# A drop-in replacement chat template for google/gemma-4-31B-it tuned for open-source agentic coding harnesses. > Source: > Published: 2026-05-23 03:01:26+00:00 | {#--------------------------------------------------------------------- | | | custom_pub_chat_template_gemma4.jinja | | | ===================================== | | | A public, harness-friendly fork of Google's Gemma 4 chat template, | | | tuned for open-source agentic coding harnesses like: | | | - anomalyco/opencode (https://github.com/anomalyco/opencode) | | | - earendil-works/pi (https://github.com/earendil-works/pi) | | | - openclaw, OpenHarness, similar Claude-Code-style harnesses | | | WHY THIS FORK EXISTS | | | -------------------- | | | The upstream chat template at google/gemma-4-31B-it is correct for | | | chat use, but four real edge cases bite agentic coding harnesses: | | | 1. tool_call.arguments arriving as a JSON string (Vercel AI SDK and | | | several OpenAI-compatible adapters serialize this way) is silently | | | wrapped in extra braces by the upstream template, producing invalid | | | Gemma 4 DSL like call:fn{{"city":"Tokyo"}} — nested braces, JSON | | | colons, and quoted keys, none of which the model was trained on. | | | Symptom: degraded tool-call accuracy, mysterious arguments collapse | | | to {} on repeated calls. | | | 2. Prior-turn reasoning is dropped from history. The model card says | | | "historical model output should only include the final response," | | | but agentic harnesses doing multi-step tool calls benefit | | | materially from keeping the prior reasoning visible. The Qwen3.6 | | | analogue of this bug is documented at: | | | https://github.com/earendil-works/pi/issues/3325 | | | Symptom (on Qwen and Gemma alike): after 2-3 turns, every tool call | | | collapses to arguments: {} even though the model's prior reasoning | | | correctly identified the parameters it needed. | | | 3. enable_thinking defaults to FALSE in the upstream template, and | | | most OpenAI-compatible adapters drop unknown request fields: | | | https://github.com/anomalyco/opencode/issues/24264 | | | So the harness ends up with thinking permanently off, agentic | | | tool-call accuracy suffers, and there's no obvious failure signal. | | | 4. JSON null values inside tool_call.arguments render as the bare | | | string "None" (Python repr of None survives Jinja). Optional | | | fields are very common in coding tools (find_files, search: | | | pattern=..., language=null) and this corrupts the prompt silently. | | | This fork is also forked from a private engineering fork used in | | | internal harnesses; the public copy reuses the same five patches but | | | adds expanded comments, removes references to private design docs, | | | and ships with a self-contained pytest conformance suite. | | | PATCH INVENTORY (full details next to each patch site below) | | | ------------------------------------------------------------ | | | P1 format_argument: emit JSON null instead of bare "None" | | | P2 enable_thinking defaults to TRUE | | | P3 tool_call.arguments as string: RAISE instead of silent corruption | | | P4 preserve_thinking kwarg (default TRUE) keeps prior <|channel> | | | P5 fix HF discussion #62 turn-tag close asymmetry | | | INVARIANT | | | --------- | | | With enable_thinking=False AND preserve_thinking=False passed | | | explicitly, this template renders byte-for-byte identical to the | | | upstream verbatim template on every input that doesn't hit P1, P3, | | | or P5's bug sites. The conformance suite at | | | tests/test_custom_pub_chat_template.py | | | locks this in across 21 representative cases. | | | USAGE | | | ----- | | | Server side (e.g. vLLM or SGLang): | | | --chat-template /path/to/custom_pub_chat_template_gemma4.jinja | | | Harness side: no changes required for the common case. If you need | | | to force defaults off (e.g. to match upstream behaviour exactly): | | | { | | | "extra_body": { | | | "chat_template_kwargs": { | | | "enable_thinking": false, | | | "preserve_thinking": false | | | } | | | } | | | } | | | For opencode-style providers, this maps to the `chat_template_args` | | | field in models config; for pi, set thinkingFormat appropriately | | | in the provider's compat block and pi will inject these kwargs. | | | PINS | | | ---- | | | Forked from google/gemma-4-31B-it @ fcf2302760ae9c6e528a8dbba9dd636e56848237 | | | Fork date: 2026-05-22 | | | License: Apache 2.0 (same as upstream) | | | Maintainer: see repo README | | | ---------------------------------------------------------------------#} | | | {%- macro format_parameters(properties, required, filter_keys=false) -%} | | | {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%} | | | {%- set ns = namespace(found_first=false) -%} | | | {%- for key, value in properties | dictsort -%} | | | {%- set add_comma = false -%} | | | {%- if not filter_keys or key not in standard_keys -%} | | | {%- if ns.found_first %},{% endif -%} | | | {%- set ns.found_first = true -%} | | | {{ key }}:{ | | | {%- if value['description'] -%} | | | description:<|"|>{{ value['description'] }}<|"|> | | | {%- set add_comma = true -%} | | | {%- endif -%} | | | {%- if value['type'] | upper == 'STRING' -%} | | | {%- if value['enum'] -%} | | | {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} | | | enum:{{ format_argument(value['enum']) }} | | | {%- endif -%} | | | {%- elif value['type'] | upper == 'ARRAY' -%} | | | {%- if value['items'] is mapping and value['items'] -%} | | | {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} | | | items:{ | | | {%- set ns_items = namespace(found_first=false) -%} | | | {%- for item_key, item_value in value['items'] | dictsort -%} | | | {%- if item_value is not none -%} | | | {%- if ns_items.found_first %},{% endif -%} | | | {%- set ns_items.found_first = true -%} | | | {%- if item_key == 'properties' -%} | | | properties:{ | | | {%- if item_value is mapping -%} | | | {{- format_parameters(item_value, value['items']['required'] | default([])) -}} | | | {%- endif -%} | | | } | | | {%- elif item_key == 'required' -%} | | | required:[ | | | {%- for req_item in item_value -%} | | | <|"|>{{- req_item -}}<|"|> | | | {%- if not loop.last %},{% endif -%} | | | {%- endfor -%} | | | ] | | | {%- elif item_key == 'type' -%} | | | {%- if item_value is string -%} | | | type:{{ format_argument(item_value | upper) }} | | | {%- else -%} | | | type:{{ format_argument(item_value | map('upper') | list) }} | | | {%- endif -%} | | | {%- else -%} | | | {{ item_key }}:{{ format_argument(item_value) }} | | | {%- endif -%} | | | {%- endif -%} | | | {%- endfor -%} | | | } | | | {%- endif -%} | | | {%- endif -%} | | | {%- if value['nullable'] %} | | | {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} | | | nullable:true | | | {%- endif -%} | | | {%- if value['type'] | upper == 'OBJECT' -%} | | | {%- if value['properties'] is defined and value['properties'] is mapping -%} | | | {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} | | | properties:{ | | | {{- format_parameters(value['properties'], value['required'] | default([])) -}} | | | } | | | {%- elif value is mapping -%} | | | {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} | | | properties:{ | | | {{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}} | | | } | | | {%- endif -%} | | | {%- if value['required'] -%} | | | {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} | | | required:[ | | | {%- for item in value['required'] | default([]) -%} | | | <|"|>{{- item -}}<|"|> | | | {%- if not loop.last %},{% endif -%} | | | {%- endfor -%} | | | ] | | | {%- endif -%} | | | {%- endif -%} | | | {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%} | | | type:<|"|>{{ value['type'] | upper }}<|"|>} | | | {%- endif -%} | | | {%- endfor -%} | | | {%- endmacro -%} | | | {%- macro format_function_declaration(tool_data) -%} | | | declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|> | | | {%- set params = tool_data['function']['parameters'] -%} | | | {%- if params -%} | | | ,parameters:{ | | | {%- if params['properties'] -%} | | | properties:{ {{- format_parameters(params['properties'], params['required']) -}} }, | | | {%- endif -%} | | | {%- if params['required'] -%} | | | required:[ | | | {%- for item in params['required'] -%} | | | <|"|>{{- item -}}<|"|> | | | {{- ',' if not loop.last -}} | | | {%- endfor -%} | | | ], | | | {%- endif -%} | | | {%- if params['type'] -%} | | | type:<|"|>{{- params['type'] | upper -}}<|"|>} | | | {%- endif -%} | | | {%- endif -%} | | | {%- if 'response' in tool_data['function'] -%} | | | {%- set response_declaration = tool_data['function']['response'] -%} | | | ,response:{ | | | {%- if response_declaration['description'] -%} | | | description:<|"|>{{- response_declaration['description'] -}}<|"|>, | | | {%- endif -%} | | | {%- if response_declaration['type'] | upper == 'OBJECT' -%} | | | type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>} | | | {%- endif -%} | | | {%- endif -%} | | | } | | | {%- endmacro -%} | | | {%- macro format_argument(argument, escape_keys=True) -%} | | | {#- P1 (public fork): emit JSON null for None values rather than the | | | bare string "None". Jinja's default coercion of Python's None | | | goes through str(None) -> "None", which then leaks into the | | | Gemma 4 DSL as a literal token the model has never been trained | | | on. Common bite path: a coding tool's optional argument | | | (language=null in a find-files call, after=null in a search, | | | etc.) → upstream emits after:None in the DSL → model | | | confusion. We emit after:null instead, matching the JSON wire | | | format the model has actually seen. | | | Branch ordering: `is none` must precede `is string`, `is | | | mapping`, `is sequence`, etc., because None matches NONE of | | | them in Jinja's type tests but the final else-branch | | | ({{ argument }}) would otherwise stringify it. -#} | | | {%- if argument is none -%} | | | {{- 'null' -}} | | | {%- elif argument is string -%} | | | {{- '<|"|>' + argument + '<|"|>' -}} | | | {%- elif argument is boolean -%} | | | {{- 'true' if argument else 'false' -}} | | | {%- elif argument is mapping -%} | | | {{- '{' -}} | | | {%- set ns = namespace(found_first=false) -%} | | | {%- for key, value in argument | dictsort -%} | | | {%- if ns.found_first %},{% endif -%} | | | {%- set ns.found_first = true -%} | | | {%- if escape_keys -%} | | | {{- '<|"|>' + key + '<|"|>' -}} | | | {%- else -%} | | | {{- key -}} | | | {%- endif -%} | | | :{{- format_argument(value, escape_keys=escape_keys) -}} | | | {%- endfor -%} | | | {{- '}' -}} | | | {%- elif argument is sequence -%} | | | {{- '[' -}} | | | {%- for item in argument -%} | | | {{- format_argument(item, escape_keys=escape_keys) -}} | | | {%- if not loop.last %},{% endif -%} | | | {%- endfor -%} | | | {{- ']' -}} | | | {%- else -%} | | | {{- argument -}} | | | {%- endif -%} | | | {%- endmacro -%} | | | {%- macro strip_thinking(text) -%} | | | {%- set ns = namespace(result='') -%} | | | {%- for part in text.split('') -%} | | | {%- if '<|channel>' in part -%} | | | {%- set ns.result = ns.result + part.split('<|channel>')[0] -%} | | | {%- else -%} | | | {%- set ns.result = ns.result + part -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {{- ns.result | trim -}} | | | {%- endmacro -%} | | | {%- macro format_tool_response_block(tool_name, response) -%} | | | {{- '<|tool_response>' -}} | | | {%- if response is mapping -%} | | | {{- 'response:' + tool_name + '{' -}} | | | {%- for key, value in response | dictsort -%} | | | {{- key -}}:{{- format_argument(value, escape_keys=False) -}} | | | {%- if not loop.last %},{% endif -%} | | | {%- endfor -%} | | | {{- '}' -}} | | | {%- else -%} | | | {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}} | | | {%- endif -%} | | | {{- '' -}} | | | {%- endmacro -%} | | | {%- set ns = namespace(prev_message_type=None) -%} | | | {%- set loop_messages = messages -%} | | | {#- P2 (public fork): default enable_thinking to TRUE. | | | Why: Gemma 4's upstream template defaults enable_thinking to False | | | (or undefined). This is wrong for agentic coding harnesses for two | | | reasons: | | | 1. Google's own model card: thinking "significantly enhances | | | function-calling accuracy" — and tool calling IS the core | | | contract that coding harnesses use the model for. Defaulting it | | | off means most opencode/pi users see degraded tool accuracy and | | | have no obvious way to fix it. | | | 2. Most OpenAI-compatible SDKs (notably Vercel AI SDK used by | | | opencode) strip unknown request fields, so a harness that tries | | | to pass chat_template_kwargs.enable_thinking=true per request | | | has it silently dropped. See: | | | https://github.com/anomalyco/opencode/issues/24264 | | | Flipping the SERVER-SIDE default to True makes "the agentic | | | happy-path" the default and lets harnesses that explicitly want | | | chat-only behaviour override it to false per request: | | | {"extra_body":{"chat_template_kwargs":{"enable_thinking":false}}} | | | After this `set`, enable_thinking is unconditionally defined as a | | | bool, so downstream `is defined` guards are dropped. -#} | | | {%- set enable_thinking = enable_thinking | default(true) -%} | | | {{- bos_token -}} | | | {#- Handle System/Tool Definitions Block -#} | | | {%- if enable_thinking or tools or messages[0]['role'] in ['system', 'developer'] -%} | | | {{- '<|turn>system\n' -}} | | | {#- Inject Thinking token at the very top of the FIRST system turn -#} | | | {%- if enable_thinking -%} | | | {{- '<|think|>\n' -}} | | | {%- set ns.prev_message_type = 'think' -%} | | | {%- endif -%} | | | {%- if messages[0]['role'] in ['system', 'developer'] -%} | | | {%- if messages[0]['content'] is string -%} | | | {{- messages[0]['content'] | trim -}} | | | {%- elif messages[0]['content'] is sequence -%} | | | {%- for item in messages[0]['content'] -%} | | | {{- item['text'] | trim + ' '-}} | | | {%- endfor -%} | | | {%- endif -%} | | | {%- set loop_messages = messages[1:] -%} | | | {%- endif -%} | | | {%- if tools -%} | | | {%- for tool in tools %} | | | {{- '<|tool>' -}} | | | {{- format_function_declaration(tool) | trim -}} | | | {{- '' -}} | | | {%- endfor %} | | | {%- set ns.prev_message_type = 'tool' -%} | | | {%- endif -%} | | | {{- '\n' -}} | | | {%- endif %} | | | {#- P4 (public fork): preserve_thinking kwarg, default TRUE. | | | Why: upstream's reasoning re-emission gate fires only when an | | | assistant message (a) carries `reasoning`/` reasoning_content`, | | | (b) has tool_calls, AND (c) is AFTER the last user message. That | | | third clause is what causes the canonical multi-turn-tool-loop | | | breakage: | | | User: "find files matching '*.py' in src" | | | Assistant: (reasoning=...calling find_files...) tool_call: | | | find_files(pattern='*.py', dir='src') | | | Tool: [result list] | | | User: "now look for '*.ts' too" | | | Assistant: (reasoning=...) tool_call: find_files(pattern={}, dir={}) | | | ↑↑↑ arguments collapse to empty here because the prior | | | reasoning the model would have learned to imitate is | | | invisible — the previous-turn <|channel> was dropped. | | | The same shape was reported on Qwen3.6 and resolved by the | | | preserve_thinking kwarg there: | | | https://github.com/earendil-works/pi/issues/3325 | | | Gemma 4's model card says "historical model output should only | | | include the final response" — that guidance is correct for plain | | | chat but actively harmful for multi-turn agentic tool calling. P4 | | | optionally drops the (c) gate so prior reasoning stays visible to | | | the model on subsequent turns. | | | Set preserve_thinking=false to recover upstream behaviour exactly | | | (used by the conformance suite to verify byte-identity). -#} | | | {%- set preserve_thinking = preserve_thinking | default(true) -%} | | | {#- Pre-scan: find last user message index for reasoning guard -#} | | | {%- set ns_turn = namespace(last_user_idx=-1) -%} | | | {%- for i in range(loop_messages | length) -%} | | | {%- if loop_messages[i]['role'] == 'user' -%} | | | {%- set ns_turn.last_user_idx = i -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {#- Loop through messages -#} | | | {%- for message in loop_messages -%} | | | {%- if message['role'] != 'tool' -%} | | | {%- set ns.prev_message_type = None -%} | | | {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%} | | | {#- Detect continuation: suppress duplicate <|turn>model when previous non-tool message was also assistant -#} | | | {%- set prev_nt = namespace(role=None, found=false) -%} | | | {%- if loop.index0 > 0 -%} | | | {%- for j in range(loop.index0 - 1, -1, -1) -%} | | | {%- if not prev_nt.found -%} | | | {%- if loop_messages[j]['role'] != 'tool' -%} | | | {%- set prev_nt.role = loop_messages[j]['role'] -%} | | | {%- set prev_nt.found = true -%} | | | {%- endif -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {%- endif -%} | | | {%- set continue_same_model_turn = (role == 'model' and prev_nt.role == 'assistant') -%} | | | {%- if not continue_same_model_turn -%} | | | {{- '<|turn>' + role + '\n' }} | | | {%- endif -%} | | | {#- Render reasoning/reasoning_content as thinking channel. | | | Upstream gate (all three required to re-emit): | | | (a) the message carries reasoning or reasoning_content, | | | (b) the message has tool_calls, | | | (c) the message is after the last user message in history. | | | P4 (public fork): when preserve_thinking is true (default), drop | | | clause (c) so prior assistant turns' <|channel> blocks survive. | | | See the long P4 comment above the pre-scan for why this matters | | | for agentic tool loops. The (b) gate stays — re-emitting a | | | <|channel> on a finalised text-only assistant turn is not in | | | the model's training distribution. -#} | | | {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%} | | | {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or preserve_thinking -%} | | | {%- if thinking_text and thinking_gate and message.get('tool_calls') -%} | | | {{- '<|channel>thought\n' + thinking_text + '\n' -}} | | | {%- endif -%} | | | {%- if message['tool_calls'] -%} | | | {%- for tool_call in message['tool_calls'] -%} | | | {%- set function = tool_call['function'] -%} | | | {{- '<|tool_call>call:' + function['name'] + '{' -}} | | | {%- if function['arguments'] is mapping -%} | | | {%- set ns_args = namespace(found_first=false) -%} | | | {%- for key, value in function['arguments'] | dictsort -%} | | | {%- if ns_args.found_first %},{% endif -%} | | | {%- set ns_args.found_first = true -%} | | | {{- key -}}:{{- format_argument(value, escape_keys=False) -}} | | | {%- endfor -%} | | | {%- elif function['arguments'] is none -%} | | | {#- P3 (public fork): None / missing arguments is | | | valid (means: call this tool with no args). | | | Emit an empty {} via the empty for-loop above. -#} | | | {%- else -%} | | | {#- P3 (public fork): refuse string (or any other | | | non-mapping) arguments rather than silently | | | corrupting the prompt. | | | Bug surface: many OpenAI-compatible SDKs (most | | | notably Vercel AI SDK, used by opencode) hand | | | tool_call.arguments back as a JSON-encoded | | | STRING — e.g. '{"city":"Tokyo"}' — rather | | | than the already-deserialized object. The | | | upstream Gemma 4 template silently emits this | | | string verbatim inside an extra pair of | | | braces, producing invalid Gemma 4 DSL: | | | call:fn{{"city":"Tokyo"}} | | | (nested braces, JSON colons, quoted keys — | | | none of which the model has been trained on). | | | The model usually still produces a plausible | | | response, which makes the bug INSIDIOUS: it | | | looks like a quality problem with the model, | | | not a prompt-corruption bug in the harness. | | | Fix: harnesses MUST deserialize | | | tool_calls[].function.arguments | | | exactly once on ingest and store the object. | | | See the canonical pi-side discussion: | | | https://github.com/earendil-works/pi/issues/3325 | | | We raise here so the bug surfaces at the | | | server (an obvious HTTP error to debug) | | | rather than as a quiet model-output | | | regression. -#} | | | {{- raise_exception( | | | "custom_pub_chat_template_gemma4: " | | | "tool_calls[].function.arguments must be a JSON " | | | "object (mapping). Got a " | | | ~ (function['arguments'] | string | length | string) | | | ~ "-char " | | | ~ (function['arguments'].__class__.__name__ if function['arguments'].__class__ is defined else 'non-mapping') | | | ~ ". This is almost always the harness handing back " | | | "a JSON-encoded STRING rather than the deserialized " | | | "object. Deserialize once on ingest and store the " | | | "object. See: github.com/earendil-works/pi/issues/3325" | | | ) -}} | | | {%- endif -%} | | | {{- '}' -}} | | | {%- endfor -%} | | | {%- set ns.prev_message_type = 'tool_call' -%} | | | {%- endif -%} | | | {%- set ns_tr_out = namespace(flag=false) -%} | | | {%- if message.get('tool_responses') -%} | | | {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#} | | | {%- for tool_response in message['tool_responses'] -%} | | | {{- format_tool_response_block(tool_response['name'] | default('unknown'), tool_response['response']) -}} | | | {%- set ns_tr_out.flag = true -%} | | | {%- set ns.prev_message_type = 'tool_response' -%} | | | {%- endfor -%} | | | {%- elif message.get('tool_calls') -%} | | | {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#} | | | {%- set ns_tool_scan = namespace(stopped=false) -%} | | | {%- for k in range(loop.index0 + 1, loop_messages | length) -%} | | | {%- if ns_tool_scan.stopped -%} | | | {%- elif loop_messages[k]['role'] != 'tool' -%} | | | {%- set ns_tool_scan.stopped = true -%} | | | {%- else -%} | | | {%- set follow = loop_messages[k] -%} | | | {#- Resolve tool_call_id to function name -#} | | | {%- set ns_tname = namespace(name=follow.get('name') | default('unknown')) -%} | | | {%- for tc in message['tool_calls'] -%} | | | {%- if tc.get('id') == follow.get('tool_call_id') -%} | | | {%- set ns_tname.name = tc['function']['name'] -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {#- Handle content as string or content-parts array -#} | | | {%- set tool_body = follow.get('content') -%} | | | {%- if tool_body is string -%} | | | {{- format_tool_response_block(ns_tname.name, tool_body) -}} | | | {%- elif tool_body is sequence and tool_body is not string -%} | | | {%- set ns_txt = namespace(s='') -%} | | | {%- for part in tool_body -%} | | | {%- if part.get('type') == 'text' -%} | | | {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}} | | | {%- for part in tool_body -%} | | | {%- if part.get('type') == 'image' -%} | | | {{- '<|image|>' -}} | | | {%- elif part.get('type') == 'audio' -%} | | | {{- '<|audio|>' -}} | | | {%- elif part.get('type') == 'video' -%} | | | {{- '<|video|>' -}} | | | {%- endif -%} | | | {%- endfor -%} | | | {%- else -%} | | | {{- format_tool_response_block(ns_tname.name, tool_body) -}} | | | {%- endif -%} | | | {%- set ns_tr_out.flag = true -%} | | | {%- set ns.prev_message_type = 'tool_response' -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {%- endif -%} | | | {%- set captured_content -%} | | | {%- if message['content'] is string -%} | | | {%- if role == 'model' -%} | | | {{- strip_thinking(message['content']) -}} | | | {%- else -%} | | | {{- message['content'] | trim -}} | | | {%- endif -%} | | | {%- elif message['content'] is sequence -%} | | | {%- for item in message['content'] -%} | | | {%- if item['type'] == 'text' -%} | | | {%- if role == 'model' -%} | | | {{- strip_thinking(item['text']) -}} | | | {%- else -%} | | | {{- item['text'] | trim -}} | | | {%- endif -%} | | | {%- elif item['type'] == 'image' -%} | | | {{- '<|image|>' -}} | | | {%- set ns.prev_message_type = 'image' -%} | | | {%- elif item['type'] == 'audio' -%} | | | {{- '<|audio|>' -}} | | | {%- set ns.prev_message_type = 'audio' -%} | | | {%- elif item['type'] == 'video' -%} | | | {{- '<|video|>' -}} | | | {%- set ns.prev_message_type = 'video' -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {%- endif -%} | | | {%- endset -%} | | | {{- captured_content -}} | | | {%- set has_content = captured_content | trim | length > 0 -%} | | | {#- P5 (public fork): symmetric continuation close-suppression | | | for HF discussion #62. | | | The bug: upstream's open suppression at the top of this | | | iteration drops the `<|turn>model\n` header when the | | | previous non-tool message was also assistant — but the | | | close below ALWAYS emits `\n`. Two back-to-back | | | text-only assistant messages therefore render as: | | | <|turn>model\npart 1\npart 2\n | | | That's one open, two closes — malformed. The model | | | (Google-confirmed in HF discussion #62) sees it as a | | | truncated and re-opened turn, which destabilises long | | | multi-step agentic histories that accumulate consecutive | | | assistant messages. | | | Fix: forward-scan for the next non-tool message. If it is | | | another assistant AND this iteration is a TEXT-ONLY | | | assistant message (no tool_calls, no tool_responses), the | | | next iteration will continue this same turn frame, so | | | suppress this iteration's close and emit a single `\n` so | | | the two contents don't byte-glue together. | | | The narrowing condition (`not message.get('tool_calls') | | | and not ns_tr_out.flag`) is critical: the tool-call + | | | tool-response chain MUST close normally so the model still | | | sees a balanced turn frame around the `<|tool_response>` | | | block. Conformance test T13 locks this in. -#} | | | {%- set next_nt = namespace(role=None, found=false) -%} | | | {%- for j in range(loop.index0 + 1, loop_messages | length) -%} | | | {%- if not next_nt.found -%} | | | {%- if loop_messages[j]['role'] != 'tool' -%} | | | {%- set next_nt.role = loop_messages[j]['role'] -%} | | | {%- set next_nt.found = true -%} | | | {%- endif -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {%- set continues_into_next = ( | | | role == 'model' | | | and next_nt.role == 'assistant' | | | and not message.get('tool_calls') | | | and not ns_tr_out.flag | | | ) -%} | | | {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%} | | | {{- '<|tool_response>' -}} | | | {%- elif continues_into_next -%} | | | {{- '\n' -}} | | | {%- elif not (ns_tr_out.flag and not has_content) -%} | | | {{- '\n' -}} | | | {%- endif -%} | | | {%- endif -%} | | | {%- endfor -%} | | | {%- if add_generation_prompt -%} | | | {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%} | | | {{- '<|turn>model\n' -}} | | | {#- When thinking is disabled, the upstream contract is to | | | pre-fill an empty `<|channel>thought\n` block so | | | the model skips reasoning. After P2's set at the top of | | | the file, `enable_thinking` is unconditionally a bool, so | | | the upstream `| default(false)` is unnecessary. (It also | | | had a Jinja precedence trap: `|` binds tighter than `not`, | | | parsing as `not (enable_thinking | default(false))`. The | | | simple `not enable_thinking` form is equivalent and | | | clearer.) -#} | | | {%- if not enable_thinking -%} | | | {{- '<|channel>thought\n' -}} | | | {%- endif -%} | | | {%- endif -%} | | | {%- endif -%} |