fix(anthropic): non-stream /v1/messages returned OpenAI shape; strip EOS sentinel from text deltas

BUG: Anthropic client (sourceFormat=CLAUDE) hitting a claude-target provider
got a raw OpenAI chat.completion body on non-streaming requests. The
needsTranslation(CLAUDE,CLAUDE) gate is false when target===source, so the
translator never ran; a claude-transport executor replying OpenAI JSON
(opencode/big-pickle) leaked choices[]/prompt_tokens to the client, which
Anthropic SDKs cannot parse (no content[] blocks, no type:"message").

FIX: shape-aware guard toClaudeMessageShape() in nonStreamingHandler — when
sourceFormat is CLAUDE, convert any OpenAI-shape body to a proper Claude
message (type, content blocks with thinking/text/tool_use, stop_reason via
finish mapping, usage input/output tokens). Claude-shaped bodies pass through.

Also: strip <|im_end|>/<|endoftext|>/<|eot_id|> EOS sentinels from Claude
text deltas in openai-to-claude and kiro-to-claude translators — the upstream
EOS token leaks into the final text_delta (observed 'OK<|im_end|>').

Tests: anthropic-nonstream-shape.test.js (6: eos strip + shape guard incl
tool_calls→tool_use, pass-through, finish mapping). 28/28 related tests green.

open-sse/handlers/chatCore/nonStreamingHandler.js

@@ -23,6 +23,20 @@ function parseToolArguments(value) {
   }
 }
 
+/**
+ * Convert an OpenAI Chat Completions body into the Claude Messages shape when
+ * the client speaks Anthropic (sourceFormat=CLAUDE) but the provider replied
+ * OpenAI JSON. This is the shape-aware guard for the non-streaming path: the
+ * needsTranslation(CLAUDE,CLAUDE) gate is false when target===source, so the
+ * raw provider body would otherwise leak to the Anthropic client with no
+ * content[] blocks and no type:"message".
+ */
+export function toClaudeMessageShape(responseBody) {
+  if (!responseBody || responseBody.type === "message") return responseBody;
+  if (responseBody?.choices) return openAICompletionToClaudeMessage(responseBody);
+  return responseBody;
+}
+
 function openAICompletionToClaudeMessage(responseBody) {
   if (!responseBody?.choices?.[0]) return responseBody;
   const choice = responseBody.choices[0];
@@ -346,9 +360,19 @@ export async function handleNonStreamingResponse({ providerResponse, provider, m
   saveUsageStats({ provider, model, tokens: usage, connectionId, apiKey, endpoint: clientRawRequest?.endpoint, silent: true });
   if (log?.line) log.line(reqTag, "📊", formatDoneLine({ usage, latency: { total: Date.now() - requestStartTime } }));
 
-  const translatedResponse = needsTranslation(targetFormat, sourceFormat)
+  let translatedResponse = needsTranslation(targetFormat, sourceFormat)
     ? translateNonStreamingResponse(responseBody, targetFormat, sourceFormat, customToolNames)
     : responseBody;
+  // Shape-aware guard: an Anthropic client (sourceFormat=CLAUDE) must receive a
+  // Claude message even when the provider answered in OpenAI Chat shape. The
+  // needsTranslation(CLAUDE,CLAUDE) gate above is false when target===source
+  // (both CLAUDE), so the raw provider body is returned — and a claude-transport
+  // executor replying OpenAI JSON leaks a chat.completion body to the client
+  // (observed with opencode/big-pickle: no content[], no type:"message").
+  if (sourceFormat === FORMATS.CLAUDE) {
+    translatedResponse = toClaudeMessageShape(translatedResponse);
+  }
+
   const isClaudeMessageResponse = sourceFormat === FORMATS.CLAUDE && translatedResponse?.type === "message";
   // Responses-format translation produces a `object:"response"` body with no
   // `choices`; skip the Chat-Completions-specific post-processing below for it.

open-sse/translator/concerns/eosStrip.js

@@ -0,0 +1,32 @@
+// EOS sentinel strip for translated streaming text.
+//
+// Some upstream executors/providers (observed: opencode/big-pickle behind a
+// claude transport) append their special end-of-stream token verbatim to the
+// final text delta — e.g. "OK<|im_end|>" — because the tokenizer emits it and
+// the executor's OpenAI-compatible decoder passes it through as content. An
+// Anthropic client receiving such a delta renders the literal marker in the
+// transcript (and a strict schema check flags the unknown character run).
+// Stripping is safe: these tokens only ever appear exactly once, as a suffix
+// of the final chunk, and never carry meaning in a translated stream.
+const EOS_SENTINELS = [
+  "<|im_end|>",
+  "<|endoftext|>",
+  "<|eot_id|>",
+  "<|end|>",
+];
+
+/** Remove a trailing EOS sentinel from streaming text (suffix-only, once). */
+export function stripEosSentinel(text) {
+  if (!text || typeof text !== "string") return text;
+  for (const sentinel of EOS_SENTINELS) {
+    if (text.endsWith(sentinel)) {
+      return text.slice(0, -sentinel.length);
+    }
+  }
+  // Some providers emit it mid-chunk boundary (sentinel split across two
+  // deltas) — strip any occurrence, since these tokens have no legitimate
+  // place in model output.
+  const hasAny = EOS_SENTINELS.find((s) => text.includes(s));
+  if (hasAny) return text.split(hasAny).join("");
+  return text;
+}

open-sse/translator/response/kiro-to-claude.js

@@ -15,6 +15,7 @@
  */
 import { register } from "../index.js";
 import { FORMATS } from "../formats.js";
+import { stripEosSentinel } from "../concerns/eosStrip.js";
 
 function stopThinkingBlock(state, results) {
   if (!state.thinkingBlockStarted) return;
@@ -153,7 +154,7 @@ export function kiroToClaudeResponse(chunk, state) {
     results.push({
       type: "content_block_delta",
       index: state.textBlockIndex,
-      delta: { type: "text_delta", text: delta.content },
+      delta: { type: "text_delta", text: stripEosSentinel(delta.content) },
     });
   }
 

open-sse/translator/response/openai-to-claude.js

@@ -3,6 +3,7 @@ import { FORMATS } from "../formats.js";
 import { ROLE, CLAUDE_BLOCK, MODEL_FALLBACK } from "../schema/index.js";
 import { fromOpenAIFinish } from "../concerns/finishReason.js";
 import { extractReasoningText } from "../concerns/reasoning.js";
+import { stripEosSentinel } from "../concerns/eosStrip.js";
 
 // Legacy "proxy_" prefix used by older request translators. Response strips it
 // defensively so tool names from such turns resolve back (e.g. proxy_Read → Read
@@ -175,7 +176,7 @@ export function openaiToClaudeResponse(chunk, state) {
     results.push({
       type: "content_block_delta",
       index: state.textBlockIndex,
-      delta: { type: "text_delta", text: delta.content }
+      delta: { type: "text_delta", text: stripEosSentinel(delta.content) }
     });
   }