pi-freeflow 1.32.1 → 1.33.0

package/CHANGELOG.md

@@ -1,5 +1,32 @@
 # Changelog
 
+## 1.33.0
+
+### Minor Changes
+
+- One new free model, picked the same way as the rest (`/model` → `freeflow` → pick):
+
+  - **Ling 3.1 Flash** on KiloCode (`inclusionai/ling-3.1-flash`, short name `ling-3.1-flash:kilo`) — 262K context, text answers.
+
+### Patch Changes
+
+- Remove four free models that stopped serving upstream, so they no longer appear in the model list:
+
+  - **MiMo V2.5** on OpenCode (`mimo-v2.5-free`) — upstream answers `401 Model is not supported`.
+  - **Exo** on OpenCode (`exo-free`) — upstream answers `503 Endpoint is unavailable`.
+  - **Space Bunny Alpha** on KiloCode (`stealth/space-bunny-alpha`) — upstream answers `404 does not exist`.
+  - **Qwen 3.8 27B** on KiloCode (`qwen/qwen3.8-27b:free`) — upstream answers `404 currently unavailable`.
+
+  All four are excluded permanently, so a stale on-disk catalog or a later catalog refresh cannot bring them back.
+
+- Requests that ask for a thinking level a model does not offer no longer fail: the level is clamped to the highest level that model supports (previously only Muse Spark was covered — now every model is), and a chat-style `response_format` left on a responses request is dropped instead of rejected upstream.
+
+## 1.32.2
+
+### Patch Changes
+
+- Prevent JSON parse errors on interrupted responses: network packet boundaries split mid-delta no longer produce unterminated string syntax errors when a turn is retried or terminated.
+
 ## 1.32.1
 
 ### Patch Changes

package/README.md

@@ -1,6 +1,6 @@
 # pi-freeflow
 
-> 33 free models with up to 1M context. No API keys to manage. Add your own relays to spread requests across more IPs.
+> 31 free models with up to 1M context. No API keys to manage. Add your own relays to spread requests across more IPs.
 
 Thin by design: a model list, a relay proxy, and a log. The host (`pi-ai`) handles thinking, normalization, and provider behavior.
 
@@ -16,7 +16,7 @@ Thin by design: a model list, a relay proxy, and a log. The host (`pi-ai`) handl
 
 | Feature | What it does | Cost |
 | :--- | :--- | :--- |
-| **33 free models** | 11 from OpenCode Zen, 17 from KiloCode Gateway, 5 from Cline, context windows up to 1M. Full list below. | **$0** |
+| **31 free models** | 10 from OpenCode Zen, 16 from KiloCode Gateway, 5 from Cline, context windows up to 1M. Full list below. | **$0** |
 | **Relay pool** | Route requests through your own Cloudflare Workers and Vercel Edge relays. Traffic spreads across the pool, or stays on one sticky relay. A relay that rate-limits, times out, or drops the connection cools down while healthy ones take its traffic. | **$0** beyond your platforms' free tiers |
 | **Automatic fallback** | When every relay is cooling down, requests go direct to upstream instead of failing. | **$0** |
 | **Short model names** | Every model has a slash-free, colon-free alias, plus an optional `:effort` suffix for thinking depth. You type `freeflow/<name>`. | **$0** |
@@ -185,13 +185,13 @@ The same command set works identically in OMP and Pi:
 
 ---
 
-### 33 models, one command
+### 31 models, one command
 
 ```bash
 /model → freeflow → pick
 ```
 
-#### OpenCode Zen (11 models), Responses and Chat API
+#### OpenCode Zen (10 models), Responses and Chat API
 
 Good defaults for long coding sessions and agentic work.
 
@@ -199,7 +199,6 @@ Good defaults for long coding sessions and agentic work.
 | :--- | :--- | :--- | :--- | :--- | :--- |
 | `muse-spark-1.2-contributor-free` | Meta Superintelligence Labs | **1M** (1.048.576) | **131K** (131.072) | `minimal … xhigh` | ✅ |
 | `muse-spark-1.3-contributor-free` | Meta Superintelligence Labs | **1M** (1.048.576) | **131K** (131.072) | `minimal … xhigh` | ✅ |
-| `mimo-v2.5-free` | Xiaomi MiMo | **1M** (1.048.576) | **131K** (131.072) | `minimal … xhigh`\* | ✅ |
 | `mimo-v2.6-flash-free` | Xiaomi MiMo | **1M** (1.048.576) | **131K** (131.072) | `minimal … xhigh`\* | ✅ |
 | `nemotron-3.5-lightning-free` | NVIDIA | **1M** (1.000.000) | **262K** (262.144) | `minimal … xhigh` | ❌ |
 | `nemotron-3-ultra-free` | NVIDIA | **1M** (1.000.000) | **128K** (128.000) | `minimal … xhigh` | ❌ |
@@ -209,7 +208,7 @@ Good defaults for long coding sessions and agentic work.
 | `fledge-alpha-free` | Stealth preview (lab undisclosed) | **1M** (1.048.576) | **131K** (131.072) | `low / high / max` | ✅ |
 | `ling-3.1-flash-free` | Inclusion AI | **262K** (262.144) | **32K** (32.768) | `minimal … max` | ❌ |
 
-#### KiloCode Gateway (17 models), OpenRouter compatible
+#### KiloCode Gateway (16 models), OpenRouter compatible
 
 Keyless access. Short aliases work for every row (the full ID is in parentheses).
 
@@ -229,9 +228,8 @@ Keyless access. Short aliases work for every row (the full ID is in parentheses)
 | `content-safety` (`nvidia/...:free`) | NVIDIA | **128K** (128.000) | **8K** (8.192) | ❌ *(non-thinking)* | ✅ |
 | `ling-3.0-flash-sante` (`inclusionai/ling-3.0-flash-sante:free`) | Inclusion AI | **262K** (262.144) | **32K** (32.768) | `minimal…xhigh`\* | ❌ |
 | `step-3.7-flash` (`stepfun/...:free`) | StepFun | **262K** (262.144) | **262K** (262.144) | `minimal…xhigh`\* | ✅ |
-| `space-bunny-alpha` (`stealth/...`) | Stealth preview (lab undisclosed) | **1M** (1.000.000) | **512K** (524.288) | `minimal…xhigh`\* | ✅ |
 | `inkling-small` (`thinkingmachines/inkling-small:free`) | Thinking Machines | **1M** (1.048.576) | **262K** (262.144) | `minimal…xhigh`\* | ✅ |
-| `qwen3.8-27b` (`qwen/...:free`) | Alibaba Qwen | **262K** (262.144) | **235K** (235.929) | `minimal…xhigh`\* | ✅ |
+| `ling-3.1-flash-kilo` (`inclusionai/ling-3.1-flash`) | Inclusion AI | **262K** (262.144) | **32K** (32.768) | `minimal…xhigh`\* | ❌ |
 
 \* Levels are forwarded as-is through the OpenRouter-style nested `reasoning` parameter; effort mapping is decided by each model. MiMo collapses `minimal→low` and `xhigh→high` upstream, so its selector shows 5 labels but only 3 distinct effort values.
 
@@ -344,7 +342,7 @@ pnpm smoke       # verifies extensions/index.ts loads without crashing
 ```
 src/
 ├── index.ts          # extension entry, lifecycle hooks
-├── models.ts         # 33-model catalog definitions
+├── models.ts         # 31-model catalog definitions
 ├── catalog.ts        # model catalog cache (24h disk)
 ├── proxy.ts          # local proxy server (127.0.0.1:28180)
 ├── relay.ts          # relay fetch, rolling on retryable failures

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "pi-freeflow",
   "type": "module",
-  "version": "1.32.1",
+  "version": "1.33.0",
   "description": "Thin provider for OMP/Pi — model list + dumb relay proxy + log; host pi-ai owns thinking/normalization",
   "main": "extensions/index.ts",
   "types": "src/index.ts",

package/src/catalog.ts

@@ -42,6 +42,10 @@ import type {
  *   reproduced across separate runs on different days while every other Zen model answers.
  * - inclusionai/ling-3.0-flash-fin:free: gone from Kilo (2026-10-01) — keyless chat 404
  *   "The requested model ... does not exist".
+ * - mimo-v2.5-free: decommissioned upstream (2026-10-07) — HTTP 401 "Model mimo-v2.5-free is not supported".
+ * - exo-free: unserved upstream (2026-10-07) — HTTP 503 "Error from provider (Console): Upstream request failed: Endpoint is unavailable.".
+ * - stealth/space-bunny-alpha: gone from Kilo (2026-10-07) — keyless chat 404 "The requested model ... does not exist".
+ * - qwen/qwen3.8-27b:free: gone from Kilo (2026-10-07) — keyless chat 404 "The requested model is currently unavailable.".
  */
 export const DEAD_MODEL_IDS = new Set<string>([
  "jev-1.13-free",
@@ -61,13 +65,17 @@ export const DEAD_MODEL_IDS = new Set<string>([
  "inclusionai/ling-3.0-flash-vl:free",
  "ling-3.0-flash-fin-free",
  "inclusionai/ling-3.0-flash-fin:free",
+ "mimo-v2.5-free",
+ "exo-free",
+ "stealth/space-bunny-alpha",
+ "qwen/qwen3.8-27b:free",
 ]);
 /**
  * Free-tier allowlist for anything entering the picker via network or stale disk.
  * Upstream lists paid models alongside free ones (e.g. claude-fable-5-1,
  * claude-opus-4-*, gemini-3-*) so a bare upstream merge leaks paid entries that
  * fail with 401 Missing API key. Known static IDs without a free suffix
- * (e.g. big-pickle, stealth/space-bunny-alpha) stay allowed via MODEL_MAP.
+ * (e.g. big-pickle, inclusionai/ling-3.1-flash) stay allowed via MODEL_MAP.
  */
 export function isFreeCatalogId(id: string): boolean {
  if (typeof id !== "string" || id.length === 0) return false;
@@ -100,7 +108,7 @@ export function sanitizeCatalogModels(models: RegisteredModel[]): RegisteredMode
 }
 /**
  * In-memory cache of currently active/available free models.
- * Initialized with all 33 verified models for 0ms instant availability.
+ * Initialized with all 31 verified models for 0ms instant availability.
  */
 let aliveCatalog: RegisteredModel[] = ALL_MODELS.map((m) => ({
  ...m,

package/src/index.ts

@@ -1,7 +1,7 @@
 /**
  * pi-freeflow — Modular, high-resiliency LLM extension for Pi & Oh My Pi (OMP)
  *
- * Provides access to 33 free models (11 OpenCode Zen + 17 KiloCode Gateway + 5 Cline) with:
+ * Provides access to 31 free models (10 OpenCode Zen + 16 KiloCode Gateway + 5 Cline) with:
  * - Single-port daemon reuse on 28180 across concurrent subagents
  * - Multi-cloud rolling egress relays (Vercel Edge, Cloudflare, Deno)
  * - 0ms instant startup with verified static catalog and background live health checks

package/src/models.ts

@@ -1,9 +1,9 @@
 /**
  * Static model definitions and upstream routing catalogs for pi-freeflow
  *
- * Defines the 33 verified free models (live-verified 2026-09-28):
+ * Defines the 31 verified free models (live-verified 2026-10-07):
  * - 10 OpenCode Zen models (2 Responses API + 8 Chat Completions)
- * - 18 KiloCode Keyless Gateway models (OpenRouter format)
+ * - 16 KiloCode Keyless Gateway models (OpenRouter format)
  * - 5 Cline direct-only models (per-user pool)
  */
 
@@ -50,23 +50,6 @@ export const OPENCODE_MODELS: ModelDef[] = [
    max: null,
   },
  },
- {
-  id: "mimo-v2.5-free",
-  name: "MiMo V2.5 [OpenCode]",
-  reasoning: true,
-  contextWindow: 1_048_576,
-  maxTokens: 131_072,
-  input: ["text", "image"],
-  thinkingLevelMap: {
-   off: null,
-   minimal: "low",
-   low: "low",
-   medium: "medium",
-   high: "high",
-   xhigh: "high",
-   max: null,
-  },
- },
  {
   id: "mimo-v2.6-flash-free",
   name: "MiMo V2.6 Flash [OpenCode]",
@@ -356,19 +339,6 @@ export const KILO_MODELS: ModelDef[] = [
   thinkingFormat: "openrouter",
   thinkingLevelMap: KILO_REASONING_MAP,
  },
- {
-  // Added 2026-09-28: live on the Kilo free list (isFree:true, 0/0
-  // pricing, 1M context, text+image in) + keyless chat 200. Passes
-  // isFreeCatalogId via MODEL_MAP (no :free suffix by design).
-  id: "stealth/space-bunny-alpha",
-  name: "Space Bunny Alpha [Kilo]",
-  reasoning: true,
-  contextWindow: 1_000_000,
-  maxTokens: 524_288,
-  input: ["text", "image"],
-  thinkingFormat: "openrouter",
-  thinkingLevelMap: KILO_REASONING_MAP,
- },
  {
   id: "thinkingmachines/inkling-small:free",
   name: "Inkling Small [Kilo]",
@@ -390,12 +360,13 @@ export const KILO_MODELS: ModelDef[] = [
   thinkingLevelMap: KILO_REASONING_MAP,
  },
  {
-  id: "qwen/qwen3.8-27b:free",
-  name: "Qwen 3.8 27B [Kilo]",
+  // Added 2026-10-07: live on Kilo free list (isFree:true, 0/0 pricing, 262K ctx)
+  id: "inclusionai/ling-3.1-flash",
+  name: "Ling 3.1 Flash [Kilo]",
   reasoning: true,
   contextWindow: 262_144,
-  maxTokens: 235_929,
-  input: ["text", "image"],
+  maxTokens: 32_768,
+  input: ["text"],
   thinkingFormat: "openrouter",
   thinkingLevelMap: KILO_REASONING_MAP,
  },
@@ -517,7 +488,6 @@ export const MODEL_ALIASES: Record<string, string> = {
  "content-safety": "nvidia/nemotron-3.5-content-safety:free",
  "ling-3.0-flash-sante": "inclusionai/ling-3.0-flash-sante:free",
  "step-3.7-flash": "stepfun/step-3.7-flash:free",
- "space-bunny-alpha": "stealth/space-bunny-alpha",
  "inkling-small": "thinkingmachines/inkling-small:free",
  "mimo-v2.6-flash": "mimo-v2.6-flash-free",
  "space-bunny": "space-bunny-free",
@@ -525,7 +495,7 @@ export const MODEL_ALIASES: Record<string, string> = {
  "longcat": "longcat-2.5-preview-free",
  "fledge-alpha": "fledge-alpha-free",
  "ling-3.1-flash": "ling-3.1-flash-free",
- "qwen3.8-27b": "qwen/qwen3.8-27b:free",
+ "ling-3.1-flash:kilo": "inclusionai/ling-3.1-flash",
  // provider-prefixed short aliases (slash-normalized)
  "laguna-s-2.1:free": "poolside/laguna-s-2.1:free",
  "laguna-xs-2.1:free": "poolside/laguna-xs-2.1:free",
@@ -559,9 +529,8 @@ export const CLINE_MODEL_IDS = new Set<string>([
   .filter(([_, target]) => CLINE_MODELS.some((cm) => cm.id === target))
   .map(([alias]) => alias),
 ]);
-
 /**
- * Combined list of all 33 static free models (canonical)
+ * Combined list of all 31 static free models (canonical)
  */
 export const ALL_MODELS: ModelDef[] = [...OPENCODE_MODELS, ...KILO_MODELS, ...CLINE_MODELS];
 

package/src/opencode-fingerprint.ts

@@ -38,6 +38,35 @@ import {
  type FindGlobRestore,
  type FingerprintToolName,
 } from "./tool-translation.ts";
+import { getModelDef } from "./models.ts";
+import type { ThinkingLevel } from "./types.ts";
+
+/** Effort levels in descending priority (off excluded — clamping to off is never correct). */
+const EFFORT_DESCENDING: ThinkingLevel[] = ["max", "xhigh", "high", "medium", "low", "minimal"];
+
+/**
+ * Find the highest supported effort for a model whose thinkingLevelMap
+ * declares the requested effort as null. Returns null when no clamp is
+ * needed (effort is supported or model has no map).
+ */
+function clampedEffortFor(model: string, effort: string): string | null {
+ const def = getModelDef(model);
+ if (!def?.thinkingLevelMap) return null;
+ const map = def.thinkingLevelMap;
+ // "off" means disable reasoning entirely — never clamp it even when map says null
+ if (effort.toLowerCase() === "off") return null;
+ // "ultra" is an unofficial alias for beyond-max; treat as "max"
+ const key = (effort.toLowerCase() === "ultra" ? "max" : effort.toLowerCase()) as ThinkingLevel;
+ // Effort is supported (non-null) → no clamp
+ if (key in map && map[key] !== null && map[key] !== undefined) return null;
+ // Effort key is unknown and not "ultra" → let upstream decide
+ if (!(key in map) && effort.toLowerCase() !== "ultra") return null;
+ // Find highest supported effort
+ for (const level of EFFORT_DESCENDING) {
+  if (map[level] !== null && map[level] !== undefined) return level;
+ }
+ return null;
+}
 
 /**
  * Safely extract the tool name whether formatted in Chat Completions style
@@ -262,14 +291,15 @@ export function ensureMessagesFingerprintTools(body: Record<string, unknown>, om
  * - `max_tokens` / `max_completion_tokens` become `max_output_tokens` (copied
  *   when absent, dropped either way).
  * - `temperature` / `top_p` are dropped (Responses rejects them).
+ * - `response_format` is dropped (Chat Completions field, Responses rejects it).
  * - `parallel_tool_calls: false` is dropped; absent-or-true rides verbatim.
  * - `tool_choice` is auto-or-absent: `"none"` (string or `{ type: "none" }`)
  *   is dropped, everything else rides (named choices still retarget find->glob).
  * - `store` defaults to false (never persisted server-side).
- * - spark reasoning effort clamps `max` / `ultra` to `xhigh`: the spark
- *   thinkingLevelMap declares `max: null` (unoffered) and upstream serves at
- *   most `xhigh`, so a stale or hand-built `max`/`ultra` effort would 400.
- *   Only spark model ids are clamped; `off: null` stays untouched everywhere.
+ * - Reasoning effort is clamped per model: when the model's thinkingLevelMap
+ *   declares the requested effort as null (unsupported), the effort is clamped
+ *   to the highest supported level. Covers all OpenCode models generically
+ *   (muse-spark, mimo, nemotron, etc.), not just a single model family.
  * Zen-responses only: chat/messages paths keep their own fields, and Kilo/Cline
  * bodies never reach here (the proxy fingerprints Zen bodies exclusively).
  */
@@ -287,23 +317,28 @@ export function normalizeResponsesBody(body: Record<string, unknown>): void {
  delete body.max_completion_tokens;
  delete body.temperature;
  delete body.top_p;
+ delete body.response_format;
  if (body.parallel_tool_calls === false) delete body.parallel_tool_calls;
  const choice = body.tool_choice;
  if (typeof choice === "string" ? choice.toLowerCase() === "none" : (
   choice !== null && typeof choice === "object" && !Array.isArray(choice)
   && (choice as Record<string, unknown>).type === "none"
  )) delete body.tool_choice;
- if (typeof body.model === "string" && /muse-spark/i.test(body.model)) {
+ if (typeof body.model === "string") {
   const reasoning = body.reasoning;
   if (reasoning !== null && typeof reasoning === "object" && !Array.isArray(reasoning)) {
    const effort = (reasoning as Record<string, unknown>).effort;
-   if (typeof effort === "string" && (effort.toLowerCase() === "max" || effort.toLowerCase() === "ultra")) {
-    (reasoning as Record<string, unknown>).effort = "xhigh";
+   if (typeof effort === "string") {
+    const clamped = clampedEffortFor(body.model as string, effort);
+    if (clamped !== null) (reasoning as Record<string, unknown>).effort = clamped;
    }
   }
   for (const key of ["reasoning_effort", "reasoningEffort"] as const) {
    const flat = body[key];
-   if (typeof flat === "string" && (flat.toLowerCase() === "max" || flat.toLowerCase() === "ultra")) body[key] = "xhigh";
+   if (typeof flat === "string") {
+    const clamped = clampedEffortFor(body.model as string, flat);
+    if (clamped !== null) body[key] = clamped;
+   }
   }
  }
  if (body.store === undefined) body.store = false;

package/src/stream-pipe.ts

@@ -745,6 +745,8 @@ export function pipeUpstreamStream(
   const cut = nodeStream;
   upstreamEnded = false;
   nodeStream = next;
+  sseBuffer = "";
+  terminalScanCarry = Buffer.alloc(0);
   watch(next);
   try {
    if (!cut.destroyed) cut.destroy();
@@ -805,11 +807,11 @@ export function pipeUpstreamStream(
    try {
     if (isError) {
      res.write(
-      `\nevent: response.failed\ndata: {"type":"response.failed","response":{"status":"failed","error":{"code":"stream_error","message":${JSON.stringify(errorMsg || "Upstream stream disconnected unexpectedly")}}}}\n\n`,
+      `\n\nevent: response.failed\ndata: {"type":"response.failed","response":{"status":"failed","error":{"code":"stream_error","message":${JSON.stringify(errorMsg || "Upstream stream disconnected unexpectedly")}}}}\n\n`,
      );
     } else {
      res.write(
-      `\nevent: response.incomplete\ndata: {"type":"response.incomplete","response":{"status":"incomplete","incomplete_details":{"reason":"cancelled"}}}\n\n`,
+      `\n\nevent: response.incomplete\ndata: {"type":"response.incomplete","response":{"status":"incomplete","incomplete_details":{"reason":"cancelled"}}}\n\n`,
      );
     }
     hasTerminalEvent = true;
@@ -822,7 +824,7 @@ export function pipeUpstreamStream(
    } catch { }
   } else if (!isResponsesApi && totalChunks > 0) {
    try {
-    res.write("\ndata: [DONE]\n\n");
+    res.write("\n\ndata: [DONE]\n\n");
     hasTerminalEvent = true;
    } catch { }
   }
@@ -860,6 +862,8 @@ export function pipeUpstreamStream(
   }
   upstreamEnded = false;
   nodeStream = next;
+  sseBuffer = "";
+  terminalScanCarry = Buffer.alloc(0);
   watch(next);
   log("debug", "mid-stream refire: continuing turn on replacement stream", { totalChunks, totalBytes }, rid);
   return true;
@@ -902,59 +906,22 @@ export function pipeUpstreamStream(
      );
     }
    }
-   if (cloakState && cloak) {
-    // Cloaked stream: buffer across chunks so event names split mid-chunk
-    // still match, rewrite complete SSE blocks, and drop injected calls.
-    sseBuffer += typeof chunk === "string" ? chunk : chunk.toString("utf8");
-    let boundary = sseBoundary.exec(sseBuffer);
-    while (boundary) {
-     const block = sseBuffer.slice(0, boundary.index);
-     sseBuffer = sseBuffer.slice(boundary.index + boundary[0].length);
+   // Buffer across chunks into complete SSE blocks delimited by double-newline.
+   // Forwarding only complete blocks guarantees that network chunk boundaries
+   // never split JSON strings on the wire or leave unterminated syntax when
+   // mid-stream refire or truncation occurs.
+   sseBuffer += typeof chunk === "string" ? chunk : chunk.toString("utf8");
+   let boundary = sseBoundary.exec(sseBuffer);
+   while (boundary) {
+    const block = sseBuffer.slice(0, boundary.index);
+    sseBuffer = sseBuffer.slice(boundary.index + boundary[0].length);
+    if (cloakState && cloak) {
      const rewritten = rewriteSseBlock(block, cloakState, cloak);
      if (rewritten !== null) forwardText(`${rewritten}\n\n`);
-     boundary = sseBoundary.exec(sseBuffer);
+    } else {
+     forwardText(`${block}\n\n`);
     }
-    return;
-   }
-   // Always scan the FULL chunk plus the small carry from the previous
-   // view. SSE chunks are KBs at most, so includes() over everything
-   // is negligible against correctness — the removed head/tail windows
-   // are exactly what let markers buried mid-chunk or split across
-   // adjacent chunks escape and cause duplicate synthetic terminals.
-   const buf =
-    typeof chunk === "string" ? Buffer.from(chunk, "utf8") : chunk;
-   const scanBuf =
-    terminalScanCarry.length > 0
-     ? Buffer.concat([terminalScanCarry, buf])
-     : buf;
-   const scanned = scanBuf.toString("utf8");
-   if (!sawToolCall && TOOL_CALL_MARKERS.some((m) => scanned.includes(m))) sawToolCall = true;
-   if (!hasTerminalEvent && checkTerminalEvent(scanned)) {
-    // Upstream-issued transient model failure on a text-only turn: hold the
-    // frame for one silent resume instead of terminating (tool turns veto
-    // inside holdModelFailureForRefire and keep the terminal below).
-    if (isTransientModelFailureFrame(scanned) && holdModelFailureForRefire(chunk)) return;
-    hasTerminalEvent = true;
-   }
-   terminalScanCarry = Buffer.from(scanBuf.subarray(Math.max(0, scanBuf.length - 32)));
-
-   const canContinue = res.write(chunk);
-   if (canContinue === false && !nodeStream.destroyed) {
-    // Backpressure: pause the upstream source until the client
-    // socket drains, instead of buffering an unbounded amount into
-    // the response. 'drain' might never fire if the client closes —
-    // the close handlers below destroy the stream regardless of
-    // its paused state.
-    nodeStream.pause();
-    const onDrain = () => {
-     res.off("drain", onDrain);
-     if (!nodeStream.destroyed) nodeStream.resume();
-    };
-    res.once("drain", onDrain);
-   }
-   const maybeFlush = res as unknown as { flush?: () => void };
-   if (typeof maybeFlush.flush === "function") {
-    maybeFlush.flush();
+    boundary = sseBoundary.exec(sseBuffer);
    }
   } catch { }
  };
@@ -985,9 +952,13 @@ export function pipeUpstreamStream(
     res.writeHead(502, { "content-type": "application/json" });
    } else if (isInternalAbort) {
     clientAborted = true;
+    sseBuffer = "";
+    terminalScanCarry = Buffer.alloc(0);
     ensureTerminalEvent(false, errorMsg || "stream interrupted", false);
    } else {
     if (await tryRefire()) return;
+    sseBuffer = "";
+    terminalScanCarry = Buffer.alloc(0);
     ensureTerminalEvent(!isSubstantial(totalChunks, totalBytes), errorMsg, true);
    }
    if (!res.writableEnded) {
@@ -1003,11 +974,17 @@ export function pipeUpstreamStream(
   upstreamEnded = true;
   // Cloaked streams hold the trailing incomplete block: flush it (rewritten)
   // ahead of the terminal-marker check so a coalesced final event still counts.
-  if (cloakState && cloak && sseBuffer.trim() !== "") {
+  if (sseBuffer.trim() !== "") {
    const tail = sseBuffer;
    sseBuffer = "";
-   const rewritten = rewriteSseBlock(tail, cloakState, cloak);
-   if (rewritten !== null) forwardText(`${rewritten}\n\n`);
+   if (checkTerminalEvent(tail)) {
+    if (cloakState && cloak) {
+     const rewritten = rewriteSseBlock(tail, cloakState, cloak);
+     if (rewritten !== null) forwardText(`${rewritten}\n\n`);
+    } else {
+     forwardText(tail);
+    }
+   }
   }
   if (hasTerminalEvent) recordOnce(false);
   const elapsed = ((Date.now() - startAt) / 1000).toFixed(1);
@@ -1054,11 +1031,15 @@ export function pipeUpstreamStream(
     if (clientAborted || upstreamEnded) {
      // Client disconnect teardown or post-end cleanup — the relay
      // is not at fault; still give the host a terminal event.
+     sseBuffer = "";
+     terminalScanCarry = Buffer.alloc(0);
      ensureTerminalEvent(false, "stream interrupted by client", false);
     } else {
      // Same offer as the clean-end path above: one silent re-fire before
      // marking this relay and terminating the turn.
      if (await tryRefire()) return;
+     sseBuffer = "";
+     terminalScanCarry = Buffer.alloc(0);
      // Upstream socket died mid-stream with no error event.
      // For muse-spark large payloads: raxtant 514KB failed but feoni 802KB
      // succeeded with same 2.6MB in — so this is edge-specific, not pure