🪢 fix: Flatten Text-Only Assistant Content To String

[misdirection]

Summary

For non-reasoning replies, formatAssistantMessage emits assistant content as
an array of text parts ([{ type: "text", text: "..." }]). The Chat Completions
spec allows both a string and an array here, so api.openai.com accepts it -
but strict OpenAI-compatible proxies validate literally and only allow a string,
rejecting the array with 400. Since the array first appears when a prior
assistant turn is replayed, this breaks the second turn of a conversation.

Observed against Langdock's OpenAI-compatible EU endpoint:

{"code":"invalid_type","expected":"string","path":["messages",1,"content"],
 "message":"Invalid input: expected string, received array"}

The fix unifies the two tail branches of formatAssistantMessage into a
single-pass buildAssistantContent helper: text-only content is joined into a
string (the canonical, universally valid shape), while content with any non-text
part keeps its array form. This also drops the now-redundant hasReasoning
branch, which did the same flattening - reasoning content is still attached via
createAIMessage. No behavior change for any other input class.

Reproduction

Reproduces only against a strict proxy (real OpenAI returns 200 for both). The
requests differ solely in assistant.content:

// 200 (post-fix)
"messages": [ {user}, {"role":"assistant","content":"Received."}, {user} ]

// 400 against a strict proxy (pre-fix)
"messages": [ {user}, {"role":"assistant","content":[{"type":"text","text":"Received."}]}, {user} ]

Scope

Strict proxies apply the same string-only constraint to the tool (and
function) roles, so a tool returning array/structured content would hit the
same 400. This PR intentionally covers only the assistant role - the
reported and most common case. The same flattening could be extended to those
roles if you'd like; flagging it here rather than widening this change
unprompted.

Change Type

• Bug fix (non-breaking change which fixes an issue)

Testing

• npx jest src/messages/ - 328 passed, 1 skipped
• npx tsc --noEmit / npx eslint "src/**/*.ts" - clean
• Added a regression test for multi-turn text-only content; updated 8 existing
  assertions that encoded the old array shape.
• Verified end-to-end through LibreChat against Langdock: with the fix
  assistant.content goes out as a string (200); reverting it reproduces the
  array and the 400 (negative control).

Test Configuration

• Node.js: v22.22.2
• Package: @librechat/agents@3.1.97

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• My changes generate no new warnings
• I have added tests that prove my fix is effective
• New and existing unit tests pass locally with my changes

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ce5248c2ec

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".