fix: merge duplicate-index entries in streaming tool_call deltas

[sharziki]

Summary

Fixes #3201

When the first streamed chunk contains multiple tool_calls entries with the same logical index (e.g. a tool-call name and its first argument fragment in one SSE event), the two early-return branches in accumulate_delta store the list as-is without merging by index. Later chunks merge into physical position 0 only, leaving the duplicate entry orphaned and producing invalid final JSON.

Root cause

Both early-return branches at the top of the per-key loop in _deltas.py (and the identical copy in _assistants.py) assign the delta list wholesale and skip the index-based merge logic that handles duplicate indexes in steady state:

if key not in acc:          # L8-10: first occurrence
    acc[key] = delta_value
    continue

acc_value = acc[key]
if acc_value is None:       # L13-15: None placeholder
    acc[key] = delta_value
    continue

Fix

Add _merge_indexed_list() that collapses list entries sharing the same index value before the list is stored for the first time. Applied in both _deltas.py and _assistants.py.

• Non-indexed lists (strings, ints) pass through unchanged
• Lists without an index field in the first entry pass through unchanged
• Lists with duplicate index values are merged using the existing accumulate_delta recursion

Test plan

• Verified with exact reproduction from issue (duplicate index: 0 in first chunk, acc_value is None)
• Verified with key not in acc variant
• Verified multiple distinct tool calls (different indexes) still work correctly
• Verified non-indexed string lists still accumulate normally
• Existing streaming test suite passes (35/35)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c2bcb65bb3

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Do not silently drop malformed indexed list entries

When _merge_indexed_list() is triggered (first item is a dict with index), entries that are not dicts or whose index is missing/non-int are skipped via continue, which silently removes data from the stream snapshot. This differs from the existing list-merge path in accumulate_delta() that explicitly raises on invalid indexed entries, so malformed-but-present payloads now become silent corruption instead of a detectable error. In streams where providers emit mixed or partially malformed tool_calls arrays, this can produce incomplete final messages without surfacing any failure.

Useful? React with 👍 / 👎.