fix: merge duplicate-index entries in streaming delta accumulator (#3201)

[C1-BA-B1-F3]

Problem

When the first streamed chunk contains multiple tool_calls entries with the same index (e.g. id/name + start of arguments, both at index 0), the accumulator stores them as separate list entries. Subsequent chunks only merge into the first one at that position, leaving the duplicate stranded with incomplete arguments.

This produces invalid final JSON for tool call arguments — some argument fragments end up in a second list entry that nobody reads.

Root cause: The accumulate_delta function stores the first encounter of a list value directly via acc[key] = delta_value without checking for duplicate indices.

Fix

Add _normalize_indexed_list() that merges list entries sharing the same integer index key before first storage. Applied to both:

• src/openai/lib/streaming/_deltas.py (chat completions)
• src/openai/lib/streaming/_assistants.py (assistants API)

Test plan

• New test file: tests/test_accumulate_delta.py with 6 test cases covering:
  • Single index in first chunk (baseline)
  • Duplicate index in first chunk (the bug)
  • Duplicate index preserves type and id
  • Multiple distinct indices in first chunk
  • Multiple indices with duplicates in first chunk
  • Non-indexed lists unchanged

All 6 new tests pass. All 15 existing streaming tests pass.

Fixes #3201

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 75b895d3d4

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Normalize duplicate tool calls in the initial chunk

When the duplicate tool_calls entries arrive in the very first chat streaming chunk, this branch is never reached: ChatCompletionStreamState._accumulate_chunk() returns _convert_initial_chunk_into_snapshot() while the snapshot is None, and that path copies choice.delta.to_dict() directly. In that provider scenario, the first chunk still stores two entries with index == 0, so the next chunk merges into tool_calls[0] and leaves the argument prefix stranded in tool_calls[1]; apply the same normalization in the initial conversion path or route the first chunk through the accumulator.

Useful? React with 👍 / 👎.