A drop-in replacement chat template for google/gemma-4-31B-it tuned for open-source agentic coding harnesses.

A developer has released a public fork of Google's Gemma 4 chat template that fixes four critical bugs affecting open-source agentic coding harnesses. The fork addresses issues including corrupted tool call arguments from nested JSON braces, dropped prior-turn reasoning, disabled thinking mode, and Python "None" strings appearing in place of JSON null values. The patched template ships with a conformance test suite and is designed for use with harnesses like OpenCode and Pi.

| { --------------------------------------------------------------------- | | | 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 '<channel| ' -%} | | | {%- 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 -%} | | | {{- '<tool response| ' -}} | | | {%- 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 -}} | | | {{- '<tool| ' -}} | | | {%- endfor %} | | | {%- set ns.prev message type = 'tool' -%} | | | {%- endif -%} | | | {{- '<turn| \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<channel| ' -}} | | | {%- 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 -%} | | | {{- '}<tool call| ' -}} | | | {%- 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 <turn| \n . Two back-to-back | | | text-only assistant messages therefore render as: | | | <|turn model\npart 1<turn| \npart 2<turn| \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 -%} | | | {{- '<turn| \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<channel| 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<channel| ' -}} | | | {%- endif -%} | | | {%- endif -%} | | | {%- endif -%} |