@gajae-code/coding-agent 0.18.12 → 0.19.0

package/src/utils/title-generator.ts

@@ -1,5 +1,5 @@
 /**
- * Generate session titles using a smol, fast model.
+ * Generate session titles with the `default` model role.
  */
 import * as path from "node:path";
 
@@ -10,6 +10,7 @@ import type { ModelRegistry } from "../config/model-registry";
 import { resolveRoleSelection } from "../config/model-resolver";
 import type { Settings } from "../config/settings";
 import titleSystemPrompt from "../prompts/system/title-system.md" with { type: "text" };
+import { collectEnvSecrets, createSecretObfuscator, loadSecrets } from "../secrets";
 import { syncCmuxWorkspaceTitle } from "./cmux-workspace";
 import { syncHerdrPaneTitle } from "./herdr-pane";
 
@@ -37,8 +38,13 @@ const MAX_TITLE_WORDS = 12;
  * Build a bounded conversation digest for title regeneration. Regeneration
  * reads several messages so a throwaway first prompt can be corrected, while
  * the automatic first-message path intentionally uses only its initial input.
+ * `obfuscateText`, when set, runs on each kept message's full text before the
+ * per-message character cut so a secret cannot be split by that cut.
  */
-export function buildConversationTitleInput(messages: readonly AgentMessage[]): string | undefined {
+export function buildConversationTitleInput(
+	messages: readonly AgentMessage[],
+	obfuscateText?: (text: string) => string,
+): string | undefined {
 	const userMessages: string[] = [];
 	for (const message of messages) {
 		if (message.role !== "user") continue;
@@ -51,7 +57,10 @@ export function buildConversationTitleInput(messages: readonly AgentMessage[]):
 						.map(block => block.text)
 						.join("\n");
 		const trimmed = text.trim();
-		if (trimmed) userMessages.push(trimmed);
+		if (!trimmed) continue;
+		// Emptiness uses the trimmed text. With an obfuscator, keep the original
+		// so a secret that includes edge whitespace still matches, then trim.
+		userMessages.push(obfuscateText ? text : trimmed);
 	}
 
 	if (userMessages.length === 0) return undefined;
@@ -60,13 +69,19 @@ export function buildConversationTitleInput(messages: readonly AgentMessage[]):
 		userMessages.length > MAX_TITLE_CONVERSATION_MESSAGES
 			? [userMessages[0]!, ...userMessages.slice(-(MAX_TITLE_CONVERSATION_MESSAGES - 1))]
 			: userMessages;
+	// Obfuscate the full kept text first. The 400-character cut below, and the
+	// later 2,000-character cap, then keep their existing limits.
+	const preparedMessages = obfuscateText
+		? keptMessages.map(message => obfuscateText(message).trim()).filter(message => message.length > 0)
+		: keptMessages;
+	if (preparedMessages.length === 0) return undefined;
 	// Reserve separator and ellipsis room so the digest is already within the
 	// generator's MAX_INPUT_CHARS bound; its existing truncation is a no-op here.
 	const maxMessageChars = Math.min(
 		MAX_TITLE_MESSAGE_CHARS,
-		Math.floor((MAX_INPUT_CHARS - (keptMessages.length - 1) - keptMessages.length) / keptMessages.length),
+		Math.floor((MAX_INPUT_CHARS - (preparedMessages.length - 1) - preparedMessages.length) / preparedMessages.length),
 	);
-	return keptMessages
+	return preparedMessages
 		.map(message => (message.length > maxMessageChars ? `${message.slice(0, maxMessageChars)}…` : message))
 		.join("\n");
 }
@@ -87,8 +102,13 @@ const setTitleTool: Tool = {
 	},
 };
 
-function getTitleModel(registry: ModelRegistry, settings: Settings, currentModel?: Model<Api>): Model<Api> | undefined {
-	const availableModels = registry.getAvailable();
+function getTitleModel(
+	registry: ModelRegistry,
+	settings: Settings,
+	currentModel?: Model<Api>,
+	credentialSessionId?: string,
+): Model<Api> | undefined {
+	const availableModels = registry.getAvailable(credentialSessionId);
 	if (availableModels.length === 0) return undefined;
 
 	const titleModel = resolveRoleSelection(["default"], settings, availableModels, registry)?.model;
@@ -99,6 +119,32 @@ function getTitleModel(registry: ModelRegistry, settings: Settings, currentModel
 	return undefined;
 }
 
+/**
+ * Obfuscate the full title input, then truncate, then wrap.
+ * A secret that crosses the character cap is replaced before the cut, and the
+ * `<user-message>` fence stays outside the truncated body.
+ */
+export function prepareTitleOutboundText(firstMessage: string, obfuscateText?: (text: string) => string): string {
+	const source = obfuscateText ? obfuscateText(firstMessage) : firstMessage;
+	const truncated = source.length > MAX_INPUT_CHARS ? `${source.slice(0, MAX_INPUT_CHARS)}…` : source;
+	return `<user-message>\n${truncated}\n</user-message>`;
+}
+
+export async function titleInputObfuscator(
+	settings: Settings,
+	activeCwd?: string,
+): Promise<((text: string) => string) | undefined> {
+	const reader = settings as Partial<Pick<Settings, "get" | "getCwd" | "getAgentDir">>;
+	if (typeof reader.get !== "function" || reader.get("secrets.enabled") !== true) return undefined;
+	const cwd = activeCwd ?? (typeof reader.getCwd === "function" ? reader.getCwd() : process.cwd());
+	const agentDir = typeof reader.getAgentDir === "function" ? reader.getAgentDir() : cwd;
+	const entries = [...collectEnvSecrets(), ...(await loadSecrets(cwd, agentDir))];
+	if (entries.length === 0) return undefined;
+	const obfuscator = createSecretObfuscator(entries);
+	if (!obfuscator.hasSecrets()) return undefined;
+	return text => obfuscator.obfuscate(text);
+}
+
 /**
  * Generate a title for a session based on the provided user-message input.
  * Automatic titles pass the first user message; regeneration passes a bounded
@@ -106,13 +152,14 @@ function getTitleModel(registry: ModelRegistry, settings: Settings, currentModel
  *
  * @param firstMessage The first user message or bounded conversation digest
  * @param registry Model registry
- * @param settings Settings used to resolve the smol role
+ * @param settings Settings used to resolve the `default` model role
  * @param sessionId Optional session id for sticky API key selection
  * @param currentModel Current model (used to derive title model)
  * @param metadataResolver Optional resolver evaluated after credential selection
  *   to produce request metadata (e.g. user_id for session attribution). Using a
  *   resolver instead of a pre-evaluated value ensures the metadata's account_uuid
  *   reflects the credential actually selected for this request.
+ * @param activeCwd Active session cwd for project-scoped secret lookup.
  */
 export async function generateSessionTitle(
 	firstMessage: string,
@@ -121,23 +168,17 @@ export async function generateSessionTitle(
 	sessionId?: string,
 	currentModel?: Model<Api>,
 	metadataResolver?: (provider: string) => Record<string, unknown> | undefined,
+	activeCwd?: string,
 ): Promise<string | null> {
-	const model = getTitleModel(registry, settings, currentModel);
+	const model = getTitleModel(registry, settings, currentModel, sessionId);
 	if (!model) {
 		logger.debug("title-generator: no title model found");
 		return null;
 	}
 
-	// Truncate message if too long
-	const truncatedMessage =
-		firstMessage.length > MAX_INPUT_CHARS ? `${firstMessage.slice(0, MAX_INPUT_CHARS)}…` : firstMessage;
-	const userMessage = `<user-message>
-${truncatedMessage}
-</user-message>`;
-
 	const apiKey = await registry.getApiKey(model, sessionId);
 	if (!apiKey) {
-		logger.debug("title-generator: no API key for smol model", {
+		logger.debug("title-generator: no API key for title model", {
 			provider: model.provider,
 			id: model.id,
 		});
@@ -147,6 +188,7 @@ ${truncatedMessage}
 	// request is already recorded; metadataResolver can then return the correct
 	// account_uuid rather than the snapshot-at-call-site value.
 	const metadata = metadataResolver?.(model.provider);
+	const userMessage = prepareTitleOutboundText(firstMessage, await titleInputObfuscator(settings, activeCwd));
 
 	// Title generation is a 3-6 word task, but some reasoning backends ignore
 	// disableReasoning. Keep the normal cheap budget for non-reasoning models
@@ -187,7 +229,10 @@ ${truncatedMessage}
 			logger.debug("title-generator: response error", {
 				model: request.model,
 				stopReason: response.stopReason,
-				errorMessage: response.errorMessage,
+				errorMessage:
+					typeof response.errorMessage === "string"
+						? redactCrashSecrets(response.errorMessage)
+						: response.errorMessage,
 			});
 			return null;
 		}
@@ -196,7 +241,7 @@ ${truncatedMessage}
 
 		logger.debug("title-generator: response", {
 			model: request.model,
-			title,
+			title: redactCrashSecrets(title),
 			usage: response.usage,
 			stopReason: response.stopReason,
 		});
@@ -209,7 +254,7 @@ ${truncatedMessage}
 	} catch (err) {
 		logger.debug("title-generator: error", {
 			model: request.model,
-			error: err instanceof Error ? err.message : String(err),
+			error: redactCrashSecrets(err instanceof Error ? err.message : String(err)),
 		});
 		return null;
 	}

package/src/web/search/index.ts

@@ -15,6 +15,7 @@ import webSearchSystemPrompt from "../../prompts/system/web-search.md" with { ty
 import webSearchDescription from "../../prompts/tools/web-search.md" with { type: "text" };
 import { discoverAuthStorage } from "../../sdk";
 import type { ToolSession } from "../../tools";
+import { wrapUntrustedContent } from "../../tools/fetch";
 import { formatAge } from "../../tools/render-utils";
 import { throwIfAborted } from "../../tools/tool-errors";
 import {
@@ -157,7 +158,7 @@ export function formatSearchResponseForLlm(response: SearchResponse): string {
 		}
 	}
 
-	return `<untrusted-content>\n${parts.join("\n").replace(/<\/untrusted-content>/gi, "&lt;/untrusted-content>")}\n</untrusted-content>`;
+	return wrapUntrustedContent(parts.join("\n"));
 }
 
 interface ExecuteSearchOptions {

package/src/web/search/providers/codex.ts

@@ -7,8 +7,9 @@
  * SQLite store, never POSTs the broker sentinel to an OpenAI token endpoint.
  */
 import * as os from "node:os";
-import { type AuthStorage, getBundledModels } from "@gajae-code/ai/core";
+import type { AuthStorage } from "@gajae-code/ai/core";
 import { OPENAI_HEADER_VALUES } from "@gajae-code/ai/providers/openai-codex/constants";
+import { fetchCodexModels } from "@gajae-code/ai/utils/discovery/codex";
 import { decodeJwt } from "@gajae-code/ai/utils/oauth/openai-codex";
 import { $env, readSseJson, sanitizeHeaderComponent } from "@gajae-code/utils";
 import packageJson from "../../../../package.json" with { type: "json" };
@@ -21,17 +22,9 @@ import { classifyProviderHttpError, withHardTimeout } from "./utils";
 
 const CODEX_BASE_URL = "https://chatgpt.com/backend-api";
 const CODEX_RESPONSES_PATH = "/codex/responses";
-const FALLBACK_MODEL = "gpt-5.5";
-const DEFAULT_MODEL_PREFERENCES = [
-	"gpt-5.5",
-	"gpt-5.4",
-	"gpt-5-codex",
-	"gpt-5",
-	"gpt-5.3-codex",
-	"gpt-5.2-codex",
-	"gpt-5.1-codex",
-	"gpt-5-codex-mini",
-];
+// Reviewed search candidates, additionally filtered by current account discovery.
+// Upstream Codex web-search metadata informs this policy, not bundle presence.
+const DEFAULT_MODEL_PREFERENCES = ["gpt-6.1-sol", "gpt-6-luna"];
 const JWT_CLAIM_PATH = "https://api.openai.com/auth";
 const DEFAULT_INSTRUCTIONS =
 	"You are a helpful assistant with web search capabilities. Search the web to answer the user's question accurately and cite your sources.";
@@ -41,21 +34,26 @@ function getConfiguredModel(): string | undefined {
 	return configuredModel ? configuredModel : undefined;
 }
 
-function getDefaultModelCandidates(): string[] {
-	const bundledModels = getBundledModels("openai-codex");
-	const bundledIds = new Set(bundledModels.map(model => model.id));
-	const candidates = DEFAULT_MODEL_PREFERENCES.filter(modelId => bundledIds.has(modelId));
-
-	if (candidates.length > 0) {
-		return candidates;
+async function getDefaultModelCandidates(
+	auth: { accessToken: string; accountId: string },
+	signal: AbortSignal | undefined,
+): Promise<string[]> {
+	signal?.throwIfAborted();
+	const discovery = await fetchCodexModels({ ...auth, signal: withHardTimeout(signal, "llm") });
+	signal?.throwIfAborted();
+	if (!discovery) {
+		throw new SearchProviderError("codex", "Codex model discovery unavailable", 503);
 	}
-
-	const nonMini = bundledModels.find(model => !model.id.includes("mini") && !model.id.includes("spark"));
-	if (nonMini) {
-		return [nonMini.id];
+	const advertisedIds = new Set(discovery.models.map(model => model.id));
+	const candidates = DEFAULT_MODEL_PREFERENCES.filter(modelId => advertisedIds.has(modelId));
+	if (candidates.length === 0) {
+		throw new SearchProviderError(
+			"codex",
+			"No preferred Codex web-search models are advertised for this account",
+			503,
+		);
 	}
-
-	return bundledModels[0]?.id ? [bundledModels[0].id] : [FALLBACK_MODEL];
+	return candidates;
 }
 
 function shouldRetryWithNextDefaultModel(error: unknown): boolean {
@@ -337,12 +335,11 @@ async function callCodexSearch(
  * Executes a web search using OpenAI code provider's built-in web search tool.
  *
  * Default-model behavior:
- * - If `PI_CODEX_WEB_SEARCH_MODEL` is set, use it exactly once and surface any
- *   upstream error verbatim.
- * - Otherwise prefer ChatGPT-account-safe bundled defaults (GPT-5.5, GPT-5.4,
- *   GPT-5 code backend, …) and retry the next candidate only when OpenAI code backend returns the
- *   known 400 "model is not supported" family. This avoids selecting
- *   `gpt-5-codex-mini` first on ChatGPT accounts, which OpenAI rejects.
+ * - Use `GJC_OPENAI_CODE_WEB_SEARCH_MODEL` (then `PI_CODEX_WEB_SEARCH_MODEL`)
+ *   exactly once when configured and surface upstream errors without model retries.
+ * - Otherwise intersect current account discovery with GPT-6.1 Sol, then GPT-6 Luna.
+ *   Fail closed when discovery is unavailable or neither candidate is advertised.
+ *   Retry the next advertised candidate only for the known unsupported-model 400 family.
  */
 export async function searchCodex(params: SearchParams): Promise<SearchResponse> {
 	const auth = await findCodexAuth(params.authStorage, params.sessionId, params.signal);
@@ -353,7 +350,7 @@ export async function searchCodex(params: SearchParams): Promise<SearchResponse>
 	}
 
 	const configuredModel = getConfiguredModel();
-	const modelCandidates = configuredModel ? [configuredModel] : getDefaultModelCandidates();
+	const modelCandidates = configuredModel ? [configuredModel] : await getDefaultModelCandidates(auth, params.signal);
 
 	let result:
 		| {