prefix-safe-json 0.4.4 → 0.4.5

package/dist/providers/openai.js

@@ -1,20 +1,122 @@
 import { OpenAICompatibleStreamAdapter } from "./openai-compatible.js";
-import { CONTENT_FILTERED_DIAGNOSTIC_CODE } from "../coordinator/diagnostic-codes.js";
+import { CONTENT_FILTERED_DIAGNOSTIC_CODE, TOOL_ARGUMENTS_AFTER_END_DIAGNOSTIC_CODE, DUPLICATE_TOOL_END_DIAGNOSTIC_CODE, INVALID_CHOICE_INDEX_DIAGNOSTIC_CODE, DUPLICATE_CHOICE_INDEX_DIAGNOSTIC_CODE, } from "../coordinator/diagnostic-codes.js";
+// Reason priority for aggregating legacy per-choice terminal reasons into
+// the ONE stream-wide reason, at adapter.finish() time. Mirrors
+// OpenAICompatibleStreamAdapter's own REASON_PRIORITY exactly (same
+// values, same order, same "worst wins" semantics: provider_error
+// strongest, complete weakest) - duplicated locally rather than
+// cross-imported so this file's production scope stays self-contained
+// (P4.3: src/providers/openai.ts only, openai-compatible.ts unmodified).
+const LEGACY_REASON_PRIORITY = [
+    "provider_error",
+    "unknown",
+    "network_error",
+    "length",
+    "cancelled",
+    "complete",
+];
+function mapLegacyFinishReason(providerReason) {
+    if (providerReason === "stop" || providerReason === "function_call" || providerReason === "tool_calls")
+        return "complete";
+    if (providerReason === "length")
+        return "length";
+    if (providerReason === "cancelled")
+        return "cancelled";
+    return "unknown";
+}
 export class OpenAIStreamAdapter {
     provider = "openai";
     sequence = 0;
     // Delegating tool_calls format to compatible adapter
     compatibleAdapter = new OpenAICompatibleStreamAdapter();
-    // For function_call (legacy)
+    // For the plural, new-style Chat Completions tool_calls format.
+    // Sticky stream-mode flag: once ANY chunk in this stream shows genuine
+    // plural tool_calls evidence (never inferred merely from finish_reason -
+    // see the "no prior plural tool call" regression), every later Chat
+    // Completions chunk continues through OpenAICompatibleStreamAdapter for
+    // the rest of the stream's lifetime, including a later chunk whose own
+    // delta carries no tool_calls at all - most commonly a separate
+    // empty-delta finish_reason terminal chunk (see
+    // test/integration/openai-official-sdk-lifecycle.test.ts for the exact
+    // shape openai@7.8.0's own SDK parser exposes for it). Before this flag
+    // existed, that terminal chunk failed the old per-chunk-only delegation
+    // check and fell through into the unrelated legacy function_call loop
+    // below, so the compatible adapter's own tracked call never received its
+    // tool_call_end (see test/providers/openai-tool-calls-terminal-routing.test.ts).
+    //
+    // Only ever set once this.hasLegacyFunctionCall is confirmed false for
+    // this stream - see push()'s own format-conflict check just above where
+    // this flag is read/written. A stream can only ever genuinely commit to
+    // ONE of the two Chat Completions tool-call shapes: conflicting evidence
+    // for the other format fails the whole stream closed instead of
+    // silently delegating to it (empirically confirmed necessary - a plural
+    // tool_calls chunk arriving after a legacy function_call had already
+    // started its own tracked call would otherwise delegate straight to
+    // OpenAICompatibleStreamAdapter, which has no knowledge of the outer
+    // adapter's legacy-tracked call and would happily track and correctly
+    // close the injected plural evidence as its own, fully independent,
+    // genuinely executable authority - see the "SYMMETRIC CASE" / "FORMAT-CONFLICT"
+    // regressions).
+    hasCompatibleToolCalls = false;
+    // For function_call (legacy). P4.3: choice-scoped identity and terminal
+    // ownership - the pre-existing design used ONE fixed, global sourceKey
+    // ("legacy-function-call") for every choice, so a genuine n>1 stream's
+    // independent per-choice function_call evidence could merge into one
+    // parser/identity (cross-choice argument/name injection - E-1), and a
+    // later, entirely separate choice's evidence could collide with and
+    // poison an earlier, already-legitimate choice's call (E-2). Ports the
+    // SAME choice-local architecture already proven for the plural path in
+    // OpenAICompatibleStreamAdapter (see that file's own class-level
+    // lifecycle-contract doc) onto this legacy singular wire shape, which
+    // has no array of tool_calls to key off - so per-choice state is
+    // tracked directly here instead of delegating.
+    //
+    // hasLegacyFunctionCall keeps its ORIGINAL meaning unchanged: "has ANY
+    // choice in this stream ever shown legacy function_call evidence" -
+    // still used by the mixed-format conflict check above and by finish()'s
+    // own branch-selection between this path and compatibleAdapter's.
     hasLegacyFunctionCall = false;
-    legacySourceKey = "legacy-function-call";
+    // choiceIndex -> is THAT choice's own legacy call's argument stream
+    // currently open (started, not yet closed by ITS OWN finish_reason).
+    legacyChoiceOpen = new Map();
+    // choiceIndex -> that choice's own first-recorded terminal StreamEndReason.
+    // Never overwritten once set for a given choice - see Phase 10 handling
+    // below. Read by finish() to aggregate one stream-wide reason.
+    legacyChoiceTerminalReasons = new Map();
+    // The same choices' own RAW (pre-normalization) finish_reason strings,
+    // keyed identically - observability only, mirrored onto the aggregated
+    // provider_stream_end's own providerReason (see decide.ts's own doc
+    // comment: "never consulted by the decision logic itself").
+    legacyChoiceTerminalProviderReasons = new Map();
     // For Responses API
     accumulatedArguments = new Map();
     receivedDeltas = new Set();
+    // Item-local terminal state (P4.1 / F-2): every output-item id that has
+    // already received its own `response.output_item.done`. Uses the exact
+    // same identity as this adapter's own `output-item:{id}` sourceKey - no
+    // second identity system. `output_item.done` closes the argument stream
+    // choice-locally (tool_call_end), but that alone does not stop the
+    // coordinator from silently merging a LATER, real
+    // `response.function_call_arguments.delta`/`.done` for the SAME item_id
+    // into the same still-"collecting" call (handleCallEnd never changes
+    // call.status - see coordinator.ts). Recognized argument evidence for an
+    // item already in this set is therefore hardened below instead of merged.
+    doneOutputItemIds = new Set();
     finished = false;
     push(rawEvent) {
-        if (this.finished)
-            return [];
+        // No `finished` early return here: silently dropping every event after
+        // the first terminal meant a late argument delta, a conflicting or
+        // duplicate finish, or a provider error that arrived even one raw event
+        // late never reached the coordinator at all - not even as a diagnostic -
+        // so an already-decided call's authority could never be revoked by it.
+        // Every case below that has ITS OWN meaningful post-terminal handling
+        // (a genuine argument delta, a second output_item.added, a second
+        // stream-terminal) still fires exactly as if the stream were open; the
+        // coordinator's own `isFinished` gate (coordinator.ts's `push()`) is
+        // what turns each of those into a sticky, stream-wide
+        // AUTHORITY_PROTOCOL_VIOLATION_CODES diagnostic once it actually
+        // receives them. Mirrors the fix already applied to AiSdkStreamAdapter
+        // for the same invariant (GHSA-3xpw-9694-2xxp).
         const events = [];
         if (!rawEvent || typeof rawEvent !== "object") {
             events.push({
@@ -63,44 +165,84 @@ export class OpenAIStreamAdapter {
                 });
             }
             else if (chunk.type === "response.function_call_arguments.delta" && chunk.item_id && chunk.delta) {
-                this.receivedDeltas.add(chunk.item_id);
-                const acc = this.accumulatedArguments.get(chunk.item_id) ?? "";
-                this.accumulatedArguments.set(chunk.item_id, acc + chunk.delta);
-                events.push({
-                    type: "tool_call_arguments_delta",
-                    sequence: ++this.sequence,
-                    provider: this.provider,
-                    callRef: { sourceKey: `output-item:${chunk.item_id}` },
-                    delta: chunk.delta,
-                });
-            }
-            else if (chunk.type === "response.function_call_arguments.done" && chunk.item_id && chunk.arguments !== undefined) {
-                const acc = this.accumulatedArguments.get(chunk.item_id) ?? "";
-                const hasDeltas = this.receivedDeltas.has(chunk.item_id);
-                if (!hasDeltas && chunk.arguments.length > 0) {
-                    // No deltas, just final arguments.
+                if (this.doneOutputItemIds.has(chunk.item_id)) {
+                    // Item-local post-terminal evidence (P4.1 / F-2): this exact
+                    // item already recorded its own output_item.done. Never merge
+                    // further argument evidence as normal - that would silently
+                    // mutate (and, if left structurally unclosed at done time,
+                    // silently CLOSE) an already-ended item's value with no record
+                    // of the anomaly. Same diagnostic code AiSdkStreamAdapter/
+                    // OpenAICompatibleStreamAdapter already use for materially
+                    // identical semantics ("tool argument evidence arrived after
+                    // that call's end"), attributed to the item's own real,
+                    // already-tool_call_start'd sourceKey so the coordinator
+                    // resolves it to the actual call, not a phantom identity.
+                    events.push({
+                        type: "provider_diagnostic",
+                        sequence: ++this.sequence,
+                        provider: this.provider,
+                        callRef: { sourceKey: `output-item:${chunk.item_id}` },
+                        code: TOOL_ARGUMENTS_AFTER_END_DIAGNOSTIC_CODE,
+                        severity: "error",
+                        message: `Tool call argument evidence for item ${chunk.item_id} arrived after that item's own output_item.done`,
+                    });
+                }
+                else {
+                    this.receivedDeltas.add(chunk.item_id);
+                    const acc = this.accumulatedArguments.get(chunk.item_id) ?? "";
+                    this.accumulatedArguments.set(chunk.item_id, acc + chunk.delta);
                     events.push({
                         type: "tool_call_arguments_delta",
                         sequence: ++this.sequence,
                         provider: this.provider,
                         callRef: { sourceKey: `output-item:${chunk.item_id}` },
-                        delta: chunk.arguments,
+                        delta: chunk.delta,
                     });
-                    this.accumulatedArguments.set(chunk.item_id, chunk.arguments);
                 }
-                else if (hasDeltas && chunk.arguments !== acc) {
+            }
+            else if (chunk.type === "response.function_call_arguments.done" && chunk.item_id && chunk.arguments !== undefined) {
+                if (this.doneOutputItemIds.has(chunk.item_id)) {
+                    // Same item-local post-terminal hardening as the `.delta` branch
+                    // above, for the `.done` (final-arguments) shape specifically.
                     events.push({
                         type: "provider_diagnostic",
                         sequence: ++this.sequence,
                         provider: this.provider,
                         callRef: { sourceKey: `output-item:${chunk.item_id}` },
-                        code: "E_FINAL_ARGUMENTS_CONFLICT",
+                        code: TOOL_ARGUMENTS_AFTER_END_DIAGNOSTIC_CODE,
                         severity: "error",
-                        message: "Final arguments do not match accumulated deltas",
+                        message: `Tool call argument evidence for item ${chunk.item_id} arrived after that item's own output_item.done`,
                     });
                 }
+                else {
+                    const acc = this.accumulatedArguments.get(chunk.item_id) ?? "";
+                    const hasDeltas = this.receivedDeltas.has(chunk.item_id);
+                    if (!hasDeltas && chunk.arguments.length > 0) {
+                        // No deltas, just final arguments.
+                        events.push({
+                            type: "tool_call_arguments_delta",
+                            sequence: ++this.sequence,
+                            provider: this.provider,
+                            callRef: { sourceKey: `output-item:${chunk.item_id}` },
+                            delta: chunk.arguments,
+                        });
+                        this.accumulatedArguments.set(chunk.item_id, chunk.arguments);
+                    }
+                    else if (hasDeltas && chunk.arguments !== acc) {
+                        events.push({
+                            type: "provider_diagnostic",
+                            sequence: ++this.sequence,
+                            provider: this.provider,
+                            callRef: { sourceKey: `output-item:${chunk.item_id}` },
+                            code: "E_FINAL_ARGUMENTS_CONFLICT",
+                            severity: "error",
+                            message: "Final arguments do not match accumulated deltas",
+                        });
+                    }
+                }
             }
             else if (chunk.type === "response.output_item.done" && chunk.item?.id) {
+                this.doneOutputItemIds.add(chunk.item.id);
                 events.push({
                     type: "tool_call_end",
                     sequence: ++this.sequence,
@@ -177,67 +319,263 @@ export class OpenAIStreamAdapter {
             }
             return events;
         }
-        // Check if it's the standard tool_calls format
-        if (chunk.choices?.[0]?.delta?.tool_calls !== undefined) {
+        // This chunk's OWN evidence for each Chat Completions tool-call format,
+        // independent of any mode already committed to by an earlier chunk.
+        // Checked across every choice, not only choices[0]: a multi-choice
+        // (n>1) stream's tool call can legitimately live in a non-zero choice.
+        const hasPluralEvidence = Array.isArray(chunk.choices) && chunk.choices.some(choice => choice.delta?.tool_calls !== undefined);
+        const hasLegacyEvidence = Array.isArray(chunk.choices) && chunk.choices.some(choice => choice.delta?.function_call !== undefined);
+        // A stream can only ever genuinely be ONE Chat Completions tool-call
+        // format. Conflicting evidence - both shapes in the SAME chunk, or one
+        // format's evidence arriving after the OTHER was already committed to
+        // by an earlier chunk - must never silently pick a winner while
+        // letting the original, already-legitimate call keep its authority
+        // (that was the exact gap the earlier hasLegacyFunctionCall-only cross
+        // check left open - see the SYMMETRIC CASE / plural-then-singular
+        // regressions). It fails the WHOLE stream closed instead: same as
+        // every other meaningful evidence this method recognizes, there is no
+        // `if (this.finished) return` guard here, so this branch also runs -
+        // and correctly revokes an already-live, not-yet-consumed decision -
+        // when the conflicting evidence arrives AFTER a clean terminal already
+        // fired. That revocation is the coordinator's own isFinished protocol
+        // converting this second stream-end-shaped event into its sticky,
+        // stream-wide diagnostic (GHSA-3xpw-9694-2xxp) - this adapter does not
+        // track post-conflict state itself beyond the ordinary `finished` flag.
+        if ((hasPluralEvidence && hasLegacyEvidence) || (hasPluralEvidence && this.hasLegacyFunctionCall) || (hasLegacyEvidence && this.hasCompatibleToolCalls)) {
+            events.push({
+                type: "provider_stream_end",
+                sequence: ++this.sequence,
+                provider: this.provider,
+                reason: "provider_error",
+                providerReason: "mixed_tool_call_formats",
+            });
+            this.finished = true;
+            return events;
+        }
+        if (hasPluralEvidence) {
+            this.hasCompatibleToolCalls = true;
+        }
+        if (this.hasCompatibleToolCalls) {
             // Delegate to OpenAICompatibleStreamAdapter
             const compatibleEvents = this.compatibleAdapter.push(rawEvent);
             // Map provider name to "openai"
             return compatibleEvents.map(e => ({ ...e, provider: this.provider, sequence: ++this.sequence }));
         }
-        // Handle legacy function_call format
+        // Handle legacy function_call format (P4.3: choice-scoped - see the
+        // field-level doc comments above for the full rationale).
         if (Array.isArray(chunk.choices)) {
+            // Duplicate-index detection within one chunk - mirrors
+            // OpenAICompatibleStreamAdapter's own choiceCounts pre-pass exactly,
+            // reused because the underlying choices[]/index wire structure is
+            // identical between plural and legacy singular tool-call shapes.
+            const choiceCounts = new Map();
             for (const choice of chunk.choices) {
-                if (choice.delta?.function_call) {
-                    const fc = choice.delta.function_call;
-                    if (!this.hasLegacyFunctionCall) {
-                        this.hasLegacyFunctionCall = true;
+                if (Number.isInteger(choice.index) && choice.index >= 0) {
+                    const choiceIndex = choice.index;
+                    choiceCounts.set(choiceIndex, (choiceCounts.get(choiceIndex) ?? 0) + 1);
+                }
+            }
+            for (const choice of chunk.choices) {
+                // Choice-index validity (Phase 5): a legacy function_call-bearing
+                // choice with no trustworthy explicit index must never be
+                // silently assumed to be choice 0 - global, unattributable, fails
+                // the whole stream closed rather than guessing (same fail-closed
+                // posture INVALID_CHOICE_INDEX_DIAGNOSTIC_CODE already has for
+                // the plural path - reused rather than inventing a new code,
+                // since the identity ambiguity it represents is identical here).
+                // A finish_reason-only choice (no function_call evidence at all)
+                // with an invalid index carries no legacy identity to protect and
+                // is simply skipped, not diagnosed - out of this fix's scope.
+                if (!Number.isInteger(choice.index) || choice.index < 0) {
+                    if (choice.delta?.function_call) {
                         events.push({
-                            type: "tool_call_start",
+                            type: "provider_diagnostic",
                             sequence: ++this.sequence,
                             provider: this.provider,
-                            callRef: { sourceKey: this.legacySourceKey },
-                            toolIndex: 0,
-                            name: fc.name,
+                            code: INVALID_CHOICE_INDEX_DIAGNOSTIC_CODE,
+                            severity: "error",
+                            message: "legacy function_call choice.index is missing, non-integer, or negative; choice identity is ambiguous",
                         });
                     }
-                    else if (fc.name) {
+                    continue;
+                }
+                const choiceIndex = choice.index;
+                if ((choiceCounts.get(choiceIndex) ?? 0) > 1) {
+                    if (choice.delta?.function_call) {
                         events.push({
-                            type: "tool_call_name_delta",
+                            type: "provider_diagnostic",
                             sequence: ++this.sequence,
                             provider: this.provider,
-                            callRef: { sourceKey: this.legacySourceKey },
-                            delta: fc.name,
+                            callRef: { sourceKey: `legacy-choice:${choiceIndex}` },
+                            code: DUPLICATE_CHOICE_INDEX_DIAGNOSTIC_CODE,
+                            severity: "error",
+                            message: `choice.index ${choiceIndex} is duplicated in one provider event`,
                         });
                     }
-                    if (fc.arguments) {
+                    continue;
+                }
+                const sourceKey = `legacy-choice:${choiceIndex}`;
+                if (choice.delta?.function_call) {
+                    const fc = choice.delta.function_call;
+                    const isKnownChoice = this.legacyChoiceOpen.has(choiceIndex);
+                    // Choice-local post-terminal evidence: this exact choice's
+                    // legacy call was already closed (its own finish_reason, or a
+                    // direct adapter.finish() force-close). Guards BOTH the
+                    // call-start path below AND the name/arguments-delta path -
+                    // `isKnownChoice` alone (the original, insufficient check) only
+                    // ever distinguished "genuinely new call" from "continuing
+                    // call," never "continuing an OPEN call" from "evidence for an
+                    // already-CLOSED one," which would otherwise let a late
+                    // argument/name delta silently mutate (and, if left
+                    // structurally unclosed at close time, silently CLOSE) an
+                    // already-ended choice's value with no record of the anomaly -
+                    // the same class of gap already hardened for OpenAI Responses
+                    // (P4.1) and Anthropic (P4.2). Same diagnostic code those two
+                    // use for materially identical semantics.
+                    if (isKnownChoice && this.legacyChoiceOpen.get(choiceIndex) === false) {
                         events.push({
-                            type: "tool_call_arguments_delta",
+                            type: "provider_diagnostic",
                             sequence: ++this.sequence,
                             provider: this.provider,
-                            callRef: { sourceKey: this.legacySourceKey },
-                            delta: fc.arguments,
+                            callRef: { sourceKey },
+                            code: TOOL_ARGUMENTS_AFTER_END_DIAGNOSTIC_CODE,
+                            severity: "error",
+                            message: `Tool call evidence for choice ${choiceIndex} arrived after that choice's own finish_reason`,
                         });
                     }
+                    else {
+                        if (!isKnownChoice) {
+                            this.hasLegacyFunctionCall = true;
+                            this.legacyChoiceOpen.set(choiceIndex, true);
+                            events.push({
+                                type: "tool_call_start",
+                                sequence: ++this.sequence,
+                                provider: this.provider,
+                                callRef: { sourceKey },
+                                toolIndex: choiceIndex,
+                                name: fc.name,
+                            });
+                        }
+                        else if (fc.name) {
+                            events.push({
+                                type: "tool_call_name_delta",
+                                sequence: ++this.sequence,
+                                provider: this.provider,
+                                callRef: { sourceKey },
+                                delta: fc.name,
+                            });
+                        }
+                        if (fc.arguments) {
+                            events.push({
+                                type: "tool_call_arguments_delta",
+                                sequence: ++this.sequence,
+                                provider: this.provider,
+                                callRef: { sourceKey },
+                                delta: fc.arguments,
+                            });
+                        }
+                    }
                 }
                 if (choice.finish_reason != null) {
-                    let reason = "unknown";
-                    if (choice.finish_reason === "stop" || choice.finish_reason === "function_call" || choice.finish_reason === "tool_calls") {
-                        reason = "complete";
+                    const alreadyTerminal = this.legacyChoiceTerminalReasons.has(choiceIndex);
+                    if (alreadyTerminal) {
+                        // Duplicate/conflicting terminal for the SAME choice (Phase
+                        // 10): never silently overwrite the first-recorded reason.
+                        // Exactly DUPLICATE_TOOL_END_DIAGNOSTIC_CODE's own documented
+                        // meaning ("more than one end part for the same call") - this
+                        // choice's own finish_reason IS what closes its legacy call,
+                        // so a second one is genuinely a repeated end signal for that
+                        // same call. Attributed to this choice's own real sourceKey
+                        // only: exact attribution is possible here, so per this
+                        // fix's own design this stays choice-local and does not
+                        // poison unrelated sibling choices - distinct from the
+                        // separate, already-approved stream-wide design used for the
+                        // analogous plural-path anomaly in openai-compatible.ts.
+                        events.push({
+                            type: "provider_diagnostic",
+                            sequence: ++this.sequence,
+                            provider: this.provider,
+                            callRef: { sourceKey },
+                            code: DUPLICATE_TOOL_END_DIAGNOSTIC_CODE,
+                            severity: "error",
+                            message: `choice ${choiceIndex} received a second finish_reason ("${choice.finish_reason}") after already terminating`,
+                        });
                     }
-                    else if (choice.finish_reason === "length") {
-                        reason = "length";
+                    else if (this.finished) {
+                        // Post-blocker-fix: this choice was already closed by a PRIOR
+                        // adapter.finish() call - force-closed via that method's own
+                        // loop, which never touches legacyChoiceTerminalReasons (only
+                        // legacyChoiceOpen - see finish()) - so `alreadyTerminal` above
+                        // is false even though the adapter has already globally
+                        // terminated. The ORIGINAL bug: silently recording this
+                        // reason and returning nothing meant this late, recognized
+                        // terminal evidence never reached the coordinator at all -
+                        // not even as a diagnostic - so an already-computed, still-
+                        // UNCONSUMED execute decision could never be revoked by it.
+                        // Exactly the class of gap this file's own push() doc comment
+                        // already promises never to reintroduce ("No `finished` early
+                        // return here: silently dropping every event after the first
+                        // terminal meant ... an already-decided call's authority
+                        // could never be revoked by it").
+                        //
+                        // Reuses DUPLICATE_TOOL_END_DIAGNOSTIC_CODE rather than
+                        // inventing a new mechanism: this choice's call already has
+                        // an end (finish()'s own synthetic close), so a further,
+                        // later terminal for the SAME choice is genuinely a second
+                        // end signal for that same call - semantically identical to
+                        // the in-stream duplicate-finish_reason case above. No new
+                        // security mechanism is implemented here: coordinator.ts's
+                        // own push() unconditionally converts ANY event arriving
+                        // after its OWN isFinished flag into a sticky, stream-wide
+                        // AUTHORITY_PROTOCOL_VIOLATION_CODES diagnostic
+                        // (EVENT_AFTER_STREAM_END_DIAGNOSTIC_CODE, or
+                        // TERMINAL_REASON_CONFLICT_DIAGNOSTIC_CODE for a genuinely
+                        // conflicting provider_stream_end) regardless of this event's
+                        // own code - this only needs to stop being silence. Never
+                        // reopens/modifies legacyChoiceOpen/legacyChoiceTerminalReasons
+                        // (that state is frozen the moment finish() ran) and never
+                        // emits a second normal tool_call_end, which would look like
+                        // ordinary in-stream evidence rather than a flagged anomaly.
+                        //
+                        // Only emitted when this exact choice had prior tracked
+                        // activity (legacyChoiceOpen.has) - a finish_reason for a
+                        // choice index that was never referenced at all carries no
+                        // existing authority to protect, and is left exactly as
+                        // before (silently ignored, matching the invalid-index
+                        // choice's own "no identity to protect" design elsewhere in
+                        // this file).
+                        if (this.legacyChoiceOpen.has(choiceIndex)) {
+                            events.push({
+                                type: "provider_diagnostic",
+                                sequence: ++this.sequence,
+                                provider: this.provider,
+                                callRef: { sourceKey },
+                                code: DUPLICATE_TOOL_END_DIAGNOSTIC_CODE,
+                                severity: "error",
+                                message: `choice ${choiceIndex} received a finish_reason ("${choice.finish_reason}") after the adapter had already globally terminated via finish()`,
+                            });
+                        }
                     }
-                    else if (choice.finish_reason === "cancelled") {
-                        reason = "cancelled";
+                    else {
+                        const reason = mapLegacyFinishReason(choice.finish_reason);
+                        this.legacyChoiceTerminalReasons.set(choiceIndex, reason);
+                        this.legacyChoiceTerminalProviderReasons.set(choiceIndex, choice.finish_reason);
+                        if (this.legacyChoiceOpen.get(choiceIndex)) {
+                            events.push({
+                                type: "tool_call_end",
+                                sequence: ++this.sequence,
+                                provider: this.provider,
+                                callRef: { sourceKey },
+                                reason,
+                                providerReason: choice.finish_reason,
+                            });
+                            this.legacyChoiceOpen.set(choiceIndex, false);
+                        }
+                        // No provider_stream_end here, and `this.finished` stays
+                        // false: this choice finishing is not, by itself, proof the
+                        // whole provider stream (which may still have other choices
+                        // actively generating) has ended. See finish().
                     }
-                    events.push({
-                        type: "provider_stream_end",
-                        sequence: ++this.sequence,
-                        provider: this.provider,
-                        reason,
-                        providerReason: choice.finish_reason,
-                    });
-                    this.finished = true;
                 }
             }
         }
@@ -251,13 +589,67 @@ export class OpenAIStreamAdapter {
         if (compatibleEvents.length > 0 && !this.hasLegacyFunctionCall) {
             return compatibleEvents.map(e => ({ ...e, provider: this.provider, sequence: ++this.sequence }));
         }
-        return [{
-                type: "provider_stream_end",
-                sequence: ++this.sequence,
-                provider: this.provider,
-                reason: meta?.reason ?? "unknown",
-                providerReason: meta?.providerReason,
-            }];
+        const events = [];
+        // Close any legacy choice whose call never got a chance to report its
+        // own finish_reason at all before the caller's raw iterator ended.
+        // Marks each one closed (mirrors OpenAICompatibleStreamAdapter's own
+        // `openSet.clear()` in its analogous finish() loop) - without this, a
+        // stray later push() for the SAME choice (still reachable: there is
+        // deliberately no top-of-push() `finished` guard, see GHSA-3xpw-9694-
+        // 2xxp) would find `legacyChoiceOpen` still true and emit a SECOND,
+        // spurious tool_call_end for a call this adapter already closed here.
+        for (const [choiceIndex, open] of this.legacyChoiceOpen.entries()) {
+            if (open) {
+                events.push({
+                    type: "tool_call_end",
+                    sequence: ++this.sequence,
+                    provider: this.provider,
+                    callRef: { sourceKey: `legacy-choice:${choiceIndex}` },
+                    reason: meta?.reason ?? "unknown",
+                    providerReason: meta?.providerReason,
+                });
+                this.legacyChoiceOpen.set(choiceIndex, false);
+            }
+        }
+        const { reason, providerReason } = this.aggregateLegacyTermination(meta?.reason, meta?.providerReason);
+        events.push({
+            type: "provider_stream_end",
+            sequence: ++this.sequence,
+            provider: this.provider,
+            reason,
+            providerReason,
+        });
+        return events;
+    }
+    // One stream-wide StreamEndReason from every legacy choice-local reason
+    // recorded so far, plus any caller-supplied override - "complete" only
+    // if every single one of them was "complete". Mirrors
+    // OpenAICompatibleStreamAdapter's own aggregateTermination() exactly
+    // (see LEGACY_REASON_PRIORITY above for why it is duplicated locally
+    // rather than cross-imported).
+    aggregateLegacyTermination(callerReason, callerProviderReason) {
+        let worst;
+        let worstRank = Infinity;
+        let worstProviderReason;
+        for (const [choiceIndex, candidate] of this.legacyChoiceTerminalReasons.entries()) {
+            const rank = LEGACY_REASON_PRIORITY.indexOf(candidate);
+            if (rank !== -1 && rank < worstRank) {
+                worst = candidate;
+                worstRank = rank;
+                worstProviderReason = this.legacyChoiceTerminalProviderReasons.get(choiceIndex);
+            }
+        }
+        if (callerReason !== undefined) {
+            const rank = LEGACY_REASON_PRIORITY.indexOf(callerReason);
+            if (rank !== -1 && rank < worstRank) {
+                worst = callerReason;
+                worstRank = rank;
+                worstProviderReason = callerProviderReason;
+            }
+        }
+        if (worst === undefined)
+            return { reason: "unknown", providerReason: callerProviderReason };
+        return { reason: worst, providerReason: worstProviderReason };
     }
 }
 //# sourceMappingURL=openai.js.map