fix(agent): split concatenated parallel tool call arguments from Gemini OpenAI-compat endpoint (#62937)

[webtecnica]

Summary

When the configured provider is Gemini via its OpenAI-compatible endpoint, the model frequently emits multiple parallel function calls concatenated into a single tool_calls entry. Google's OpenAI-compat layer does not increment tool_calls[].index for parallel calls, so the streaming assembler concatenates all argument fragments into one string. json.loads() rejects with "Extra data", and _repair_tool_call_arguments replaces everything with {}. Tools execute with empty arguments — silently doing nothing.

Changes

1. agent/chat_completion_helpers.py: When json.loads fails with "Extra data" on tool call arguments, uses json.JSONDecoder().raw_decode to peel off each complete object and emit one mock_tool_call per object (with distinct _split_N ids)
2. agent/message_sanitization.py: Added Repair pass 5 — when payload contains ≥2 concatenated JSON objects, returns the first decoded object instead of {}

Verification

99 tool_call/streaming/repair tests pass.

[alt-glitch]

This was generated by AI during triage.

Fix PR for #62937. This targets the Gemini OpenAI-compat chat_completions streaming path (raw_decode split + a new repair pass), which is a distinct code path from the closed native-adapter concatenated-JSON cluster (#25333/#25347/#36039/#33402) and the open native-adapter #54355 -- related, not a duplicate.

[teknium1]

Thanks for isolating the Gemini OpenAI-compatible streaming path, and for documenting why it differs from the native-adapter reports linked in the discussion.

This automated hermes-sweeper review is closing this as not planned because the proposed mechanism matches the standing model-output-repair policy: Hermes does not add new repair or reconstruction passes for malformed model tool-call arguments.

• agent/chat_completion_helpers.py:2319 does currently concatenate argument fragments for a repeated stream index, and :2362-2377 routes invalid JSON through repair handling.
• PR commit 98ae9d57dc8c0318ec17c770bfb4c31e64190bb4 adds a JSONDecoder.raw_decode() loop that reconstructs multiple tool calls and a new sanitization repair pass that selects decoded content.
• The standing policy rejects new parsing/reconstruction of malformed tool output, including concatenated/mangled tool-call JSON, even where the observed failure is real.

A provider-side correction that emits distinct OpenAI-compatible tool-call indices/entries is the appropriate direction.

---

Closed as not-planned per standing maintainer policy (model-output-repair). This is a design-direction decision, not a code-quality judgment — see the Contribution Rubric in AGENTS.md for what the project is looking for. If you believe this policy was misapplied to your change, comment here and a maintainer will take a look.