feat(prompt): implement prompt injection slice 4

2026-06-11 22:54:51 +02:00
parent 19666c900e
commit 58e8bd31d3
5 changed files with 965 additions and 18 deletions
@@ -10,6 +10,8 @@ import {
 	modeAllowsPreInitHint,
 	modeAllowsInjection,
 	modeRequiresProtocolPath,
+	hasProtocolPath,
+	buildAdvisoryReminder,
 	discoverContextWindow,
 	estimateTokens,
 	findAllArtifactPairs,
@@ -20,9 +22,15 @@ import {
 	isRelevantTurnForReinjection,
 	detectEditIntent,
 	detectArchitectureSensitiveReasoning,
+	isSensitiveAction,
+	messagesHaveBypass,
+	extractBypassReason,
+	evaluateStrictBypass,
+	buildStrictBypassGuard,
 	ROOT_PAIR_START_MARKER,
 	ROOT_PAIR_END_MARKER,
 	TRUST_BOUNDARY_TEXT,
+	BYPASS_MARKER_PREFIX,
 } from "../src/prompt-injection.js";

 describe("prompt-injection helpers", () => {
@@ -708,6 +716,73 @@ describe("prompt-injection helpers", () => {
 		});
 	});

+	describe("protocol path detection", () => {
+		it("detects protocol path when messages contain marker and trust boundary", () => {
+			const messages = [
+				{
+					content: `Start ${ROOT_PAIR_START_MARKER} ${TRUST_BOUNDARY_TEXT} end`,
+				},
+			];
+			expect(hasProtocolPath(messages)).toBe(true);
+		});
+
+		it("detects protocol path when payload contains marker and trust boundary", () => {
+			const payload = {
+				body: `${ROOT_PAIR_START_MARKER}\n${TRUST_BOUNDARY_TEXT}`,
+			};
+			expect(hasProtocolPath(undefined, payload)).toBe(true);
+		});
+
+		it("returns false when only marker is present", () => {
+			const messages = [{ content: `hello ${ROOT_PAIR_START_MARKER} world` }];
+			expect(hasProtocolPath(messages)).toBe(false);
+		});
+
+		it("returns false when only trust boundary is present", () => {
+			const messages = [{ content: `hello ${TRUST_BOUNDARY_TEXT} world` }];
+			expect(hasProtocolPath(messages)).toBe(false);
+		});
+
+		it("returns false when neither marker nor trust boundary is present", () => {
+			const messages = [{ content: "plain text" }];
+			expect(hasProtocolPath(messages)).toBe(false);
+		});
+
+		it("returns false for empty messages and undefined payload", () => {
+			expect(hasProtocolPath([])).toBe(false);
+			expect(hasProtocolPath(undefined, undefined)).toBe(false);
+		});
+
+		it("handles array content blocks", () => {
+			const messages = [
+				{
+					content: [
+						{ type: "text", text: ROOT_PAIR_START_MARKER },
+						{ type: "text", text: TRUST_BOUNDARY_TEXT },
+					],
+				},
+			];
+			expect(hasProtocolPath(messages)).toBe(true);
+		});
+
+		it("requires both elements in the same combined source", () => {
+			const messages = [{ content: ROOT_PAIR_START_MARKER }];
+			const payload = { body: TRUST_BOUNDARY_TEXT };
+			expect(hasProtocolPath(messages, payload)).toBe(true);
+		});
+	});
+
+	describe("advisory reminder", () => {
+		it("returns a visible lightweight reminder with no synthetic artifacts", () => {
+			const reminder = buildAdvisoryReminder();
+			expect(reminder).toContain("advisory mode active");
+			expect(reminder).toContain("📋");
+			expect(reminder).not.toContain(ROOT_PAIR_START_MARKER);
+			expect(reminder).not.toContain("### Root index");
+			expect(reminder).not.toContain("### Root map");
+		});
+	});
+
 	describe("injection payload", () => {
 		it("always includes the root pair", () => {
 			const dir = mkdtempSync(join(tmpdir(), "pi-payload-test-"));
@@ -769,4 +844,352 @@ describe("prompt-injection helpers", () => {
 			expect(payload.content).not.toContain("# src map x");
 		});
 	});
+
+	describe("sensitive action detection", () => {
+		it("detects explicit edit_intent event type", () => {
+			expect(isSensitiveAction("edit_intent", [])).toBe(true);
+		});
+
+		it("detects explicit architecture_sensitive event type", () => {
+			expect(isSensitiveAction("architecture_sensitive", [])).toBe(true);
+		});
+
+		it("falls back to edit-intent heuristic for generic events", () => {
+			expect(
+				isSensitiveAction("user_chat", [{ content: "edit the file" }]),
+			).toBe(true);
+		});
+
+		it("falls back to architecture heuristic for generic events", () => {
+			expect(
+				isSensitiveAction("user_chat", [
+					{ content: "this is an architecture decision" },
+				]),
+			).toBe(true);
+		});
+
+		it("treats compaction as non-sensitive", () => {
+			expect(isSensitiveAction("compaction", [{ content: "compacting" }])).toBe(
+				false,
+			);
+		});
+
+		it("treats agent_start as non-sensitive", () => {
+			expect(isSensitiveAction("agent_start", [{ content: "hi" }])).toBe(false);
+		});
+
+		it("returns false for irrelevant messages without event type", () => {
+			expect(
+				isSensitiveAction(undefined, [{ content: "what is the weather?" }]),
+			).toBe(false);
+		});
+
+		it("returns false when no event type or messages are provided", () => {
+			expect(isSensitiveAction(undefined, undefined)).toBe(false);
+		});
+
+		it("handles array content blocks", () => {
+			expect(
+				isSensitiveAction(undefined, [
+					{
+						content: [
+							{ type: "text", text: "please " },
+							{ type: "text", text: "refactor the module" },
+						],
+					},
+				]),
+			).toBe(true);
+		});
+
+		it("detects sensitive action from payload when messages are empty", () => {
+			expect(isSensitiveAction("user_chat", [], "edit the file")).toBe(true);
+		});
+
+		it("detects architecture-sensitive action from payload", () => {
+			expect(isSensitiveAction("user_chat", [], "architecture decision")).toBe(
+				true,
+			);
+		});
+
+		it("returns false for irrelevant payload without messages or event type", () => {
+			expect(
+				isSensitiveAction(undefined, undefined, "what is the weather?"),
+			).toBe(false);
+		});
+
+		it("combines messages and payload for heuristic detection", () => {
+			expect(
+				isSensitiveAction(
+					"user_chat",
+					[{ content: "please consider" }],
+					"the architecture decision",
+				),
+			).toBe(true);
+		});
+	});
+
+	describe("bypass marker parsing", () => {
+		it("detects a valid bypass marker in messages", () => {
+			expect(
+				messagesHaveBypass([{ content: "[PI_MAP_BYPASS: temporary fix]" }]),
+			).toBe(true);
+		});
+
+		it("detects a valid bypass marker prefix", () => {
+			expect(BYPASS_MARKER_PREFIX).toBe("[PI_MAP_BYPASS:");
+		});
+
+		it("returns false when no marker is present", () => {
+			expect(messagesHaveBypass([{ content: "plain text" }])).toBe(false);
+		});
+
+		it("extracts the bypass reason", () => {
+			expect(
+				extractBypassReason([{ content: "[PI_MAP_BYPASS: temporary fix]" }]),
+			).toBe("temporary fix");
+		});
+
+		it("returns undefined when marker is absent", () => {
+			expect(extractBypassReason([{ content: "plain text" }])).toBeUndefined();
+		});
+
+		it("rejects an empty bypass reason", () => {
+			expect(
+				extractBypassReason([{ content: "[PI_MAP_BYPASS:]" }]),
+			).toBeUndefined();
+		});
+
+		it("rejects a whitespace-only bypass reason", () => {
+			expect(
+				extractBypassReason([{ content: "[PI_MAP_BYPASS:   ]" }]),
+			).toBeUndefined();
+		});
+
+		it("extracts the first reason when multiple markers are present", () => {
+			expect(
+				extractBypassReason([
+					{
+						content: "[PI_MAP_BYPASS: first] [PI_MAP_BYPASS: second]",
+					},
+				]),
+			).toBe("first");
+		});
+
+		it("handles array content blocks", () => {
+			expect(
+				extractBypassReason([
+					{
+						content: [
+							{ type: "text", text: "[PI_MAP_BYPASS: " },
+							{ type: "text", text: "array reason]" },
+						],
+					},
+				]),
+			).toBe("array reason");
+		});
+
+		it("detects bypass marker in payload string", () => {
+			expect(
+				messagesHaveBypass(undefined, "[PI_MAP_BYPASS: payload reason]"),
+			).toBe(true);
+		});
+
+		it("detects bypass marker in payload object", () => {
+			expect(
+				messagesHaveBypass(undefined, {
+					body: "[PI_MAP_BYPASS: object reason]",
+				}),
+			).toBe(true);
+		});
+
+		it("extracts reason from payload when messages have none", () => {
+			expect(
+				extractBypassReason(undefined, "[PI_MAP_BYPASS: payload reason]"),
+			).toBe("payload reason");
+		});
+
+		it("prefers first marker across messages and payload", () => {
+			expect(
+				extractBypassReason(
+					[{ content: "[PI_MAP_BYPASS: msg reason]" }],
+					"[PI_MAP_BYPASS: payload reason]",
+				),
+			).toBe("msg reason");
+		});
+
+		it("extracts reason with colon inside", () => {
+			expect(
+				extractBypassReason([{ content: "[PI_MAP_BYPASS: fix: edge case]" }]),
+			).toBe("fix: edge case");
+		});
+
+		it("rejects bypass marker with only newlines/spaces", () => {
+			expect(
+				extractBypassReason([{ content: "[PI_MAP_BYPASS:\n\t ]" }]),
+			).toBeUndefined();
+		});
+	});
+
+	describe("strict bypass evaluation", () => {
+		it("never guards in non-strict modes", () => {
+			for (const mode of ["off", "advisory", "strong"] as const) {
+				const decision = evaluateStrictBypass(
+					{ messages: [{ content: "edit the file" }], type: "edit_intent" },
+					mode,
+				);
+				expect(decision.guard).toBe(false);
+			}
+		});
+
+		it("does not guard non-sensitive turns in strict mode", () => {
+			const decision = evaluateStrictBypass(
+				{ messages: [{ content: "hi" }], type: "user_chat" },
+				"strict",
+			);
+			expect(decision.guard).toBe(false);
+		});
+
+		it("guards sensitive turns when protocol path is missing", () => {
+			const decision = evaluateStrictBypass(
+				{ messages: [{ content: "edit the file" }], type: "edit_intent" },
+				"strict",
+			);
+			expect(decision.guard).toBe(true);
+			expect(decision.reason).toContain("protocol path missing");
+		});
+
+		it("does not guard sensitive turns when protocol path is present", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [
+						{
+							content: `${ROOT_PAIR_START_MARKER} ${TRUST_BOUNDARY_TEXT}`,
+						},
+					],
+					type: "edit_intent",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(false);
+		});
+
+		it("uses heuristic fallback for generic sensitive events", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [{ content: "refactor the module" }],
+					type: "user_chat",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(true);
+		});
+
+		it("clears the guard when a valid inline bypass marker is present", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [
+						{
+							content: "[PI_MAP_BYPASS: emergency patch] edit the file",
+						},
+					],
+					type: "edit_intent",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(false);
+			expect(decision.bypassMarker).toBe("emergency patch");
+		});
+
+		it("does not clear the guard for an empty bypass marker", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [{ content: "[PI_MAP_BYPASS:] edit the file" }],
+					type: "edit_intent",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(true);
+		});
+
+		it("does not guard when event is null or undefined", () => {
+			expect(evaluateStrictBypass(undefined, "strict").guard).toBe(false);
+			expect(evaluateStrictBypass(null, "strict").guard).toBe(false);
+		});
+
+		it("does not guard when payload-only protocol path is present", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [{ content: "edit the file" }],
+					payload: `${ROOT_PAIR_START_MARKER}\n${TRUST_BOUNDARY_TEXT}`,
+					type: "edit_intent",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(false);
+		});
+
+		it("clears guard when bypass marker is only in payload", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [{ content: "edit the file" }],
+					payload: "[PI_MAP_BYPASS: emergency patch]",
+					type: "edit_intent",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(false);
+			expect(decision.bypassMarker).toBe("emergency patch");
+		});
+
+		it("guards when payload-only bypass marker has empty reason", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [{ content: "edit the file" }],
+					payload: "[PI_MAP_BYPASS:]",
+					type: "edit_intent",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(true);
+		});
+
+		it("uses heuristic fallback for payload-only sensitive events", () => {
+			const decision = evaluateStrictBypass(
+				{
+					messages: [],
+					payload: "refactor the module",
+					type: "user_chat",
+				},
+				"strict",
+			);
+			expect(decision.guard).toBe(true);
+		});
+	});
+
+	describe("strict bypass guard message", () => {
+		it("includes the provided reason", () => {
+			const guard = buildStrictBypassGuard("custom reason");
+			expect(guard).toContain("custom reason");
+			expect(guard).toContain("🛑");
+		});
+
+		it("includes a default reason when none is provided", () => {
+			const guard = buildStrictBypassGuard();
+			expect(guard).toContain("sensitive action");
+			expect(guard).toContain("protocol path");
+		});
+
+		it("tells the user how to proceed", () => {
+			const guard = buildStrictBypassGuard();
+			expect(guard).toContain("PI_MAP_BYPASS");
+			expect(guard).toContain("restore the project-map context");
+		});
+
+		it("does not contain synthetic root-pair content", () => {
+			const guard = buildStrictBypassGuard();
+			expect(guard).not.toContain(ROOT_PAIR_START_MARKER);
+			expect(guard).not.toContain("### Root index");
+			expect(guard).not.toContain("### Root map");
+		});
+	});
 });