@oh-my-pi/pi-ai 18.0.1 → 18.0.3

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## [Unreleased]
 
+## [18.0.3] - 2026-08-23
+
+### Fixed
+
+- Fixed a Fireworks-hosted model aborting mid-generation with an HTTP 400 `Floating point NaN (not-a-number) is detected in generation` killing the turn instead of retrying; this model-side numerical fault is now classified transient and retried, matching the existing treatment of Copilot fleet-skew 400s ([#9458](https://github.com/can1357/oh-my-pi/issues/9458)).
+
+## [18.0.2] - 2026-08-23
+
+### Fixed
+
+- Fixed OpenAI-compatible completions hosts that stream content then terminate with the `[DONE]` sentinel while omitting (or `null`ing) `finish_reason` failing every turn with `OpenAI completions stream closed before a finish_reason was received`; a `[DONE]`-terminated stream now finalizes as a clean stop and only a genuine transport EOF (no `[DONE]`, no finish reason) surfaces the incomplete-stream error ([#9433](https://github.com/can1357/oh-my-pi/issues/9433)).
+
 ## [18.0.1] - 2026-08-23
 
 ### Changed
@@ -29,7 +41,6 @@
 
 - Added Amazon Bedrock Converse guardrail configuration with provider-scoped identifier, version, and trace settings.
 
-### Changed
 ## [18.0.0] - 2026-08-22
 
 ### Added

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "18.0.1",
+	"version": "18.0.3",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://omp.sh",
 	"author": "Stencil Labs, Inc.",
@@ -37,10 +37,10 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.0.1",
-		"@oh-my-pi/pi-catalog": "18.0.1",
-		"@oh-my-pi/pi-utils": "18.0.1",
-		"@oh-my-pi/pi-wire": "18.0.1"
+		"@oh-my-pi/omptype": "18.0.3",
+		"@oh-my-pi/pi-catalog": "18.0.3",
+		"@oh-my-pi/pi-utils": "18.0.3",
+		"@oh-my-pi/pi-wire": "18.0.3"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/error/flags.ts

@@ -165,6 +165,15 @@ const COPILOT_TRANSIENT_MODEL_CODES: Record<string, true> = {
 	model_not_supported: true,
 };
 const COPILOT_TRANSIENT_MODEL_PATTERN = /model_not_supported/i;
+// Fireworks (and other OpenAI-compat backends) can abort mid-generation with an
+// HTTP 400 `invalid_request_error` whose body reports a model-side numerical
+// fault: "Floating point NaN (not-a-number) is detected in generation". Despite
+// the request-validation wrapper this is a decode-time logits overflow, not a
+// bad request — a byte-identical replay of the same payload succeeds. The fault
+// fires before any content is emitted, so retry is replay-safe. Kept narrow
+// (both the NaN wording and "detected in generation") so it cannot swallow
+// genuine request-validation 400s.
+const GENERATION_NAN_PATTERN = /floating[ _-]?point nan\b.*\bdetected in generation/is;
 // Anthropic strict-tool grammar too large / schema too complex (400 invalid_request_error).
 // Feature-gated deployments (Azure Foundry, Baseten, …) reject `strict: true`
 // tools outright when the hosted model lacks structured outputs, e.g.
@@ -435,6 +444,9 @@ function classifyText(
 
 		// Copilot's `model_not_supported` fleet-skew rejection is transient.
 		if (statusClean === 400 && COPILOT_TRANSIENT_MODEL_PATTERN.test(cleanMessage)) kinds |= Flag.Transient;
+		// Fireworks mid-generation NaN 400 is a model-side decode fault, not a bad
+		// request; a byte-identical replay succeeds, so treat it as transient.
+		if (statusClean === 400 && GENERATION_NAN_PATTERN.test(cleanMessage)) kinds |= Flag.Transient;
 		if (matchesStrictToolsRejection(cleanMessage, statusClean)) kinds |= Flag.Grammar;
 		if (matchesFastModeUnsupported(cleanMessage, statusClean)) kinds |= Flag.FastModeUnsupported;
 	}

package/src/providers/openai-completions.ts

@@ -603,26 +603,31 @@ const streamOpenAICompletionsOnce = (
 		);
 		const { requestAbortController, requestSignal } = abortTracker;
 		const onSseEvent = options?.onSseEvent;
-		const rawSseObserver = onSseEvent
-			? (event: RawSseEvent) => {
-					if (!event.event && event.data && event.data !== "[DONE]") {
-						try {
-							const parsed = JSON.parse(event.data);
-							const resolvedEvent =
-								typeof parsed.type === "string"
-									? parsed.type
-									: typeof parsed.object === "string"
-										? parsed.object
-										: null;
-							if (resolvedEvent) {
-								event.event = resolvedEvent;
-								event.raw = [`event: ${resolvedEvent}`, ...event.raw];
-							}
-						} catch {}
-					}
-					onSseEvent(event, model);
+		// Track the OpenAI `[DONE]` sentinel independently of `onSseEvent`: it is
+		// the streaming protocol's terminal signal, so a stream that ends with it
+		// completed by server agreement even when no `finish_reason` chunk arrived.
+		let sawDoneSentinel = false;
+		const rawSseObserver = (event: RawSseEvent) => {
+			if (event.data === "[DONE]") sawDoneSentinel = true;
+			if (onSseEvent) {
+				if (!event.event && event.data && event.data !== "[DONE]") {
+					try {
+						const parsed = JSON.parse(event.data);
+						const resolvedEvent =
+							typeof parsed.type === "string"
+								? parsed.type
+								: typeof parsed.object === "string"
+									? parsed.object
+									: null;
+						if (resolvedEvent) {
+							event.event = resolvedEvent;
+							event.raw = [`event: ${resolvedEvent}`, ...event.raw];
+						}
+					} catch {}
 				}
-			: undefined;
+				onSseEvent(event, model);
+			}
+		};
 		// Assigned once the block helpers exist (they are scoped to the `try`);
 		// the catch handler uses it to close open blocks before emitting the
 		// terminal error so both exit paths obey the same block lifecycle.
@@ -1301,7 +1306,15 @@ const streamOpenAICompletionsOnce = (
 			// Detect premature stream closure before the normal block-finalization
 			// sweep. Throwing after that sweep would make the error handler emit a
 			// second text_end/thinking_end for the same partial block.
-			if (streamFinishedAt === undefined && output.content.length > 0) {
+			//
+			// Only a genuine truncation — transport EOF with neither a
+			// `finish_reason` chunk nor the `[DONE]` sentinel — is incomplete. A
+			// stream terminated by `[DONE]` completed by server agreement; some
+			// OpenAI-compatible hosts omit or `null` the `finish_reason` and rely on
+			// `[DONE]` alone, so finalize those as the default `stop`
+			// (mapStopReason(null)) instead of surfacing a false incomplete-stream
+			// error and retrying every turn.
+			if (streamFinishedAt === undefined && !sawDoneSentinel && output.content.length > 0) {
 				throw new AIError.ProviderResponseError(
 					"OpenAI completions stream closed before a finish_reason was received",
 					{ provider: model.provider, kind: "incomplete-stream" },