String/typed content caps differ from llama.cpp for SmolVLM, TranslateGemma 2B, Ministral 3 and LFM2.5 #720

TemplateCaps.detect reports string/typed content support that differs from llama.cpp's /props chat_template_caps for five templates. llamadart uses these caps to decide how to shape message content: whether to take the typed-content path when media is present. Because the caps are wrong, prompts differ from llama-server.

The other 92 templates in the #719 differential that llama-server loads agree.

How the repros were run

• Each repro renders with ChatTemplateEngine.render(templateSource: <template>, messages: …, metadata: {bos '<s>', eos '</s>'}) at ad8dd330d.
• The llama-server reference is 7fe450e19 POST /apply-template with the same conversation, run as --jinja --chat-template-file <template>:
  • text conversations on stories15M.gguf;
  • image conversations on LFM2-VL-450M-Q4_0.gguf with --mmproj mmproj-LFM2-VL-450M-Q8_0.gguf.
• In the server output, the media marker is normalized to <__media__>, and the model's <|im_end|> EOS is compared as </s>.

SmolVLM: text-only chats lose all message text

Conversation, no tools: user Weather in Paris?; assistant Checking. plus a get_weather call; tool result {"temp":21.5,"sky":"clear","note":"it's"}.

llamadart (same at base eeb382b55):

<|im_start|>User: <end_of_utterance>
Assistant: <end_of_utterance>
Tool: <end_of_utterance>
Assistant:

llama-server:

<|im_start|>User: Weather in Paris?<end_of_utterance>
Assistant: Checking.<end_of_utterance>
Tool: {"temp":21.5,"sky":"clear","note":"it's"}<end_of_utterance>
Assistant:

• The template iterates message['content'] as parts.
• llama.cpp converts every string to one text part for typed-only templates (messages_inp_normalizer, common/chat.cpp:474).
• llamadart passes strings, and uses the typed path only when a message has media (chat_template_engine.dart:254).
• So the fix needs both corrected caps and routing typed-only templates to parts even without media. I tried that locally: the same render becomes identical to llama-server.

Ministral 3: image plus a reasoning-only turn throws or drops the reasoning

Conversation: user image + Describe, assistant with reasoning Let me look. only, user More?.

• Ministral-3-14B throws Assistant message must have a string or a list of chunks in content or a list of tool calls. llama-server renders …[INST]<__media_…__>Describe[/INST][THINK]Let me look.[/THINK]<|im_end|>[INST]More?[/INST] (<|im_end|> is LFM2-VL's EOS).
• Ministral 3 3B renders …[INST]<__media__>Describe[/INST]</s>[INST]More?[/INST], without the [THINK] block.
• Cause: because llamadart thinks the template takes typed content, the image path goes through renderWithMultimodalContent, which skips MinistralHandler's block serialization.
• With llama.cpp's caps (string-only) forced locally, all 12 image renders of each Ministral template in the differential are identical to llama-server. Today it is 2 of 12, plus 4 throws for 14B.

LFM2.5-8B-A1B: correcting the caps alone makes it worse

• With an image, llamadart's typed path currently happens to match llama-server: 8 of 12 image renders are identical.
• With llama.cpp's caps (string-only) forced, only 4 of 12 match.
• The string path goes through Lfm2Handler, which flattens the tool list: List of tools: [{"name": "get_weather", … instead of [{"type": "function", "function": {"name": "get_weather", ….
• This is part 2 of #716. Fix it together with this issue.

TranslateGemma 2B: correcting the caps alone makes it worse

• With llama.cpp's caps (typed-only) forced, the empty assistant turn gains [{'type': 'text', 'text': ''}] in the prompt. llama-server prints nothing there.
• TranslateGemmaHandler's content normalization needs to change with the caps.

Related: Kimi-K2 tool results on the image path (separate cause, blocked by this issue)

• With an image in the conversation, moonshotai-Kimi-K2.jinja renders the tool result as ## Return of call_0\n[{'type': 'text', 'text': 'sunny'}]. llama-server renders ## Return of call_0\nsunny.
• Kimi-K2's caps match llama.cpp (both true). The cause is that llamadart's typed path turns every message's content, including tool results, into parts. llama.cpp keeps strings unless the template is typed-only.
• Making tool results strings on the typed path fixes Kimi-K2, but moved 8 SmolVLM image renders away from llama-server, because of SmolVLM's wrong caps. So it has to follow this fix.

All of this exists at base eeb382b55, before #719. It was found in #719's multimodal differential.

Expected

• TemplateCaps string/typed content caps match llama.cpp /props for these templates, pinned in template_caps_cache_fixtures_test.dart.
• Typed-only templates get text parts for every message, as llama.cpp's normalizer does.
• The Ministral, LFM2.5 and TranslateGemma handlers produce llama-server's prompts under the corrected caps. Pin each with a multimodal /apply-template fixture in test/fixtures/template_message_json_upstream.json.
• Then give tool results on the typed path as strings, unless the template is typed-only (Kimi-K2).