fix(prompt): preserve non-string content in message placeholders

[winklemad]

Summary

ChatPromptClient.compile() passes every placeholder-history message's content straight into the string template parser (TemplateParser.compile_template), which assumes a str. Two standard OpenAI message shapes therefore crash instead of compiling:

• content=None — an assistant tool-call turn — raises TypeError: object of type 'NoneType' has no len() (compile_template does while curr_idx < len(content)).
• content=[...] — a multimodal / vision message (a list of content parts) — raises AttributeError: 'list' object has no attribute 'find'.

Both are exactly the shapes the message-placeholder feature is meant to carry when replaying a real conversation:

prompt.compile(message_history=[
    {"role": "user", "content": "What's the weather in SF?"},
    {"role": "assistant", "content": None,
     "tool_calls": [{"id": "call_1", "type": "function",
                     "function": {"name": "get_weather", "arguments": "{...}"}}]},
    {"role": "tool", "content": "72F sunny", "tool_call_id": "call_1"},
])

get_langchain_prompt(...) hits the same path since it calls compile() internally.

The existing test_tool_calls_preservation_in_message_placeholder covers a tool-call turn but sets content="", so the real None shape (and multimodal list content) were never exercised.

Fix

Only run the template parser when the content is a string; pass non-string content (None, or a multimodal list) through unchanged. One guard covers both cases and preserves tool_calls and every other field on the message.

Verification

• uv run --frozen pytest tests/unit/test_prompt_compilation.py — all pass, including a new regression test (test_placeholder_message_with_non_string_content) that fails on main with the TypeError above and passes with this change.
• Full tests/unit offline suite passes locally (only pre-existing unrelated collection errors from optional integration deps / server-credentialed fixtures).
• uv run --frozen ruff check langfuse/model.py and ruff format --check — no new findings.
• uv run --frozen mypy langfuse/model.py --no-error-summary — clean.

Confidence Score: 4/5

The PR is not yet safe to merge because get_langchain_prompt() still crashes on the non-string placeholder values this change intends to support.

Summary

This PR updates chat-placeholder compilation to preserve non-string message content instead of sending it through the string template parser.

• Preserves None content used by assistant tool-call turns.
• Preserves list-based multimodal content.
• Adds direct compile() regression coverage for both shapes.
• The corresponding LangChain conversion path remains string-only and still fails for these values.

_{Reviews (1) · Last reviewed commit: "fix(prompt): preserve non-string content..."}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

LangChain conversion still crashes

Non-string placeholder content now survives compile(), but get_langchain_prompt() then passes every compiled message’s content to the string-only _get_langchain_prompt_string(). When the history contains an assistant tool-call turn with content=None or multimodal list content, this public path raises TypeError, so the fix remains incomplete for LangChain callers.

Knowledge Base Used:

• Prompts and framework integrations
• Prompt retrieval, compilation, and caching

Prompt To Fix With AI

This is a comment left during a code review.
Path: langfuse/model.py
Line: 364-369

Comment:
**LangChain conversion still crashes**

Non-string placeholder content now survives `compile()`, but `get_langchain_prompt()` then passes every compiled message’s content to the string-only `_get_langchain_prompt_string()`. When the history contains an assistant tool-call turn with `content=None` or multimodal list content, this public path raises `TypeError`, so the fix remains incomplete for LangChain callers.

**Knowledge Base Used:**
- [Prompts and framework integrations](https://app.greptile.com/personal-org-4986/-/custom-context/knowledge-base/langfuse/langfuse-python/-/docs/prompts-and-framework-integrations.md)
- [Prompt retrieval, compilation, and caching](https://app.greptile.com/personal-org-4986/-/custom-context/knowledge-base/langfuse/langfuse-python/-/docs/prompt-retrieval-compilation-and-cache.md)

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[winklemad]

Good catch — get_langchain_prompt() hits the same string-only path (_escape_json_for_langchain does len(text)), so guarding compile() alone wasn't enough. Pushed a fix that applies the same isinstance-str guard there too: multimodal list content passes through (LangChain accepts it), and a tool-call turn's None is normalized to "" — LangChain rejects a (role, None) template tuple, and "" is its convention for a tool-call message. Added a get_langchain_prompt regression test that also asserts the result is consumable by ChatPromptTemplate.from_messages(...).