chat_template.jinja

{#- ===== HELPER: raise_exception macro =====
     Jinja2 doesn't have a built-in raise_exception.
     This macro outputs an error marker in the rendered output.
     Callers should check output for "ERROR:" pattern to detect validation failures.
-#}
{%- macro raise_exception(message) -%}
  {{- '\n[ERROR: ' ~ message ~ ']' -}}
{%- endmacro -%}

{#- ===== SECTION 1A: MACRO render_content =====
     Handles string, list (image/video/text items), or None/undefined.
     count_vision=true: increments ns.image_count / ns.video_count.
     is_system_content=false: Set true when rendering system/developer content
                            to enable media validation (raises exception).
     count_vision=true: increments vision counters.
-#}
{%- macro render_content(content, count_vision=false, is_system_content=false) -%}
  {#- VALIDATION: System messages cannot contain images or videos (from v18) -#}
  {#- FIX: also exclude strings and handle None - llama.cpp treats strings as non-iterable in for loops -#}
  {%- if is_system_content and content is iterable and content is not mapping and content is not string and content is not none -%}
    {%- for item in content -%}
      {%- if item.type == 'image' or 'image' in item or 'image_url' in item -%}
        {{- raise_exception('System message cannot contain images.') -}}
      {%- endif -%}
      {%- if item.type == 'video' or 'video' in item -%}
        {{- raise_exception('System message cannot contain videos.') -}}
      {%- endif -%}
    {%- endfor -%}
  {%- endif -%}

  {#- Main content rendering -#}
  {#- Handle None/undefined content -#}
  {%- if content is none or content is defined == false -%}
    {{- '' -}}
  {%- elif content is string -%}
    {{- content -}}
  {#- FIX: also exclude strings - llama.cpp treats strings as non-iterable in for loops -#}
  {%- elif content is iterable and content is not mapping and content is not string -%}
    {%- for item in content -%}
      {#- Handle different item types -#}
      {%- if item.type == 'image' or 'image' in item or 'image_url' in item -%}
        {%- if count_vision -%}{%- set ns.image_count = ns.image_count + 1 -%}{%- endif -%}
        {%- if add_vision_id is defined and add_vision_id -%}
          {{- 'Picture ' ~ ns.image_count ~ ': ' -}}
        {%- endif -%}
        {{- '<|vision_start|><|image_pad|><|vision_end|>' -}}
      {%- elif item.type == 'video' or 'video' in item -%}
        {%- if count_vision -%}{%- set ns.video_count = ns.video_count + 1 -%}{%- endif -%}
        {%- if add_vision_id is defined and add_vision_id -%}
          {{- 'Video ' ~ ns.video_count ~ ': ' -}}
        {%- endif -%}
        {{- '<|vision_start|><|video_pad|><|vision_end|>' -}}
      {%- elif item.type == 'text' or 'text' in item -%}
        {{- item.text -}}
      {#- ERROR: Unknown content type - raise explicit exception (from v18) -#}
      {%- else -%}
        {{- raise_exception('Unexpected content type in message content.') -}}
      {%- endif -%}
    {%- endfor -%}
  {#- ERROR: Unknown content type - raise explicit exception (from v18) -#}
  {%- elif content is not none and content is defined -%}
    {{- raise_exception('Unexpected content type.') -}}
  {%- endif -%}
{%- endmacro -%}

{#- ===== SECTION 1B: MACRO detect_tool_error (NEW in v0.7) =====
     Detects if a tool response contains error indicators.
     Uses heuristics from v18:
     - Checks for error keywords (error, exception, traceback, failed to)
     - Ignores responses with '$ ' (shell output prefix) or 'took ' (timing info)
     - Ignores responses > 500 chars (likely valid output, not error)

     Returns: ns.last_tool_failed (true/false)
     Side effect: Updates ns.consecutive_failures counter
-#}
{%- macro detect_tool_error(content) -%}
  {#- Type guard: ensure content is string (llama.cpp compatibility) -#}
  {%- set content = content if content is string else '' -%}
  {%- set content_lower = content | lower -%}
  {%- set content_length = content | length -%}

  {#- Error detection heuristics: short response + no shell prefix + has error keywords -#}
  {%- if content_length < 500
      and '$ ' not in content
      and 'took ' not in content_lower
      and ('"error":' in content_lower or 'error:' in content_lower
           or 'exception:' in content_lower or 'traceback' in content_lower
           or 'command not found' in content_lower or 'invalid syntax' in content_lower
           or 'failed to' in content_lower or 'permission denied' in content_lower) -%}
    {#- Error detected - update failure tracking -#}
    {%- set ns.last_tool_failed = true -%}
    {%- set ns.consecutive_failures = ns.consecutive_failures + 1 -%}
  {%- else -%}
    {#- No error - reset failure tracking -#}
    {%- set ns.last_tool_failed = false -%}
    {%- set ns.consecutive_failures = 0 -%}
  {%- endif -%}
{%- endmacro -%}

{#- ===== SECTION 2: NAMESPACE INITIALISATION =====
     Single ns object for all mutable state.

     enable_thinking:  default=true (controls think-block in generation prompt)
     preserve_thinking: default=true (controls think-block display in conversation history)
     image_count:      Vision counter for images
     video_count:      Vision counter for videos

     NEW in v0.7:
     - consecutive_failures: Tracks consecutive tool call failures (from v18)
     - last_tool_failed: Boolean flag for current tool response (from v18)
-#}
{%- set ns = namespace(
    enable_thinking=false,
    preserve_thinking=false,
    image_count=0,
    video_count=0,
    consecutive_failures=0,
    last_tool_failed=false
) -%}

{#- Resolve enable_thinking kwarg -#}
{%- if enable_thinking is defined -%}
  {%- if enable_thinking -%}
    {%- set ns.enable_thinking = true -%}
  {%- else -%}
    {%- set ns.enable_thinking = false -%}
  {%- endif -%}
{%- endif -%}

{#- Resolve preserve_thinking kwarg (FIXED in v0.7: now also affects conversation history, not just generation prompt).
    preserve_thinking=false  => force non-thinking mode (same as enable_thinking=false).
    preserve_thinking=true   => default, no override (thinking controlled by enable_thinking).
    When not defined         => default, no override.
-#}
{%- if preserve_thinking is defined -%}
  {%- if not preserve_thinking -%}
    {%- set ns.enable_thinking = false -%}
    {%- set ns.preserve_thinking = false -%}
  {%- else -%}
    {%- set ns.preserve_thinking = true -%}
  {%- endif -%}
{%- endif -%}

{#- ===== SECTION 3: PRE-SCAN =====
     Track last /no_think or /think flag in user messages.
     Also scan system messages for <|think_off|> / <|think_on|> markers
     (allows apps to control thinking mode via system prompt injection).
     The model follows the last flag encountered in multi-turn conversations.
-#}
{%- for i in range(messages | length) -%}
  {%- set _msg = messages[i] -%}
  {%- if _msg.role == 'user' -%}
    {%- set _u = _msg.content if _msg.content is string else '' -%}
    {%- if _u.rstrip().endswith('/no_think') -%}
      {%- set ns.enable_thinking = false -%}
    {%- elif _u.rstrip().endswith('/think') -%}
      {%- set ns.enable_thinking = true -%}
    {%- endif -%}
  {%- elif _msg.role == 'system' or _msg.role == 'developer' -%}
    {%- set _s = _msg.content if _msg.content is string else '' -%}
    {%- if '<|think_off|>' in _s -%}
      {%- set ns.enable_thinking = false -%}
    {%- elif '<|think_on|>' in _s -%}
      {%- set ns.enable_thinking = true -%}
    {%- endif -%}
  {%- endif -%}
{%- endfor -%}

{#- ===== SECTION 4: VALIDATE MESSAGES (NEW in v0.7) =====
     Validate that messages is provided and not empty.
     From v18: raises exception if no messages provided.
-#}
{%- if not messages -%}
  {{- raise_exception('No messages provided.') -}}
{%- endif -%}

{#- ===== SECTION 5: COLLECT SYSTEM CONTENT =====
     Merge all system/developer messages with \n\n separator.
     <|think_off|> / <|think_on|> markers are stripped from output.

     FIXED in v0.7: Pass is_system_content=true to render_content to trigger
     media validation (raises exception if system contains images/videos).
-#}
{%- set ns_sys = namespace(content='') -%}
{%- for msg in messages -%}
  {%- if msg.role == 'system' or msg.role == 'developer' -%}
    {#- Pass is_system_content=true for media validation -#}
    {%- set _c = render_content(msg.content | default(''), false, true) | trim -%}
    {%- set _c = _c | replace('<|think_off|>', '') | replace('<|think_on|>', '') | trim -%}
    {%- if _c -%}
      {%- if ns_sys.content == '' -%}
        {%- set ns_sys.content = _c -%}
      {%- else -%}
        {%- set ns_sys.content = ns_sys.content + '\n\n' + _c -%}
      {%- endif -%}
    {%- endif -%}
  {%- endif -%}
{%- endfor -%}

{#- ===== SECTION 6: BUILD TOOLS LIST =====
     Normalise each tool to {"type":"function","function":{...}} format.
     Serialisation happens later at output time (avoids Markup + str escaping bugs).
-#}
{%- set _has_tools = tools is defined and tools -%}
{%- if _has_tools -%}
  {%- set ns_tb = namespace(list=[]) -%}
  {%- for tool in tools -%}
    {%- if tool.function is defined -%}
      {%- set ns_tb.list = ns_tb.list + [tool] -%}
    {%- else -%}
      {%- set ns_tb.list = ns_tb.list + [{"type": "function", "function": tool}] -%}
    {%- endif -%}
  {%- endfor -%}
{%- endif -%}

{#- ===== SECTION 7: OUTPUT SYSTEM TURN =====
     Each fragment output via its own {{ }} block so tojson Markup objects are
     never Python-concatenated with plain strings (would trigger HTML-escaping).
     User system content appears BEFORE the tools block (correct ordering).
     No default system prompt injected.
-#}
{%- if ns_sys.content or _has_tools -%}
  {{- '<|im_start|>system\n' -}}
  {%- if ns_sys.content -%}
    {{- ns_sys.content -}}
    {%- if _has_tools -%}{{- '\n\n' -}}{%- endif -%}
  {%- endif -%}
  {%- if _has_tools -%}
    {{- '# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n' -}}
    {%- for tool in ns_tb.list -%}
      {{- tool | tojson -}}
      {%- if not loop.last -%}{{- '\n' -}}{%- endif -%}
    {%- endfor -%}
    {{- '\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{"name": <function-name>, "arguments": <args-json-object>}\n</tool_call>' -}}
  {%- endif -%}
  {{- '<|im_end|>\n' -}}
{%- endif -%}

{#- ===== SECTION 8: MAIN MESSAGE LOOP =====
     FIXED in v0.7:
     - Tool responses now have error detection via detect_tool_error macro
     - Warning messages injected for failed tool calls
     - consecutive_failures tracking for escalating warnings
-#}
{%- for message in messages -%}

  {#- 8a: System / Developer — already rendered above, skip -#}
  {%- if message.role == 'system' or message.role == 'developer' -%}

  {#- 8b: User messages -#}
  {%- elif message.role == 'user' -%}
    {%- set _uc = render_content(message.content | default(''), true, false) -%}
    {{- '<|im_start|>user\n' + _uc + '<|im_end|>\n' -}}

  {#- 8c: Assistant messages -#}
  {%- elif message.role == 'assistant' -%}
    {#- Safely extract content as string — guard against absent key.
        Also support message.reasoning_content as an explicit think-block source
        (used by some frameworks that store thinking separately from content). -#}
    {%- if message.content is defined and message.content is string -%}
      {%- set _ac = message.content -%}
    {#- FIX: also exclude strings - llama.cpp treats strings as non-iterable in for loops -#}
    {%- elif message.content is defined and message.content is iterable and message.content is not mapping and message.content is not string -%}
      {%- set _ac = render_content(message.content, false, false) -%}
    {%- else -%}
      {%- set _ac = '' -%}
    {%- endif -%}

    {#- Reconstruct content from reasoning_content + content when the framework
        stores thinking separately (e.g. OpenAI-style reasoning_content field).
        Only apply when no think-block already present in _ac. -#}
    {%- if message.reasoning_content is defined and message.reasoning_content is string
        and message.reasoning_content | trim
        and '<think>' not in _ac -%}
      {%- set _ac = '<think>\n' + message.reasoning_content | trim + '\n</think>\n\n' + _ac -%}
    {%- endif -%}

    {#- Collect tool_calls if present -#}
    {#- Type check: ensure tool_calls is a list, not string (llama.cpp compatibility) -#}
    {%- set _tc = message.tool_calls if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls is not string else [] -%}

    {#- Strip <tool_call> prefix from content when tool_calls also present
        (some frameworks duplicate the data in both fields) -#}
    {%- if _tc and '<tool_call>' in _ac -%}
      {%- set _ac = _ac.split('<tool_call>')[0] | trim -%}
    {%- endif -%}

    {#- FIXED in v0.7: Think-block handling with preserve_thinking support

       New logic (from v18): preserve_thinking controls think-block display on ALL
       assistant messages, not just generation prompt:

       - Tool-call turns   : never strip (think block is part of the tool-call format)
       - preserve_thinking : if true, show think blocks on ALL messages
       - Last-history turn : if preserve_thinking false, apply last-turn handling
       - Historical turns  : if preserve_thinking false, strip think blocks

       The old behavior (strip unless add_generation_prompt) is now controlled
       by preserve_thinking parameter.
    -#}
    {%- set _show_think = false -%}
    {%- if _tc -%}
      {#- Tool calls: always show think block -#}
      {%- set _show_think = true -%}
    {%- elif ns.preserve_thinking -%}
      {#- preserve_thinking=true: show think blocks on all messages -#}
      {%- set _show_think = true -%}
    {%- elif loop.last -%}
      {#- Last message without preserve_thinking: show if thinking enabled -#}
      {%- set _show_think = ns.enable_thinking -%}
    {%- endif -%}

    {#- Apply think-block stripping based on _show_think flag -#}
    {%- if not _show_think -%}
      {#- Fuzzy end-tag detection for stripping -#}
      {%- set _think_end = '' -%}
      {%- if '</think>' in _ac -%}
        {%- set _think_end = '</think>' -%}
      {%- elif '</thinking>' in _ac -%}
        {%- set _think_end = '</thinking>' -%}
      {%- elif '</ think>' in _ac -%}
        {%- set _think_end = '</ think>' -%}
      {%- elif '</think >' in _ac -%}
        {%- set _think_end = '</think >' -%}
      {%- endif -%}
      {%- if _think_end -%}
        {%- set _ac = _ac.split(_think_end)[-1].lstrip('\n') -%}
      {%- endif -%}
    {%- elif not _tc and loop.last and '<think>' not in _ac and not ns.enable_thinking -%}
      {#- Last turn, non-thinking: inject empty think block if missing -#}
      {%- set _ac = '<think>\n\n</think>\n\n' + _ac -%}
    {%- endif -%}

    {#- Emit the assistant turn -#}
    {{- '<|im_start|>assistant\n' -}}
    {%- if _ac -%}
      {{- _ac -}}
      {%- if _tc -%}{{- '\n' -}}{%- endif -%}
    {%- endif -%}

    {#- Render tool calls in Hermes format.
        Each value output via its own {{ }} block — never concatenated with plain strings
        in Python, which would trigger Markup HTML-escaping. -#}
    {%- if _tc -%}
      {%- for tc in _tc -%}
        {{- '<tool_call>\n' -}}
        {{- '{"name": ' -}}{{- tc.function.name | tojson -}}
        {%- if tc.function.arguments is string -%}
          {{- ', "arguments": ' + tc.function.arguments -}}
        {%- else -%}
          {{- ', "arguments": ' -}}{{- tc.function.arguments | tojson -}}
        {%- endif -%}
        {{- '}' -}}
        {%- if not loop.last -%}
          {{- '\n</tool_call>\n' -}}
        {%- else -%}
          {{- '\n</tool_call>' -}}
        {%- endif -%}
      {%- endfor -%}
    {%- endif -%}
    {{- '<|im_end|>\n' -}}

  {#- 8d: Tool results — with error detection (NEW in v0.7) -#}
  {%- elif message.role == 'tool' -%}
    {%- set _prev_role = messages[loop.index0 - 1].role if loop.index0 > 0 else '' -%}
    {%- set _next_role = messages[loop.index0 + 1].role if not loop.last else '' -%}

    {#- NEW in v0.7: Detect errors in tool response -#}
    {%- set _tool_content = message.content | default('') -%}
    {{- detect_tool_error(_tool_content) -}}

    {%- if _prev_role != 'tool' -%}
      {{- '<|im_start|>user\n' -}}
    {%- endif -%}
    {{- '<tool_response>\n' -}}
    {{- _tool_content -}}

    {#- NEW in v0.7: Inject warning if tool error detected -#}
    {#- v0.8: Replaced emoji with text-only for tokenization safety -#}
    {%- if ns.last_tool_failed -%}
      {%- if ns.consecutive_failures >= 2 -%}
        {{- '\n\n[SYSTEM WARNING: ' ~ ns.consecutive_failures ~ ' consecutive tool errors detected. Your previous approach is incorrect.]' -}}
      {%- else -%}
        {{- '\n\n[SYSTEM WARNING: The previous tool call returned an error. Diagnose the failure and retry with corrected arguments.]' -}}
      {%- endif -%}
    {%- endif -%}

    {%- if _next_role == 'tool' -%}
      {{- '\n</tool_response>\n' -}}
    {%- else -%}
      {{- '\n</tool_response>' -}}
      {{- '<|im_end|>\n' -}}
    {%- endif -%}

  {#- 8e: Unknown role - explicit error (from v18) -#}
  {%- else -%}
    {{- raise_exception('Unexpected message role: ' + message.role) -}}
  {%- endif -%}

{%- endfor -%}

{#- ===== SECTION 9: GENERATION PROMPT =====
     FIXED in v0.7: preserve_thinking now affects conversation history (Section 8),
     so generation prompt logic is simplified.

     enable_thinking=True  → open <think>\n prefill so llama.cpp reasoning-budget
                             and other inference engines can hook into the think-stream.
                             The model continues generating inside the open block.
     enable_thinking=False → exact non-thinking prefill: </think>\n\n

     NOTE: The <think>\n opener is EPHEMERAL — it lives only in the generation
     prompt, never in chat history. Historical think-block stripping is handled
     in Section 8 based on preserve_thinking setting.
-#}
{%- if add_generation_prompt -%}
  {{- '<|im_start|>assistant\n' -}}
  {%- if ns.enable_thinking -%}
    {{- '<think>\n' -}}
  {%- else -%}
    {{- '<think>\n\n</think>\n\n' -}}
  {%- endif -%}
{%- endif -%}