fix(nim): NVIDIA NIM provider hardening and response normalization

[Zireael]

Issue for this PR

Closes #10884, #6290, #19947, #25786, #13900, #24264, #22493

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

NVIDIA NIM returns OpenAI-incompatible responses in several ways that crash agent sessions: numeric or missing tool_call.id, malformed JSON arguments, thinking tokens leaking into message content, and HTTP 200 payloads that look like errors but pass through as valid responses. Reasoning models like DeepSeek V4 also hang without chat_template_kwargs in the right place.

This adds a defense layer in the provider fetch pipeline that fixes these at the transport level. It works in three parts: normalize outgoing requests (inject kwargs, fix model IDs), normalize incoming responses (fix IDs, repair JSON, strip thinking leakage), and retry transient failures (HTTP 429, 5xx, HTTP 200 errors).

The normalization intercept runs after the raw HTTP response comes back but before the AI SDK validates it. Response reconstruction preserves original headers. Streaming passes through untouched.

How did you verify your code works?

58 unit tests cover the normalizer, JSON repair, request enrichment, and retry wrapper. The full provider test suite passes (393/394, one pre-existing timeout in unrelated plugin config tests). Typecheck is clean.

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Related PRs Found

Here are PRs addressing overlapping issues that PR #27830 consolidates:

1. fix(provider): inject chat_template_kwargs for Nvidia NIM deepseek-v4 models #24262 - fix(provider): inject chat_template_kwargs for Nvidia NIM deepseek-v4 models

   • Addresses the chat_template_kwargs injection for reasoning models (Ring 1 of PR fix(nim): NVIDIA NIM provider hardening and response normalization #27830)

2. fix(provider): coerce numeric tool call IDs for OpenAI-compatible providers #24026 - fix(provider): coerce numeric tool call IDs for OpenAI-compatible providers

   • Handles numeric tool_call.id coercion (part of Ring 2 response normalization)

3. fix: generate fallback tool call IDs for providers missing id field #12585 - fix: generate fallback tool call IDs for providers missing id field

   • Addresses missing tool_call.id values (overlaps with PR fix(nim): NVIDIA NIM provider hardening and response normalization #27830's Ring 2 defense)

4. fix(provider): propagate options.extraBody for openai-compatible providers #26233 - fix(provider): propagate options.extraBody for openai-compatible providers

   • Related to request body enrichment for NIM providers

Note: PR #27830 appears to be a comprehensive consolidation that addresses all these issues in a unified defense layer (nim-defense.ts). These older PRs may be superseded or complementary to the current PR's approach.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍