From c8230be4a76bdc8a2b2b4f59dc3766da3d358ba7 Mon Sep 17 00:00:00 2001
From: can1357 <me@can.ac>
Date: Fri, 20 Feb 2026 11:00:38 +0100
Subject: [PATCH] fix(coding-agent): fail fast when RPC startup exits early

---
 packages/ai/CHANGELOG.md                        | 12 +++++-------
 .../coding-agent/src/modes/rpc/rpc-client.ts    | 14 +++++---------
 .../coding-agent/test/rpc-client.start.test.ts  | 17 +++++++++++++++++
 3 files changed, 27 insertions(+), 16 deletions(-)
 create mode 100644 packages/coding-agent/test/rpc-client.start.test.ts

diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
index 0167d8ab1..d5b42d658 100644
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -1,17 +1,15 @@
 # Changelog
 
 ## [Unreleased]
-
-### Changed
-
-- Implemented time-budget-based retry strategy for Google Gemini rate limits (429 errors), allowing retries within a 5-minute window instead of a fixed attempt cap
-- Set default `maxRetries` to 5 for Anthropic, Azure OpenAI, and OpenAI client configurations to improve resilience against transient failures
-
 ### Fixed
 
 - Improved error messages for OAuth token refresh failures by including detailed error information from the provider
 - Separated rate limit and usage limit error handling to provide distinct user-friendly messages for ChatGPT rate limits vs subscription usage limits
-- Changed 429 retry strategy for OpenAI Codex to use a 5-minute time budget when the server provides a retry delay, instead of a fixed 5-attempt cap
+
+### Changed
+
+- Increased SDK retry attempts to 5 for OpenAI, Azure OpenAI, and Anthropic clients (was SDK default of 2)
+- Changed 429 retry strategy for OpenAI Codex and Google Gemini CLI to use a 5-minute time budget when the server provides a retry delay, instead of a fixed attempt cap
 
 ## [12.14.0] - 2026-02-19
 ### Added
diff --git a/packages/coding-agent/src/modes/rpc/rpc-client.ts b/packages/coding-agent/src/modes/rpc/rpc-client.ts
index 6a47be4a0..7596cb2c2 100644
--- a/packages/coding-agent/src/modes/rpc/rpc-client.ts
+++ b/packages/coding-agent/src/modes/rpc/rpc-client.ts
@@ -134,15 +134,11 @@ export class RpcClient {
 		// Wait a moment for process to initialize
 		await Bun.sleep(100);
 
-		try {
-			const exitCode = await Promise.race([this.#process.exited, Bun.sleep(500).then(() => null)]);
-			if (exitCode !== null) {
-				throw new Error(
-					`Agent process exited immediately with code ${exitCode}. Stderr: ${this.#process.peekStderr()}`,
-				);
-			}
-		} catch {
-			// Process still running, which is what we want
+		const exitCode = await Promise.race([this.#process.exited, Bun.sleep(500).then(() => null)]);
+		if (exitCode !== null) {
+			throw new Error(
+				`Agent process exited immediately with code ${exitCode}. Stderr: ${this.#process.peekStderr()}`,
+			);
 		}
 	}
 
diff --git a/packages/coding-agent/test/rpc-client.start.test.ts b/packages/coding-agent/test/rpc-client.start.test.ts
new file mode 100644
index 000000000..7009b93ef
--- /dev/null
+++ b/packages/coding-agent/test/rpc-client.start.test.ts
@@ -0,0 +1,17 @@
+import { describe, expect, test } from "bun:test";
+import * as path from "node:path";
+import { RpcClient } from "@oh-my-pi/pi-coding-agent/modes/rpc/rpc-client";
+
+describe("RpcClient.start", () => {
+	test("rejects when RPC process exits immediately", async () => {
+		using client = new RpcClient({
+			cliPath: path.join(import.meta.dir, "..", "src", "cli.ts"),
+			cwd: path.join(import.meta.dir, ".."),
+			provider: "__missing_provider__",
+			model: "claude-sonnet-4-5",
+			env: { PI_NO_TITLE: "1" },
+		});
+
+		await expect(client.start()).rejects.toThrow('Unknown provider "__missing_provider__"');
+	});
+});