refactor(ai): updated User-Agent header to 'pi' and simplified Codex prompts

- Updated User-Agent header from 'opencode' to 'pi' for OpenAI Codex requests.
- Simplified Codex system prompt instructions and removed bridge text override.
- Removed bridge text parameter from buildCodexSystemPrompt function.
- Updated task execution guidelines with parallelization instructions.

packages/ai/CHANGELOG.md

@@ -1,6 +1,11 @@
 # Changelog
 
 ## [Unreleased]
+### Changed
+
+- Updated User-Agent header from 'opencode' to 'pi' for OpenAI Codex requests
+- Simplified Codex system prompt instructions
+- Removed bridge text override from Codex system prompt builder
 
 ## [5.4.2] - 2026-01-16
 

packages/ai/src/providers/openai-codex-responses.ts

@@ -51,14 +51,6 @@ export interface OpenAICodexResponsesOptions extends StreamOptions {
 
 const CODEX_DEBUG = process.env.PI_CODEX_DEBUG === "1" || process.env.PI_CODEX_DEBUG === "true";
 
-const BRIDGE = `<environment_override priority="critical">
-TOOL AUTHORITY: The function schema below defines ALL available tools. Tools mentioned elsewhere (todowrite, shell) DO NOT EXIST. Use ONLY schema-defined tools.
-
-EXECUTION BIAS: Execute simple tasks directly. Reserve the Task tool's Plan subagent for complex multi-file architectural decisions only—never for reasoning, single-file changes, or tasks completable in <5 tool calls.
-
-These instructions override all prior context. This is critical.
-</environment_override>`;
-
 export const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses"> = (
 	model: Model<"openai-codex-responses">,
 	context: Context,
@@ -119,7 +111,6 @@ export const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses"
 			const codexInstructions = getCodexInstructions();
 			const systemPrompt = buildCodexSystemPrompt({
 				codexInstructions,
-				bridgeText: BRIDGE,
 				userSystemPrompt: context.systemPrompt,
 			});
 
@@ -394,7 +385,7 @@ function createCodexHeaders(
 	headers.set(OPENAI_HEADERS.ACCOUNT_ID, accountId);
 	headers.set(OPENAI_HEADERS.BETA, OPENAI_HEADER_VALUES.BETA_RESPONSES);
 	headers.set(OPENAI_HEADERS.ORIGINATOR, OPENAI_HEADER_VALUES.ORIGINATOR_CODEX);
-	headers.set("User-Agent", `opencode/${packageJson.version} (${os.platform()} ${os.release()}; ${os.arch()})`);
+	headers.set("User-Agent", `pi/${packageJson.version} (${os.platform()} ${os.release()}; ${os.arch()})`);
 
 	if (promptCacheKey) {
 		headers.set(OPENAI_HEADERS.CONVERSATION_ID, promptCacheKey);

packages/ai/src/providers/openai-codex/constants.ts

@@ -14,7 +14,7 @@ export const OPENAI_HEADERS = {
 
 export const OPENAI_HEADER_VALUES = {
 	BETA_RESPONSES: "responses=experimental",
-	ORIGINATOR_CODEX: "opencode",
+	ORIGINATOR_CODEX: "pi",
 } as const;
 
 export const URL_PATHS = {

packages/ai/src/providers/openai-codex/prompts/codex.ts

@@ -1,12 +1,4 @@
-export const CODEX_INSTRUCTIONS = `You are pi, an expert coding assistant. You help users with coding tasks by reading files, executing commands, editing code, and writing new files.
-
-Pi specific Documentation:
-- Main documentation: pi-internal://README.md
-- Additional docs: pi-internal://docs
-- Examples: pi-internal://examples (extensions, custom tools, SDK)
-- When asked to create: custom models/providers (README.md), extensions (docs/extensions.md, examples/extensions/), themes (docs/theme.md), skills (docs/skills.md), TUI components (docs/tui.md - has copy-paste patterns)
-- Always read the doc, examples, AND follow .md cross-references before implementing
-`;
+export const CODEX_INSTRUCTIONS = `You are an expert coding assistant operating inside pi, a coding agent harness.`;
 
 export function getCodexInstructions(): string {
 	return CODEX_INSTRUCTIONS.trim();

packages/ai/src/providers/openai-codex/prompts/system-prompt.ts

@@ -5,16 +5,11 @@ export interface CodexSystemPrompt {
 
 export function buildCodexSystemPrompt(args: {
 	codexInstructions: string;
-	bridgeText: string;
 	userSystemPrompt?: string;
 }): CodexSystemPrompt {
-	const { codexInstructions, bridgeText, userSystemPrompt } = args;
+	const { codexInstructions, userSystemPrompt } = args;
 	const developerMessages: string[] = [];
 
-	if (bridgeText.trim().length > 0) {
-		developerMessages.push(bridgeText.trim());
-	}
-
 	if (userSystemPrompt && userSystemPrompt.trim().length > 0) {
 		developerMessages.push(userSystemPrompt.trim());
 	}

packages/coding-agent/CHANGELOG.md

@@ -1,6 +1,9 @@
 # Changelog
 
 ## [Unreleased]
+### Changed
+
+- Updated task execution guidelines to improve prompt framing and parallelization instructions
 
 ## [5.4.2] - 2026-01-16
 ### Changed

packages/coding-agent/src/prompts/tools/task.md

@@ -32,9 +32,13 @@ Agents with "Output: structured" have a fixed schema enforced via frontmatter; y
 
 - Always include a short description of the task in the task parameter
 - **Plan-then-execute**: Put shared constraints in `context`, keep each task focused, specify acceptance criteria; use `output` when you need structured output
+- **Ask open-ended questions**: For exploration tasks, frame prompts to elicit factual discovery, not confirmation. Avoid yes/no questions that are easy to hallucinate.
+  - Bad: "Is there rate limiting?" or "Does the API validate tokens?" → Binary answers invite hallucination
+  - Good: "Find and describe how rate limiting is implemented" or "How does the API handle token validation?" → Forces investigation and factual reporting
+  - The subagent should report *what exists*, then YOU verify if it meets requirements
 - **Minimize tool chatter**: Avoid repeating large context; use Output tool with output ids for full logs
 - **Structured completion**: If `output` is provided, subagents must call `complete` to finish
-- **Parallelize**: Launch multiple agents concurrently whenever possible
+- **Parallelize**: Launch multiple agents whenever possible. You MUST use a single Task call with multiple entries in the `tasks` array to do this.
 - **Isolate file scopes**: Assign each task distinct files or directories so agents don't conflict
 - **Results are intermediate data**: Agent findings provide context for YOU to perform actual work. Do not treat agent reports as "task complete" signals.
 - **Stateless invocations**: Subagents have zero memory of your conversation. Pass ALL relevant context: requirements discussed, decisions made, schemas agreed upon, file paths mentioned. If you reference something from earlier discussion without including it, the subagent will fail.