@henryqw/pi-multi-codex 1.0.1 → 1.0.3

package/README.md

@@ -1,13 +1,9 @@
 # `@henryqw/pi-multi-codex`
 
-Add multiple ChatGPT Codex OAuth accounts and start new Pi work on the slot with the most weekly quota.
+Add multiple ChatGPT Codex OAuth accounts and start Pi work on the eligible slot with the most weekly quota. The extension switches away from active five-hour blocks and can retry an HTTP 429 on another eligible slot.
 
 ![Pi showing Codex account quotas and the active footer slot](./example.png)
-
-## Why
-
-- **Created for**: Pi users who work across more than one Codex subscription.
-- **Advantage**: Quota-aware routing starts with the eligible slot that has the most remaining seven-day quota. With automatic switching enabled, Pi can retry an HTTP 429 response on another eligible slot.
+![Flowchart separating fresh-quota startup ranking from broader HTTP 429 failover](./docs/codex-routing-flow.svg)
 
 ## Install
 
@@ -15,7 +11,7 @@ Add multiple ChatGPT Codex OAuth accounts and start new Pi work on the slot with
 pi install npm:@henryqw/pi-multi-codex
 ```
 
-## With
+## Works with
 
 | Package | Why |
 | --- | --- |
@@ -37,17 +33,17 @@ Run `/login` and authenticate `OpenAI Codex` for slot 1 first. Run `/codex-add`,
 
 A numbered slot is one Codex account position in Pi.
 
-![Flowchart separating fresh-quota startup ranking from broader HTTP 429 failover](./docs/codex-routing-flow.svg)
+## Flow
 
-- The extension reads `auth.json`. It never writes or refreshes credentials.
 - Before the first agent start, only fresh quota snapshots enter startup ranking. The slot with the most seven-day quota wins.
+- An explicit model selection always wins for the next agent start.
+- Before later agent starts, a known active five-hour block switches to the eligible fresh slot with the most seven-day quota.
 - After HTTP 429, failover can use authenticated, registered, scope-allowed, untried slots with stale or missing quota.
 - Failover skips known active five-hour blocks. It ranks fresh known quota first, then unranked slots by slot number.
 - Routing preserves the model ID.
 - During one agent run, each eligible slot is tried at most once after HTTP 429 responses.
 - Automatic retry stops when no untried eligible slot remains.
 - The footer shows the active slot's fresh quota or five-hour block.
-- Scoped sessions can switch only to exact scoped aliases.
 
 ## Config
 
@@ -61,4 +57,14 @@ Automatic HTTP 429 switching is on by default. To disable it, create `~/.pi/agen
 
 The config must contain only `autoSwitchOn429` as a boolean. Invalid config is preserved and disables automatic switching.
 
-Generated credential-free quota cache: `~/.pi/agent/config/pi-multi-codex/usage.json`. The extension maintains it.
+## State and storage
+
+The extension maintains a generated, credential-free quota cache at `~/.pi/agent/config/pi-multi-codex/usage.json`.
+
+## Data, cost, and privacy
+
+The extension reads `auth.json`. It never writes or refreshes credentials.
+
+## Limits and recovery
+
+Scoped sessions can switch only to exact scoped aliases.

package/extensions/multi-codex.ts

@@ -835,6 +835,8 @@ export default function multiCodex(pi: ExtensionAPI): void {
 	let sessionContext: ExtensionContext | undefined;
 	let automaticOpen = false;
 	let automaticCandidate: Model<any> | undefined;
+	let preserveModelChoice = false;
+	let selectingAutomatically = false;
 	let autoSwitchOn429 = true;
 	let configWarned = false;
 	let providerRequest: { provider?: string; status?: number } | undefined;
@@ -921,8 +923,17 @@ export default function multiCodex(pi: ExtensionAPI): void {
 		if (ctx) ctx.ui.setStatus?.("pi-multi-codex", footerText(ctx));
 	};
 
+	const setModelAutomatically = async (model: Model<any>): Promise<boolean> => {
+		selectingAutomatically = true;
+		try {
+			return await pi.setModel(model);
+		} finally {
+			selectingAutomatically = false;
+		}
+	};
+
 	const updatePendingCandidate = (ctx: ExtensionContext): void => {
-		if (!automaticOpen || !ctx.model || !isManagedProvider(ctx.model.provider)) return;
+		if (!ctx.model || !isManagedProvider(ctx.model.provider)) return;
 		const slot = selectFreshSlot(ctx, ctx.model);
 		automaticCandidate = slot ? { ...ctx.model, provider: providerForSlot(slot) } : ctx.model;
 	};
@@ -941,6 +952,7 @@ export default function multiCodex(pi: ExtensionAPI): void {
 		sessionContext = ctx;
 		triedSlots.clear();
 		providerRequest = undefined;
+		preserveModelChoice = false;
 		try {
 			autoSwitchOn429 = configStore.loadSync().value.autoSwitchOn429;
 		} catch {
@@ -962,14 +974,17 @@ export default function multiCodex(pi: ExtensionAPI): void {
 	pi.on("before_agent_start", async (_event, ctx) => {
 		triedSlots.clear();
 		providerRequest = undefined;
-		if (!automaticOpen) return;
 		const model = ctx.model;
+		const slot = slotForProvider(model?.provider ?? "");
+		const identity = slot ? currentIdentity(slot) : undefined;
+		const snapshot = slot ? quota.snapshot(slot) : undefined;
+		if (!automaticOpen && (preserveModelChoice || !identity || snapshot?.accountHash !== identity.accountHash || !isFiveHourLimited(snapshot, Date.now()))) return;
 		updatePendingCandidate(ctx);
 		// Close before await: no timer, refresh, queued turn, or provider event can route later.
 		automaticOpen = false;
 		const candidate = automaticCandidate;
 		if (!model || !candidate || !isManagedProvider(model.provider) || candidate.id !== model.id || candidate.provider === model.provider) return;
-		await pi.setModel(candidate);
+		await setModelAutomatically(candidate);
 		updateFooter(ctx);
 	});
 
@@ -1002,7 +1017,7 @@ export default function multiCodex(pi: ExtensionAPI): void {
 		syncSlots();
 		for (const slot of failoverSlots(ctx, model)) {
 			const provider = providerForSlot(slot);
-			if (!await pi.setModel({ ...model, provider })) continue;
+			if (!await setModelAutomatically({ ...model, provider })) continue;
 			ctx.ui.notify(`Codex #${failedSlot} returned HTTP 429. Retrying with Codex #${slot}.`, "warning");
 			updateFooter(ctx);
 			return { message: { ...event.message, errorMessage: `HTTP 429: ${event.message.errorMessage ?? "Codex request was rate limited."}` } };
@@ -1018,6 +1033,7 @@ export default function multiCodex(pi: ExtensionAPI): void {
 
 	pi.on("agent_start", (_event, ctx) => {
 		automaticOpen = false;
+		preserveModelChoice = false;
 		if (!sessionHasAgentWork(ctx)) pi.appendEntry(AGENT_STARTED_ENTRY);
 	});
 	pi.on("agent_settled", () => {
@@ -1032,8 +1048,9 @@ export default function multiCodex(pi: ExtensionAPI): void {
 		quota.stop();
 	});
 	pi.on("model_select", (_event, ctx) => {
-		// Explicit selector choice wins over startup routing.
+		// Explicit selector choice wins over routing at the next agent boundary.
 		automaticOpen = false;
+		if (!selectingAutomatically) preserveModelChoice = true;
 		syncSlots();
 		quota.requestRefresh();
 		updateFooter(ctx);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@henryqw/pi-multi-codex",
-  "version": "1.0.1",
+  "version": "1.0.3",
   "description": "Use multiple ChatGPT Codex OAuth accounts in Pi.",
   "keywords": [
     "pi-package",