+		}
 		if (completion === "continue_tools") {
+			if (outputRepairAttempted && loopCoordinator.peek()) {
+				ctx.ui.notify("PR review was not posted: automatic output correction attempted to call tools", "error");
+				ctx.abort();
+				loopCoordinator.clear();
+				outputRepairAttempted = false;
+				outputRepairCancelled = false;
+				persistTelemetry("cleared");
+				return;
+			}
 			const toolCalls = Array.isArray(event.message.content)
 				? event.message.content.filter((part) => part.type === "toolCall")
 				: [];
@@ -729,6 +760,8 @@ export default function registerReviewTable(
 		}
 		if (completion === "clear_invocation") {
 			loopCoordinator.clear();
+			outputRepairAttempted = false;
+			outputRepairCancelled = false;
 			persistTelemetry("cleared");
 			return;
 		}
@@ -744,15 +777,42 @@ export default function registerReviewTable(
 			return;
 		}
 
-		// Every other terminal response consumes authority, whether valid, empty, or unrelated.
+		let publishable = active ? parsePublishableReview(text) : undefined;
+		if (active && !publishable?.review && !outputRepairAttempted) {
+			const error = publishable?.error ?? "final review JSON is invalid";
+			if (loopCoordinator.suspendToolsForRepair()) {
+				outputRepairAttempted = true;
+				outputRepairCancelled = false;
+				ctx.ui.notify(`PR review output is invalid (${error}); automatically correcting it once`, "warning");
+				pi.sendMessage(
+					{
+						customType: "pr-review-output-repair",
+						content: [
+							`The completed PR review could not be accepted because ${error}.`,
+							"This is the only automatic correction attempt. Do not call tools, redo the review, or change its substance.",
+							"Re-emit the same completed review as exactly one JSON object satisfying the required OUTPUT FORMAT. Emit no Markdown fences, prose, headings, or trailing text.",
+						].join(" "),
+						display: false,
+						details: { error, attempt: 1 },
+					},
+					{ triggerTurn: true, deliverAs: "followUp" },
+				);
+				return;
+			}
+			ctx.ui.notify("Automatic PR review output correction was skipped because tools could not be disabled", "warning");
+		}
+
+		// A valid response, or a failed single correction, consumes authority.
 		// Persist timing before publication so network/write latency is never coupled to review wall time.
 		const invocation = active ? loopCoordinator.consume() : undefined;
-		if (invocation) persistTelemetry("terminal_response");
-		if (!text.trim()) return;
-		const review = parseReview(text);
-		if (!review) return; // not a /pr-review JSON payload — leave untouched
 		if (invocation) {
-			const publishable = parsePublishableReview(text);
+			outputRepairAttempted = false;
+			outputRepairCancelled = false;
+			persistTelemetry("terminal_response");
+		}
+		const review = text.trim() ? parseReview(text) : null;
+		if (invocation) {
+			publishable ??= parsePublishableReview(text);
 			let repository: RepositoryBinding | undefined;
 			let record: CompletedReviewRecord | undefined;
 			let session: CompletedReviewSessionIdentity | undefined;
@@ -775,6 +835,7 @@ export default function registerReviewTable(
 			}
 			pendingCompletion = { text, invocation, repository, record, session };
 		}
+		if (!review) return; // not a renderable /pr-review JSON payload — leave untouched
 
 		// Keep raw JSON for automation; only prettify for interactive terminals.
 		if (ctx.mode !== "tui") return;

package/lib/pr-review-loop.ts

@@ -76,11 +76,13 @@ export class ReviewLoopCoordinator {
 	private readonly invocationGate = new ReviewInvocationGate();
 	private readonly focusRegistry = new ReviewFocusRegistry();
 	private binding?: ReviewLoopBinding;
+	private suspendedTools?: string[];
 	private nextGeneration = 1;
 
 	constructor(private readonly pi: Pick<ExtensionAPI, "getActiveTools" | "setActiveTools">) {}
 
 	private setToolsEnabled(enabled: boolean): void {
+		if (this.suspendedTools !== undefined) return;
 		try {
 			const current = this.pi.getActiveTools();
 			const next = enabled
@@ -150,7 +152,9 @@ export class ReviewLoopCoordinator {
 
 	acquire(ctx: Pick<ExtensionContext, "cwd" | "sessionManager">): ReviewLoopLease | undefined {
 		const phase = this.invocationGate.phase();
-		if ((phase !== "reviewing" && phase !== "confirmed") || !this.binding) return undefined;
+		if (this.suspendedTools !== undefined || (phase !== "reviewing" && phase !== "confirmed") || !this.binding) {
+			return undefined;
+		}
 		if (!sameBinding(this.binding, ctx) || this.binding.controller.signal.aborted) {
 			this.clear();
 			return undefined;
@@ -165,6 +169,7 @@ export class ReviewLoopCoordinator {
 		lease: ReviewLoopLease,
 		ctx: Pick<ExtensionContext, "cwd" | "sessionManager">,
 	): boolean {
+		if (this.suspendedTools !== undefined) return false;
 		const active = !!this.binding &&
 			this.binding.generation === lease.generation &&
 			!lease.signal.aborted &&
@@ -219,11 +224,48 @@ export class ReviewLoopCoordinator {
 		this.setToolsEnabled(false);
 	}
 
+	/** Hide every tool for a format-only repair turn while retaining invocation authority. */
+	suspendToolsForRepair(): boolean {
+		const phase = this.invocationGate.phase();
+		if (
+			this.suspendedTools !== undefined ||
+			!this.binding ||
+			(phase !== "reviewing" && phase !== "confirmed")
+		) {
+			return false;
+		}
+		let baseTools: string[] = [];
+		try {
+			baseTools = this.pi.getActiveTools().filter((name) => !REVIEW_LOOP_TOOL_SET.has(name));
+			this.suspendedTools = baseTools;
+			this.pi.setActiveTools([]);
+			return true;
+		} catch {
+			this.suspendedTools = undefined;
+			try {
+				this.pi.setActiveTools(baseTools);
+			} catch {
+				// Failure to restore remains fail-closed with no repair turn started.
+			}
+			return false;
+		}
+	}
+
 	private revokeBinding(): void {
 		const generation = this.binding?.generation;
 		if (generation !== undefined) this.focusRegistry.close(generation);
 		this.binding?.controller.abort();
 		this.binding = undefined;
+		const suspendedTools = this.suspendedTools;
+		this.suspendedTools = undefined;
+		if (suspendedTools !== undefined) {
+			try {
+				this.pi.setActiveTools(suspendedTools);
+			} catch {
+				// Keep tools fail-closed if the extension runtime is shutting down.
+			}
+			return;
+		}
 		this.setToolsEnabled(false);
 	}
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-pr-review",
-  "version": "1.10.2",
+  "version": "1.10.3",
   "description": "Parallel AI code review for GitHub pull requests in the Pi coding agent, with model-agnostic tiered subagents, structured findings, optional verification, and safe COMMENT-only publishing.",
   "keywords": [
     "pi-package",

package/tests/pr-review-extension-lifecycle.test.ts

@@ -69,6 +69,7 @@ interface Harness {
 	tools: Map<string, any>;
 	branch: any[];
 	notifications: string[];
+	sentMessages: Array<{ message: any; options: any }>;
 	activeTools(): string[];
 	abortCount(): number;
 	selfReviewCoordinator: SelfReviewPermitCoordinator;
@@ -156,6 +157,7 @@ function createHarness(initialBranch: any[] = [], identity = session): Harness {
 	const tools = new Map<string, any>();
 	const branch = [...initialBranch];
 	const notifications: string[] = [];
+	const sentMessages: Array<{ message: any; options: any }> = [];
 	let activeTools = ["read", "bash", ...REVIEW_LOOP_TOOL_NAMES];
 	let aborts = 0;
 	let promptPath = ownPromptPath;
@@ -191,6 +193,9 @@ function createHarness(initialBranch: any[] = [], identity = session): Harness {
 		appendEntry: (customType: string, data: unknown) => {
 			branch.push({ type: "custom", id: `custom-${nextId++}`, customType, data });
 		},
+		sendMessage: (message: any, options: any) => {
+			sentMessages.push({ message, options });
+		},
 		getActiveTools: () => [...activeTools],
 		setActiveTools: (next: string[]) => {
 			activeTools = [...next];
@@ -210,6 +215,7 @@ function createHarness(initialBranch: any[] = [], identity = session): Harness {
 		tools,
 		branch,
 		notifications,
+		sentMessages,
 		activeTools: () => [...activeTools],
 		abortCount: () => aborts,
 		selfReviewCoordinator,
@@ -241,6 +247,132 @@ function persistedInlineReview(identity = session, allowStalePublish = true): an
 }
 
 describe("completed review extension lifecycle", () => {
+	test("automatically corrects invalid final JSON once and attempts publication", async () => {
+		const harness = createHarness();
+		await harness.emit("input", { text: "/pr-review 7 --comment", source: "interactive" });
+		const wrapped = {
+			role: "assistant",
+			stopReason: "stop",
+			content: [{ type: "text", text: `\`\`\`json\n${JSON.stringify(review)}\n\`\`\`` }],
+		};
+		await harness.emit("message_end", { message: wrapped });
+		expect(harness.sentMessages).toHaveLength(1);
+		expect(harness.sentMessages[0]?.message).toMatchObject({
+			customType: "pr-review-output-repair",
+			display: false,
+		});
+		expect(harness.sentMessages[0]?.message.content).toContain("exactly one JSON object");
+		expect(harness.sentMessages[0]?.options).toEqual({ triggerTurn: true, deliverAs: "followUp" });
+		expect(harness.notifications.some((message) => message.includes("automatically correcting"))).toBeTrue();
+		expect(harness.activeTools()).toEqual([]);
+
+		harness.appendMessage(wrapped, "wrapped-review");
+		await harness.emit("turn_end", { message: wrapped, toolResults: [] });
+		expect(harness.notifications.some((message) => message.includes("was not posted"))).toBeFalse();
+
+		const payloadPath = installFakePublishingGh();
+		const corrected = {
+			role: "assistant",
+			stopReason: "stop",
+			content: [{ type: "text", text: JSON.stringify(review) }],
+		};
+		await harness.emit("message_end", { message: corrected });
+		harness.appendMessage(corrected, "corrected-review");
+		await harness.emit("turn_end", { message: corrected, toolResults: [] });
+
+		expect(harness.sentMessages).toHaveLength(1);
+		expect(harness.notifications.some((message) => message.includes("PR review posted"))).toBeTrue();
+		expect(JSON.parse(readFileSync(payloadPath, "utf8")).body).toContain("Checks lifecycle persistence");
+		expect(harness.activeTools()).toEqual(BASE_ACTIVE_TOOLS);
+	});
+
+	test("aborts and revokes repair authority when the correction attempts a tool call", async () => {
+		const harness = createHarness();
+		await harness.emit("input", { text: "/pr-review 7 --comment", source: "interactive" });
+		const invalid = {
+			role: "assistant",
+			stopReason: "stop",
+			content: [{ type: "text", text: "not json" }],
+		};
+		await harness.emit("message_end", { message: invalid });
+		expect(harness.activeTools()).toEqual([]);
+
+		await harness.emit("message_end", {
+			message: {
+				role: "assistant",
+				stopReason: "toolUse",
+				content: [{ type: "toolCall", id: "repair-bash", name: "bash", arguments: { command: "echo unsafe" } }],
+			},
+		});
+
+		expect(harness.abortCount()).toBe(1);
+		expect(harness.sentMessages).toHaveLength(1);
+		expect(harness.notifications.some((message) => message.includes("correction attempted to call tools"))).toBeTrue();
+		expect(harness.activeTools()).toEqual(BASE_ACTIVE_TOOLS);
+	});
+
+	test("keeps tools suspended until a cancelled repair definitively settles", async () => {
+		const harness = createHarness();
+		await harness.emit("input", { text: "/pr-review 7 --comment", source: "interactive" });
+		const invalid = {
+			role: "assistant",
+			stopReason: "stop",
+			content: [{ type: "text", text: "not json" }],
+		};
+		await harness.emit("message_end", { message: invalid });
+		expect(harness.activeTools()).toEqual([]);
+
+		const overlap = await harness.emit("input", {
+			text: "do something unrelated",
+			source: "interactive",
+			streamingBehavior: "steer",
+		});
+		expect(overlap).toContainEqual({ action: "handled" });
+		expect(harness.abortCount()).toBe(1);
+		expect(harness.activeTools()).toEqual([]);
+		expect(harness.notifications.some((message) => message.includes("was not queued"))).toBeTrue();
+
+		const staleCorrection = {
+			role: "assistant",
+			stopReason: "stop",
+			content: [{ type: "text", text: JSON.stringify(review) }],
+		};
+		await harness.emit("message_end", { message: staleCorrection });
+		harness.appendMessage(staleCorrection, "stale-correction");
+		await harness.emit("turn_end", { message: staleCorrection, toolResults: [] });
+		expect(harness.notifications.some((message) => message.includes("PR review posted"))).toBeFalse();
+		expect(harness.activeTools()).toEqual([]);
+
+		await harness.emit("agent_settled", {});
+		expect(harness.activeTools()).toEqual(BASE_ACTIVE_TOOLS);
+	});
+
+	test("stops after one automatic correction attempt", async () => {
+		const harness = createHarness();
+		await harness.emit("input", { text: "/pr-review 7 --comment", source: "interactive" });
+		const invalid = {
+			role: "assistant",
+			stopReason: "stop",
+			content: [{ type: "text", text: "not json" }],
+		};
+		await harness.emit("message_end", { message: invalid });
+		expect(harness.sentMessages).toHaveLength(1);
+		expect(harness.activeTools()).toEqual([]);
+		await harness.emit("turn_end", { message: invalid, toolResults: [] });
+
+		await harness.emit("message_end", { message: invalid });
+		harness.appendMessage(invalid, "invalid-retry");
+		await harness.emit("turn_end", { message: invalid, toolResults: [] });
+
+		expect(harness.sentMessages).toHaveLength(1);
+		expect(
+			harness.notifications.some((message) =>
+				message.includes("PR review was not posted: final response is not exactly one JSON object"),
+			),
+		).toBeTrue();
+		expect(harness.activeTools()).toEqual(BASE_ACTIVE_TOOLS);
+	});
+
 	test("persists a reference before publishing after Pi stores exact assistant JSON", async () => {
 		const harness = createHarness();
 		await harness.emit("input", { text: "/pr-review 7 --comment", source: "interactive" });

package/tests/pr-review-loop.test.ts

@@ -55,6 +55,29 @@ describe("review-loop authority", () => {
 		expect(h.coordinator.acquire(h.ctx as any)).toBeDefined();
 	});
 
+	test("suspends every tool for output repair and restores only base tools", () => {
+		const h = harness();
+		h.coordinator.begin(parsePublishMode("/pr-review 7"), autoOff, "interactive", h.ctx as any);
+		const lease = h.coordinator.acquire(h.ctx as any)!;
+
+		expect(h.coordinator.suspendToolsForRepair()).toBeTrue();
+		expect(h.activeTools()).toEqual([]);
+		expect(h.coordinator.acquire(h.ctx as any)).toBeUndefined();
+		expect(h.coordinator.isLeaseActive(lease, h.ctx as any)).toBeFalse();
+		expect(lease.signal.aborted).toBeFalse();
+		expect(h.coordinator.suspendToolsForRepair()).toBeFalse();
+
+		expect(h.coordinator.consume()?.prNumber).toBe(7);
+		expect(lease.signal.aborted).toBeTrue();
+		expect(h.activeTools()).toEqual(["read", "bash"]);
+
+		h.coordinator.begin(parsePublishMode("/pr-review 8"), autoOff, "interactive", h.ctx as any);
+		expect(h.coordinator.suspendToolsForRepair()).toBeTrue();
+		h.coordinator.clear();
+		expect(h.activeTools()).toEqual(["read", "bash"]);
+		expect(h.coordinator.suspendToolsForRepair()).toBeFalse();
+	});
+
 	test("never authorizes extension-originated commands", () => {
 		const h = harness();
 		h.coordinator.hideTools();

package/tests/pr-review-publish.test.ts

@@ -553,6 +553,7 @@ fi
 		expect(extension).toContain("Publishing never starts or reruns a review");
 		expect(extension).toContain("review was cancelled");
 		expect(readme).toContain("handles that request directly");
+		expect(readme).toContain("automatically asks the same agent to correct its completed output once");
 		expect(readme).toContain("`allowStalePublish: true`");
 		expect(readme).toContain("/pr-review-publish 123 --allow-stale");
 		expect(readme).toContain("Inline comments are always disabled for stale reviews");