@gajae-code/coding-agent 0.18.5 → 0.18.7

package/src/tools/bash.ts

@@ -19,8 +19,10 @@ import {
 } from "../gjc-runtime/managed-owner-admission";
 import {
 	MANAGED_OWNER_CHILD_TOKEN_ENV,
+	MANAGED_OWNER_COMMAND_ENV,
 	MANAGED_OWNER_GENERATION_ENV,
 	MANAGED_OWNER_INCARNATION_ENV,
+	MANAGED_OWNER_REDACT_COMMAND_ENV,
 	MANAGED_OWNER_RUN_ID_ENV,
 	MANAGED_OWNER_STATE_DIR_ENV,
 } from "../gjc-runtime/managed-owner-supervisor";
@@ -118,6 +120,8 @@ export const MANAGED_OWNER_BASH_ENV = [
 	MANAGED_OWNER_RUN_ID_ENV,
 	MANAGED_OWNER_INCARNATION_ENV,
 	MANAGED_OWNER_CHILD_TOKEN_ENV,
+	MANAGED_OWNER_COMMAND_ENV,
+	MANAGED_OWNER_REDACT_COMMAND_ENV,
 	MANAGED_OWNER_PREDECESSOR_TOKEN_ENV,
 	MANAGED_OWNER_PREDECESSOR_GENERATION_ENV,
 	MANAGED_OWNER_PREDECESSOR_RUN_ID_ENV,
@@ -1278,7 +1282,7 @@ export class BashTool implements AgentTool<BashToolSchema, BashToolDetails> {
 	 */
 	async #waitForManagedBashJob(
 		job: ManagedBashJobHandle,
-		thresholdMs: number,
+		thresholdMs: number | undefined,
 		signal?: AbortSignal,
 		backgroundRequest?: Promise<FoldReason>,
 		foldAdapter?: FoldAdapter,
@@ -1288,31 +1292,36 @@ export class BashTool implements AgentTool<BashToolSchema, BashToolDetails> {
 		}
 
 		const startedAt = Date.now();
-		const threshold = Promise.withResolvers<{ kind: "running"; reason: FoldReason }>();
-		const thresholdTimer = setTimeout(
-			() => {
-				const requestFold = this.session.requestForegroundBashBackground;
-				if (!foldAdapter || !requestFold) {
-					threshold.resolve({ kind: "running", reason: "timer" });
-					return;
-				}
-				void requestFold("timer", foldAdapter)
-					.then(folded => {
-						if (!folded) threshold.resolve({ kind: "running", reason: "timer" });
-					})
-					.catch(error => {
-						logger.warn("Timer-triggered fold failed", {
-							jobId: foldAdapter.jobId,
-							error: error instanceof Error ? error.message : String(error),
-						});
-						threshold.resolve({ kind: "running", reason: "timer" });
-					});
-			},
-			Math.max(0, thresholdMs),
-		);
+		const threshold =
+			thresholdMs === undefined ? undefined : Promise.withResolvers<{ kind: "running"; reason: FoldReason }>();
+		const thresholdTimer =
+			threshold === undefined
+				? undefined
+				: setTimeout(
+						() => {
+							const requestFold = this.session.requestForegroundBashBackground;
+							if (!foldAdapter || !requestFold) {
+								threshold.resolve({ kind: "running", reason: "timer" });
+								return;
+							}
+							void requestFold("timer", foldAdapter)
+								.then(folded => {
+									if (!folded) threshold.resolve({ kind: "running", reason: "timer" });
+								})
+								.catch(error => {
+									logger.warn("Timer-triggered fold failed", {
+										jobId: foldAdapter.jobId,
+										error: error instanceof Error ? error.message : String(error),
+									});
+									threshold.resolve({ kind: "running", reason: "timer" });
+								});
+						},
+						Math.max(0, thresholdMs ?? 0),
+					);
 		const waiters: Array<
 			Promise<ManagedBashJobCompletion | { kind: "running"; reason: FoldReason } | { kind: "aborted" }>
-		> = [job.completion, threshold.promise];
+		> = [job.completion];
+		if (threshold) waiters.push(threshold.promise);
 		if (backgroundRequest) {
 			waiters.push(backgroundRequest.then(reason => ({ kind: "running" as const, reason })));
 		}
@@ -1333,7 +1342,7 @@ export class BashTool implements AgentTool<BashToolSchema, BashToolDetails> {
 			return await Promise.race(waiters);
 		} finally {
 			stopSteerWatch();
-			clearTimeout(thresholdTimer);
+			if (thresholdTimer !== undefined) clearTimeout(thresholdTimer);
 			if (signal && onAbort) signal.removeEventListener("abort", onAbort);
 		}
 	}
@@ -1894,11 +1903,11 @@ export class BashTool implements AgentTool<BashToolSchema, BashToolDetails> {
 		// longer has to be bypassed to make folding work. A capable ACP session keeps
 		// its terminal contract and still folds.
 		if (!pty && ownedManager && !clientTerminalActive) {
-			// With auto-background off, wait past the command's own timeout so the job only
-			// leaves the foreground on an explicit Ctrl+B fold, never on an auto-background timer.
+			// With auto-background off, the foreground wait has no timer threshold, so the job
+			// only leaves the foreground on an explicit Ctrl+B or steer fold.
 			const autoBackgroundWaitMs = this.#autoBackgroundEnabled
 				? this.#resolveAutoBackgroundWaitMs(timeoutMs)
-				: timeoutMs + 1_000;
+				: undefined;
 			const startBackgrounded = autoBackgroundWaitMs === 0;
 			let managedForegroundSettled = false;
 			const job = this.#startManagedBashJob({
@@ -2600,7 +2609,7 @@ export class BashTool implements AgentTool<BashToolSchema, BashToolDetails> {
 			try {
 				bridgeWait = await this.#waitForManagedBashJob(
 					bridgeHandle,
-					this.#autoBackgroundEnabled ? this.#resolveAutoBackgroundWaitMs(timeoutMs) : timeoutMs + 1_000,
+					this.#autoBackgroundEnabled ? this.#resolveAutoBackgroundWaitMs(timeoutMs) : undefined,
 					signal,
 					bridgeFoldRequest.promise,
 					bridgeFoldAdapter,

package/src/tools/descriptors.test.ts

@@ -736,6 +736,35 @@ describe("tool descriptor compatibility gate", () => {
 			}),
 		).toThrow('Validation failed for tool "ask"');
 	});
+	test("loaded and deferred asks reject missing interview metadata and empty siblings with actionable corrections", async () => {
+		for (const stage of ["topology", "post-topology"] as const) {
+			const session = makeSession({ "tools.discoveryMode": "all" });
+			session.hasUI = true;
+			session.getDeepInterviewAskStage = () => stage;
+			const eager = await BUILTIN_TOOL_DESCRIPTORS.ask.load(session);
+			const lazy = (await createTools(session)).find(tool => tool.name === "ask");
+			if (!eager || !lazy) throw new Error("expected eager and deferred asks");
+			const ordinary = {
+				id: "dot-gjc-topology",
+				question: "Round 0 | Topology confirmation\n\nDoes this scope match?",
+				options: [{ label: "Looks right" }],
+				workflowGate: { stage: "deep-interview", kind: "question" },
+			};
+			for (const [arguments_, correction] of [
+				[{ questions: [ordinary] }, "require structured deepInterview metadata"],
+				[{ questions: [{ ...ordinary, deepInterview: null }] }, "require structured deepInterview metadata"],
+				[{ questions: [ordinary, { id: "metadata", question: "", options: [] }] }, "remove empty extra questions"],
+				[{ questions: [ordinary, ordinary] }, "require exactly one question"],
+			] as const) {
+				for (const tool of [eager, lazy]) {
+					expect(() =>
+						validateToolArguments(tool, { type: "toolCall", id: "incident", name: "ask", arguments: arguments_ }),
+					).toThrow(correction);
+				}
+			}
+		}
+	});
+
 	test("deferred ask validation matches eager AskTool on adversarial contracts", async () => {
 		const session = makeSession({ "tools.discoveryMode": "all" });
 		session.hasUI = true;

package/src/tools/tool-catalog.generated.ts

@@ -310,7 +310,7 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 	"ask": {
 		"name": "ask",
 		"label": "Ask",
-		"description": "Asks user when you need clarification or input during task execution.\n\n<conditions>\n- Multiple approaches exist with significantly different tradeoffs user should weigh\n</conditions>\n\n<instruction>\n- Use `recommended: <index>` to mark default (0-indexed); \" (Recommended)\" added automatically\n- Use `questions` for multiple related questions instead of asking one at a time\n- Set `multi: true` on question to allow multiple selections\n</instruction>\n\n<caution>\n- Provide 2-5 concise, distinct options\n</caution>\n\n<critical>\n- **Default to action.** Resolve ambiguity yourself using repo conventions, existing patterns, and reasonable defaults. Exhaust existing sources (code, configs, docs, history) before asking. Only ask when options have materially different tradeoffs the user must decide.\n- **If multiple choices are acceptable**, pick the most conservative/standard option and proceed; state the choice.\n- **Do NOT include the automatic custom-input control or labels that imitate it** such as \"Other (type your own)\", \"Other (specify)\", \"type your own\", \"custom input\", or \"직접 입력\" — UI automatically adds \"Other (type your own)\" to every question. A plain domain value such as \"Other\" is allowed when it is a genuine answer, not a request for free text. To collect free text, ask for it in the question body and rely on the automatic Other/custom-input control instead of adding a fake option.\n</critical>\n\n<examples>\n# Single question\nquestions: [{\"id\": \"auth_method\", \"question\": \"Which authentication method should this API use?\", \"options\": [{\"label\": \"JWT\"}, {\"label\": \"OAuth2\"}, {\"label\": \"Session cookies\"}], \"recommended\": 0}]\n\n# Multiple questions\nquestions: [{\"id\": \"storage_type\", \"question\": \"Which storage backend?\", \"options\": [{\"label\": \"SQLite\"}, {\"label\": \"PostgreSQL\"}]}, {\"id\": \"auth_method\", \"question\": \"Which auth method?\", \"options\": [{\"label\": \"JWT\"}, {\"label\": \"Session cookies\"}]}]\n</examples>",
+		"description": "Asks user when you need clarification or input during task execution.\n\n<conditions>\n- Multiple approaches exist with significantly different tradeoffs user should weigh\n</conditions>\n\n<instruction>\n- Use `recommended: <index>` to mark default (0-indexed); \" (Recommended)\" added automatically\n- Use `questions` for multiple related questions instead of asking one at a time\n- Set `multi: true` on question to allow multiple selections\n- Every question body must contain meaningful, non-whitespace text. Never add empty `metadata`/`ignore` questions.\n- While deep-interview is in the `interviewing` phase, submit exactly one question with the required `deepInterview` metadata on that question. Round 0 needs `round: 0`, `component: \"review-topology\"`, `dimension: \"topology\"`, `ambiguity`, and `intent_contract` containing the displayed items and affirmative labels. Later questions need a positive round, component, dimension, and ambiguity. Question prose and `workflowGate` alone cannot record an interview answer. After final-spec persistence transitions to `handoff`, next-workflow/approval choices are ordinary asks without interview round metadata.\n</instruction>\n\n<caution>\n- Provide 2-5 concise, distinct options\n</caution>\n\n<critical>\n- **Default to action.** Resolve ambiguity yourself using repo conventions, existing patterns, and reasonable defaults. Exhaust existing sources (code, configs, docs, history) before asking. Only ask when options have materially different tradeoffs the user must decide.\n- **If multiple choices are acceptable**, pick the most conservative/standard option and proceed; state the choice.\n- **Do NOT include the automatic custom-input control or labels that imitate it** such as \"Other (type your own)\", \"Other (specify)\", \"type your own\", \"custom input\", or \"직접 입력\" — UI automatically adds \"Other (type your own)\" to every question. A plain domain value such as \"Other\" is allowed when it is a genuine answer, not a request for free text. To collect free text, ask for it in the question body and rely on the automatic Other/custom-input control instead of adding a fake option.\n</critical>\n\n<examples>\n# Single question\nquestions: [{\"id\": \"auth_method\", \"question\": \"Which authentication method should this API use?\", \"options\": [{\"label\": \"JWT\"}, {\"label\": \"OAuth2\"}, {\"label\": \"Session cookies\"}], \"recommended\": 0}]\n\n# Multiple questions\nquestions: [{\"id\": \"storage_type\", \"question\": \"Which storage backend?\", \"options\": [{\"label\": \"SQLite\"}, {\"label\": \"PostgreSQL\"}]}, {\"id\": \"auth_method\", \"question\": \"Which auth method?\", \"options\": [{\"label\": \"JWT\"}, {\"label\": \"Session cookies\"}]}]\n</examples>",
 		"parameters": {
 			"type": "object",
 			"properties": {
@@ -326,7 +326,8 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 							},
 							"question": {
 								"type": "string",
-								"description": "question text"
+								"minLength": 1,
+								"description": "non-empty question text"
 							},
 							"options": {
 								"type": "array",
@@ -2660,7 +2661,8 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 					"description": "recency filter"
 				},
 				"limit": {
-					"type": "number",
+					"type": "integer",
+					"minimum": 1,
 					"description": "max results"
 				},
 				"max_tokens": {
@@ -2672,7 +2674,8 @@ export const TOOL_CATALOG: Readonly<Record<string, ToolCatalogEntry>> = {
 					"description": "sampling temperature"
 				},
 				"num_search_results": {
-					"type": "number",
+					"type": "integer",
+					"minimum": 1,
 					"description": "number of search results"
 				},
 				"xai_search_mode": {

package/src/web/search/index.ts

@@ -44,10 +44,10 @@ import {
 export const webSearchSchema = z.object({
 	query: z.string().describe("search query"),
 	recency: z.enum(["day", "week", "month", "year"]).describe("recency filter").optional(),
-	limit: z.number().describe("max results").optional(),
+	limit: z.number().int().min(1).describe("max results").optional(),
 	max_tokens: z.number().describe("max output tokens").optional(),
 	temperature: z.number().describe("sampling temperature").optional(),
-	num_search_results: z.number().describe("number of search results").optional(),
+	num_search_results: z.number().int().min(1).describe("number of search results").optional(),
 	xai_search_mode: z
 		.enum(["web", "x", "web_and_x"])
 		.describe("xAI only: use web_search, x_search, or both")