stack 3/5: carry six contributor bug fixes with authorship intact

src/adapters/google-truncation.ts

@@ -11,3 +11,14 @@ export function vertexTruncationErrorMessage(reason?: string): string {
   const suffix = reason ? ` (${redactSecretString(reason).slice(0, 160)})` : "";
   return `Vertex AI response truncated upstream before the turn completed${suffix}`;
 }
+
+/**
+ * Whether a finished turn must fail closed. A truncation reason arriving mid tool call always
+ * does. MALFORMED_FUNCTION_CALL fails closed even with zero started calls: the malformed call
+ * is dropped upstream and usually never materializes as a part, so the turn is incomplete
+ * despite looking empty. MAX_TOKENS with no started call stays a plain token-limit stop.
+ */
+export function isVertexTruncatedTurn(finishReason: string | undefined, toolCallsStarted: number): boolean {
+  if (!isVertexTruncationReason(finishReason)) return false;
+  return toolCallsStarted > 0 || finishReason === "MALFORMED_FUNCTION_CALL";
+}

src/adapters/google.ts

@@ -17,7 +17,7 @@ import { contentPartsToText, parseDataUrl } from "./image";
 import { getVertexAccessToken } from "../lib/gcp-adc";
 import { fetchAntigravityWithRetry, fetchVertexWithRetry } from "./google-http";
 import { safeAntigravityHttpErrorMessage, safeVertexHttpErrorMessage } from "./google-errors";
-import { isVertexTruncationReason, vertexTruncationErrorMessage } from "./google-truncation";
+import { isVertexTruncatedTurn, vertexTruncationErrorMessage } from "./google-truncation";
 import { ANTIGRAVITY_REQUEST_UA, antigravitySessionId, isLikelyRealThoughtSignature, sanitizeAntigravityClaudeSignatures } from "./google-antigravity-wire";
 import { compileGoogleWireBody } from "./google-wire-compiler";
 import { identifyRoutedModel } from "./identity";
@@ -596,7 +596,7 @@ export function createGoogleAdapter(provider: OcxProviderConfig): ProviderAdapte
         // Fail-closed: a turn cut off mid tool call (MAX_TOKENS / MALFORMED_FUNCTION_CALL) surfaces
         // an error instead of a silently-incomplete done. Mirrors kiro-truncation.
         if ((provider.googleMode === "vertex" || provider.googleMode === "cloud-code-assist")
-          && toolCallsStarted > 0 && isVertexTruncationReason(lastFinishReason)) {
+          && isVertexTruncatedTurn(lastFinishReason, toolCallsStarted)) {
           yield { type: "error", message: vertexTruncationErrorMessage(lastFinishReason) };
           return;
         }
@@ -752,7 +752,7 @@ export function createGoogleAdapter(provider: OcxProviderConfig): ProviderAdapte
       // Fail-closed truncation, same as the stream path: a non-stream turn cut off mid tool call
       // (MAX_TOKENS / MALFORMED_FUNCTION_CALL) surfaces an error instead of a silent done.
       if ((provider.googleMode === "vertex" || provider.googleMode === "cloud-code-assist")
-        && toolCallsStarted > 0 && isVertexTruncationReason(candidates?.[0]?.finishReason)) {
+        && isVertexTruncatedTurn(candidates?.[0]?.finishReason, toolCallsStarted)) {
         return finish([{ type: "error", message: vertexTruncationErrorMessage(candidates?.[0]?.finishReason) }]);
       }
 

src/adapters/kiro-events.ts

@@ -3,7 +3,8 @@ import { kiroTruncationReason } from "./kiro-truncation";
 
 export type ParsedKiroEvent =
   | { type: "content"; data?: string; modelId?: string }
-  | { type: "reasoning"; data?: string }
+  | { type: "reasoning"; data?: string; redactedContent?: string }
+  | { type: "context_usage"; contextUsagePercentage: number }
   | { type: "tool"; name?: string; toolUseId?: string; input?: string; stop?: boolean }
   | { type: "truncation"; data: string }
   | { type: "metadata"; usage?: OcxUsage; contextUsagePercentage?: number; stopReason?: string }
@@ -17,6 +18,10 @@ const KNOWN_EVENT_TYPES = new Set([
   "toolUseEvent",
   "messageMetadataEvent",
   "metadataEvent",
+  // Authoritative context pressure. Every capture (kiro-cli 2.14.1 and 2.16.0) put the percentage
+  // HERE and left `metadataEvent` carrying only `stopReason`; metadataEvent's own
+  // contextUsagePercentage stays supported as a fallback rather than being dropped.
+  "contextUsageEvent",
   "invalidStateEvent",
   "error",
 ]);
@@ -114,11 +119,17 @@ export function parseKiroEvent(eventType: string, payload: Uint8Array): ParsedKi
           : {}),
       };
     case "reasoningContentEvent":
+      // `text` is plaintext reasoning; `redactedContent` is the encrypted blob the GPT-5.6 family
+      // (sol/terra/luna) actually returns — they never send `text`. Keyed off the wire field, not
+      // the model id. Both may be absent on a bare event.
       return {
         type: "reasoning",
         ...(optionalString(eventType, parsed, "text") !== undefined
           ? { data: optionalString(eventType, parsed, "text") }
           : {}),
+        ...(optionalString(eventType, parsed, "redactedContent") !== undefined
+          ? { redactedContent: optionalString(eventType, parsed, "redactedContent") }
+          : {}),
       };
     case "toolUseEvent":
       return {
@@ -161,6 +172,13 @@ export function parseKiroEvent(eventType: string, payload: Uint8Array): ParsedKi
         ...(stopReason !== undefined ? { stopReason } : {}),
       };
     }
+    case "contextUsageEvent": {
+      const contextUsagePercentage = parsed.contextUsagePercentage;
+      if (typeof contextUsagePercentage !== "number" || !Number.isFinite(contextUsagePercentage)) {
+        return malformed(eventType, "contextUsagePercentage must be a finite number");
+      }
+      return { type: "context_usage", contextUsagePercentage };
+    }
     case "invalidStateEvent":
       return { type: "invalid_state", message: optionalString(eventType, parsed, "message") };
     case "error":

src/adapters/kiro.ts

@@ -99,7 +99,11 @@ interface KiroUserInputMessage {
 }
 interface KiroHistoryEntry {
   userInputMessage?: KiroUserInputMessage;
-  assistantResponseMessage?: { content: string; toolUses?: KiroToolUse[] };
+  assistantResponseMessage?: {
+    content: string;
+    toolUses?: KiroToolUse[];
+    reasoningContent?: { redactedContent: string };
+  };
 }
 
 function kiroToolWireNames(tools: readonly unknown[]): string[] {
@@ -326,7 +330,7 @@ function validateKiroCapabilities(parsed: OcxParsedRequest): void {
 
 type KiroTurn =
   | { kind: "user"; content: string; images: KiroImage[]; toolResults: KiroToolResult[] }
-  | { kind: "assistant"; content: string; toolUses: KiroToolUse[] };
+  | { kind: "assistant"; content: string; toolUses: KiroToolUse[]; redactedReasoning?: string };
 
 function appendTurnText(target: string, next: string): string {
   if (!next) return target;
@@ -471,13 +475,15 @@ export function buildKiroPayload(
       turns.push({ kind: "user", content, images: [...images], toolResults: [...toolResults] });
     }
   };
-  const pushAssistant = (content: string, toolUses: KiroToolUse[]): void => {
+  const pushAssistant = (content: string, toolUses: KiroToolUse[], redactedReasoning?: string): void => {
     const last = turns.at(-1);
     if (last?.kind === "assistant") {
       last.content = appendTurnText(last.content, content);
       last.toolUses.push(...toolUses);
+      // Merged turns keep the newest blob: it covers the reasoning up to the merged turn's end.
+      if (redactedReasoning) last.redactedReasoning = redactedReasoning;
     } else {
-      turns.push({ kind: "assistant", content, toolUses: [...toolUses] });
+      turns.push({ kind: "assistant", content, toolUses: [...toolUses], ...(redactedReasoning ? { redactedReasoning } : {}) });
     }
   };
 
@@ -507,7 +513,7 @@ export function buildKiroPayload(
         const hasReasoning = aMsg.content.some(part => part.type === "thinking" && part.thinking.trim());
         if (hasReasoning) continue;
       }
-      pushAssistant(text, toolUses);
+      pushAssistant(text, toolUses, aMsg.kiroRedactedReasoning);
     } else if (msg.role === "toolResult") {
       const tr = msg as OcxToolResultMessage;
       if (tr.containsEncryptedContent) {
@@ -561,6 +567,7 @@ export function buildKiroPayload(
         assistantResponseMessage: {
           content: turn.content,
           ...(turn.toolUses.length > 0 ? { toolUses: turn.toolUses } : {}),
+          ...(turn.redactedReasoning ? { reasoningContent: { redactedContent: turn.redactedReasoning } } : {}),
         },
       }
     : {
@@ -1183,6 +1190,12 @@ async function* parseKiroAttemptEvents(
           if (ev.data) {
             yield* emitRetained(stage({ type: "reasoning_raw_delta", text: ev.data }));
           }
+          if (ev.redactedContent) {
+            yield* emitRetained(stage({ type: "kiro_redacted_reasoning", data: ev.redactedContent }));
+          }
+          break;
+        case "context_usage":
+          if (ev.contextUsagePercentage > 0) contextUsagePercentage = ev.contextUsagePercentage;
           break;
         case "tool": {
           for (const contentEvent of thinking.flush()) {

src/bridge.ts

@@ -442,6 +442,29 @@ export function bridgeToResponsesSSE(
         retainFinishedItem(item as OutputItem, bytesOf(encrypted), "reasoning");
         outputIndex++;
       };
+      // Kiro reasoning round-trip. Kiro sends its encrypted blob at the END of a turn, while the
+      // assistant message is still open, so this CANNOT emit on arrival: the open message still
+      // owns `outputIndex` (it only advances on close), and an item emitted here would both reuse
+      // that index and land BEFORE the message — where the parser's backwards pairing drops it as
+      // orphaned. Stash it and flush after `done` has closed every open item instead.
+      let pendingKiroRedacted: string | undefined;
+      let pendingKiroRedactedBytes = 0;
+      const flushKiroRedactedReasoning = () => {
+        if (!pendingKiroRedacted) return;
+        const previousBytes = pendingKiroRedactedBytes;
+        const encrypted = encodeReasoningEnvelope({ krc: pendingKiroRedacted });
+        const reservation = budget?.reserveTransient(bytesOf(encrypted), { kind: "reasoning" });
+        pendingKiroRedacted = undefined;
+        pendingKiroRedactedBytes = 0;
+        reservation?.commitRetained();
+        budget?.releaseRetained(previousBytes, { kind: "reasoning" });
+        const itemId = `rs_${uuid()}`;
+        const item = { type: "reasoning", id: itemId, summary: [] as never[], encrypted_content: encrypted };
+        emit("response.output_item.added", { output_index: outputIndex, item });
+        emit("response.output_item.done", { output_index: outputIndex, item });
+        retainFinishedItem(item as OutputItem, bytesOf(encrypted), "reasoning");
+        outputIndex++;
+      };
       // Full assistant text of a compaction turn (across message boundaries) — becomes the
       // synthetic compaction item's payload on done.
       let compactionText = "";
@@ -869,6 +892,12 @@ export function bridgeToResponsesSSE(
               pendingRedacted.push(event.data);
               break;
             }
+            case "kiro_redacted_reasoning": {
+              // Stash only — see flushKiroRedactedReasoning. One blob per turn, so last wins.
+              pendingKiroRedactedBytes = replaceRetainedString(pendingKiroRedactedBytes, event.data, "reasoning");
+              pendingKiroRedacted = event.data;
+              break;
+            }
             case "reasoning_raw_delta": {
               if (options?.hideThinkingSummary) {
                 ({ value: hiddenRawReasoningText, bytes: hiddenRawReasoningBytes } = appendString(
@@ -1039,6 +1068,9 @@ export function bridgeToResponsesSSE(
               // Redacted-only turns (or hidden thinking without a trailing signature event) still
               // need their envelope-only reasoning item so the blocks replay next turn.
               flushHiddenReasoningEnvelope();
+              // After every close above, so the blob lands AFTER the assistant message it belongs
+              // to and the parser's backwards pairing finds it.
+              flushKiroRedactedReasoning();
               if (options?.compaction) {
                 // Exactly one compaction item per turn; codex-rs takes the first and fatals on 0.
                 const item = {
@@ -1361,6 +1393,10 @@ function buildResponseJSONWithBudget(
   let batchSignatureBytes = 0;
   let batchRedacted: string[] = [];
   let batchRedactedBytes = 0;
+  // Kiro reasoning blob, held until after the trailing flushes so it lands AFTER the assistant
+  // message (see the streaming path). Retained because it outlives releaseTranslatedEvent.
+  let batchKiroRedacted: string | undefined;
+  let batchKiroRedactedBytes = 0;
   let currentToolCallId = "";
   let currentToolCallName = "";
   let currentToolCallArgs = "";
@@ -1528,6 +1564,16 @@ function buildResponseJSONWithBudget(
         }
         batchRedacted.push(e.data);
         break;
+      case "kiro_redacted_reasoning":
+        // Stash only — pushed after the trailing flushes. One blob per turn, so last wins.
+        {
+          const dataBytes = bytesOf(e.data);
+          budget?.chargeRetained(dataBytes, { kind: "reasoning" });
+          if (batchKiroRedactedBytes > 0) budget?.releaseRetained(batchKiroRedactedBytes, { kind: "reasoning" });
+          batchKiroRedactedBytes = dataBytes;
+        }
+        batchKiroRedacted = e.data;
+        break;
       case "reasoning_raw_delta":
         if (currentText) flushText("commentary");
         if (currentSummaryReasoning) flushSummaryReasoning();
@@ -1626,6 +1672,15 @@ function buildResponseJSONWithBudget(
   flushRawReasoning();
   // Open tool call on a failed/incomplete turn must not land as status:"completed".
   if (currentToolCallId) flushToolCall(errorEvent || incompleteEvent ? "incomplete" : "completed");
+  if (batchKiroRedacted) {
+    // pushOutput reserves the item itself and releases the retained raw blob it replaces.
+    pushOutput({
+      type: "reasoning", id: `rs_${uuid()}`, summary: [],
+      encrypted_content: encodeReasoningEnvelope({ krc: batchKiroRedacted }),
+    }, batchKiroRedactedBytes, "reasoning");
+    batchKiroRedacted = undefined;
+    batchKiroRedactedBytes = 0;
+  }
   // A truncated turn must never be installed as replacement history: emit the
   // compaction item only when the turn actually completed (#422).
   if (

src/claude/inbound.ts

@@ -106,6 +106,10 @@ function toolResultOutput(block: Rec): string | Rec[] {
       } else if (item.type === "image") {
         const img = imageBlockToInputImage(item);
         if (img) out.push(img);
+      } else if (item.type === "document") {
+        // Same marker as the user-message document case below: the model should see the
+        // attachment happened instead of an empty tool output.
+        out.push({ type: "input_text", text: `[document${typeof item.title === "string" ? `: ${item.title}` : ""}]` });
       }
     }
     if (isError) out.unshift({ type: "input_text", text: "[tool error]" });

src/claude/outbound.ts

@@ -187,6 +187,8 @@ interface OpenBlock {
   argsBufBytes?: number;
   webSearchArgsEmitted?: boolean;
   callId?: string;
+  /** Last reasoning part identity (item + summary/content index) seen by this thinking block. */
+  reasoningPartKey?: string;
 }
 
 /** Streaming: Responses SSE bytes -> Anthropic Messages SSE bytes. */
@@ -353,6 +355,21 @@ export function responsesSseToAnthropicSse(
           case "response.reasoning_text.delta": {
             if (typeof data.delta !== "string" || data.delta.length === 0) break;
             ensureBlock("thinking");
+            // The JSON path joins reasoning summary/content parts with "\n\n"
+            // (responsesJsonToAnthropicMessage); mirror that at part and item boundaries
+            // so multi-part summaries do not glue into one run-on paragraph. Frames
+            // without part indices produce a constant key and never get a separator.
+            const slot = eventName === "response.reasoning_summary_text.delta"
+              ? `s${String(data.summary_index)}`
+              : `c${String(data.content_index)}`;
+            const partKey = `${String(data.item_id)}:${slot}`;
+            if (open!.reasoningPartKey !== undefined && open!.reasoningPartKey !== partKey) {
+              emit("content_block_delta", {
+                type: "content_block_delta", index: open!.index,
+                delta: { type: "thinking_delta", thinking: "\n\n" },
+              });
+            }
+            open!.reasoningPartKey = partKey;
             emit("content_block_delta", {
               type: "content_block_delta", index: open!.index,
               delta: { type: "thinking_delta", thinking: data.delta },

src/lib/bun-stream-caps.ts

@@ -6,7 +6,8 @@
  * PR #32120, merged 2026-06-21). No RELEASED Bun version is proven to carry
  * that fix yet, so `MIN_FIXED_BUN_VERSION` is null: every runtime is
  * "known-bad" until a bundle-bump commit sets it. Windows no-rewrite traffic
- * follows this runtime/config decision. Darwin no-rewrite traffic stays on tee
+ * follows this runtime/config decision, preserving the explicit legacy-tee
+ * safety pin. Darwin no-rewrite traffic stays on tee
  * for `auto` regardless of runtime capability and reaches eager relay only via
  * explicit `streamMode: "eager-relay"` opt-in (see
  * devlog/_plan/260731_macos_rss_retention/100_darwin_eager_optin.md).

src/responses/parser.ts

@@ -416,6 +416,18 @@ export function parseRequest(body: unknown): OcxParsedRequest {
           : null;
         const thinkingText = envelope?.txt || text;
 
+        // Kiro reasoning round-trip: a krc-only item carries nothing renderable — it is provider
+        // state for the assistant turn that ALREADY closed, because Kiro emits its
+        // reasoningContentEvent at the END of a turn (after content AND tool calls, verified
+        // against kiro-cli 2.14.1/2.16.0). Folding it into the FOLLOWING turn like ordinary
+        // reasoning would attach turn N's blob to turn N+1, so attach it backwards instead. With
+        // no assistant turn to own it the blob is dropped rather than mis-paired.
+        if (envelope?.krc && thinkingText.length === 0) {
+          const previous = messages[messages.length - 1];
+          if (previous?.role === "assistant") previous.kiroRedactedReasoning = envelope.krc;
+          continue;
+        }
+
         // Native/non-ocxr1 encrypted-only reasoning is opaque here. Do not create a detached
         // assistant turn or invent replayable plaintext/signatures from the encrypted payload.
         if (thinkingText.length > 0) {

src/responses/reasoning-envelope.ts

@@ -24,6 +24,12 @@ export interface ReasoningEnvelope {
    * so replay needs it even though the visible summary was suppressed.
    */
   txt?: string;
+  /**
+   * Kiro `reasoningContentEvent.redactedContent`: a KMS-encrypted reasoning blob that is opaque to
+   * the proxy. Kiro's own CLI replays it on the matching `assistantResponseMessage` to preserve
+   * model reasoning across turns, so it round-trips here the same way a signature does.
+   */
+  krc?: string;
 }
 
 export function encodeReasoningEnvelope(envelope: ReasoningEnvelope): string {
@@ -45,7 +51,9 @@ export function decodeReasoningEnvelope(encryptedContent: string): ReasoningEnve
     }
     const txt = (parsed as { txt?: unknown }).txt;
     if (typeof txt === "string" && txt.length > 0) envelope.txt = txt;
-    return envelope.sig || envelope.red || envelope.txt ? envelope : null;
+    const krc = (parsed as { krc?: unknown }).krc;
+    if (typeof krc === "string" && krc.length > 0) envelope.krc = krc;
+    return envelope.sig || envelope.red || envelope.txt || envelope.krc ? envelope : null;
   } catch {
     return null;
   }