fix(responses): keep translated output indexes unique

[ryanngit]

Summary

Allocate one monotonically increasing Responses output_index for every reasoning, assistant-message, and function-call item, then reuse that index across added, delta, done, and terminal events.

The fix covers both callable conversion paths:

• active translator pipeline in open-sse/translator
• legacy createResponsesApiTransformStream implementation

OpenAI Chat choice indexes and tool-call indexes are separate source namespaces. Reusing them directly can assign index 0 to reasoning, text, and the first function call in one response, producing an invalid item stream.

Scope

This recut contains output-index allocation only. Custom tools, deferred tool search, terminal payloads, cache details, Claude schema normalization, refusal-history repair, and model fallback are separate review surfaces documented in the split comment.

Verification

Tested on pristine v0.5.45 (6fcd2733):

• index parity suite: 3/3 passed
• adjacent terminal regression suite: 4/4 passed
• changed-file ESLint and git diff --check passed
• Gitleaks found no leaks

Fixes #2611.

[zierocode]

Fresh cross-check on 2026-07-23 at 30082c65c84c76a96f7212656710ee6990345765: both focused files pass, 12/12 tests total. This is the implementation I plan to rely on rather than opening a parallel custom-tool PR.

One provider-neutral gap remains outside the current suite: route-wide restoration of a synthetic _ide suffix to the unique client-declared tool name. This matters when a translated Responses, Chat, or Messages path returns emit_probe_ide even though the request declared only emit_probe.

The fail-closed contract is small:

// Unique declaration: restore the known synthetic alias.
declared = ["emit_probe"];
providerName = "emit_probe_ide";
expect(clientName).toBe("emit_probe");

// Ambiguous declaration: preserve the provider name.
declared = ["emit_probe", "emit_probe_ide"];
providerName = "emit_probe_ide";
expect(clientName).toBe("emit_probe_ide");

The same mapping should apply to Anthropic tool_use, Responses output items, and Chat tool_calls, while leaving explicitly declared _ide names untouched. This is separate from custom/freeform input preservation and should not block the current PR if you prefer it as a follow-up.

[ryanngit]

Added d1494630 to handle empty OpenAI function-tool schemas in the Claude bridge.

GitHub Copilot /v1/messages requires input_schema.type; Responses clients can legally send parameters: {}. The bridge now normalizes missing, empty, non-object, and array schemas to { type: "object", properties: {} }, while preserving schemas that already declare a root type.

Tests:

• responses-custom-tool-roundtrip.test.js: 11/11 passed
• Adjacent Claude/Responses matrix: 30 passed, 1 pre-existing unrelated failure in openai-to-claude.test.js
• The same unrelated failure reproduces at prior PR head e7d8350c without this commit.

[ryanngit]

Updated in 25bd48bb with centralized Claude tool-schema compatibility.

• Normalizes missing/invalid input schemas to object schemas.
• Removes unsupported root oneOf, anyOf, and allOf while preserving branch object properties.
• Merges required only for allOf; nested unions remain unchanged.
• Applies same helper during OpenAI-to-Claude translation and final Claude request formatting.
• Adds request-translation, final-format, allOf, nested-union, and invalid-root-metadata regressions.

Public compatibility references:

• CLIProxyAPI MIT implementation for root-union normalization: router-for-me/CLIProxyAPI@59aa35a
• LiteLLM Anthropic root-key filtering: BerriAI/litellm@56cda9f674d8

Verification:

• responses-custom-tool-roundtrip.test.js: 14/14 passed
• final Claude-format union guard: 1/1 passed
• provider canary: 25 tools with root oneOf at index 20 returned HTTP 200, a function call, one completion event, and one [DONE]

[ryanngit]

Updated with Claude refusal history closure (8780011a). Claude stop_reason=refusal now maps to content_filter, then emits visible assistant refusal text with one response.completed instead of leaving an unanswered failed turn that poisons every continuation. Other reasoning-only/protocol-only/whitespace terminals remain upstream_empty_response. Verification: Responses matrix 43/43, changed-file ESLint clean, git diff --check clean.

[ryanngit]

Updated with post-refusal continuation repair (f94a01bd).

A completed synthetic Claude refusal is still shown to the client, but its exact generated _refusal assistant item and immediately paired direct-user message are omitted from later Claude provider history. This prevents unrelated later turns from inheriting the refusal while preserving ordinary assistant text, current input, tool-bearing history, model/account selection, and the original refusal policy decision.

Verification:

• focused semantic terminal suite: 11/11
• broader Responses/Claude matrix: 45/45
• changed-file ESLint: clean
• git diff --check: clean

[ryanngit]

v0.5.45 split and recut

The prior 17-commit head mixed nine independent behaviors. It has been replaced by one reviewable output-index commit.

• Base: v0.5.45 (6fcd2733)
• New head: f705201d
• Merge state: CLEAN
• Active and legacy transformer parity: covered
• Focused tests: 7/7 passed
• ESLint, git diff --check, and Gitleaks: clean

Displaced commit map:

• custom/function tool round-trip: 7486afb6, 19947790, a163dfa8, 30082c65
• complete terminal payload: 8894c2bf
• cache details: f761272a
• deferred tool search: 69bf5fbd
• semantic-empty rejection: e7d8350c
• Claude schema normalization: d1494630, 25bd48bb
• Claude refusal closure/history pruning: 8780011a, f94a01bd, 1af56075, 04902276, 682e1cac
• GitHub Fable-to-Opus fallback: 8f0a4deb

Each retained behavior needs its own PR and focused tests. Refusal-history deletion and silent model fallback remain blocked on maintainer policy decisions; they are not part of this recut.