[Joel Stidley (joelst)]

Independent of #1129. The two PRs can merge in either order and merge cleanly together. This PR fixes the second Gemma 4 symptom from #1128: an extra <|turn>model after tool responses.

Minimal repro (Jinja2 vs. ORT Extensions before this fix)

Template	Jinja2	Before
{% set ns = namespace(v='outer') %}{% macro m() %}{% set ns = namespace(v='inner') %}{{ ns.v }}{% endmacro %}{{ m() }}|{{ ns.v }}	inner|outer	inner|inner
{% set x = 'outer' %}{% macro m(x) %}{{ x }}{% endmacro %}{{ m('arg') }}|{{ x }}	arg|outer	arg|arg
{% macro r(n, tag) %}{% if n > 0 %}{{ r(n - 1, 'inner') }}{% endif %}{{ tag }}{% endmacro %}{{ r(2, 'outer') }}	innerinnerouter	innerinnerinner

Change

One line in shared/api/minja.hpp: each macro call renders in a child context of the defining context instead of the defining context itself. This matches Jinja2 semantics and upstream google/minja (execution_context = Context::make(Value::object(), macro_context)).

Scope / behavior

• Fixed: macro parameters and {% set %} inside a macro no longer overwrite outer variables. Recursive macros no longer overwrite their callers' arguments.
• Unchanged: {% set ns.attr = ... %} inside a macro still mutates an outer namespace, because ns is resolved through parent lookup. Default parameter values are still evaluated per call. Templates that don't use macros are unaffected.
• Possible break: only templates that read a macro's parameters or locals after calling the macro. That relied on this bug and fails in Jinja2.

Validation

• New OrtxTokenizerTest.MinjaMacroCallsUseLocalScope: the 3 scope assertions fail before the fix (inner|inner, arg|arg, innerinnerinner) and pass after. The namespace-mutation and default-argument assertions pass both before and after (guarding against over-isolation).
• New OrtxTokenizerTest.Gemma4ToolResponseSuppressesGenerationPrompt, which uses the Gemma 4 template already on main, with the expected string taken from Jinja2. It fails before the fix (trailing <|turn>model) and passes after.
• All OrtxTokenizerTest.* tests pass (89/89) on macOS arm64.
• Downstream: ONNX Runtime GenAI rel-0.17.1 built against this change (via FETCHCONTENT_SOURCE_DIR_ONNXRUNTIME_EXTENSIONS) renders the Gemma 4 E2B template byte-identical to Jinja2 for user+tools, no-tools, and tool-call history (when combined with Preserve Gemma 4 tool definitions and add Minja upper filter #1129).
• Blast radius (differential test): I rendered 45 real chat templates (21 in this repo plus 24 popular HF models: Llama 3.x, Qwen2.5/3, Phi-3.5/4, Mistral, DeepSeek, Granite, SmolLM3, gpt-oss, Gemma 3/4, GLM-4.5, LFM2, Nemotron, Falcon3, OLMo2, Kimi-K2, ERNIE), each in 8 conversation scenarios, before and after this change through GenAI rel-0.17.1. 356 of the 360 renders were byte-identical. The only 4 differences are Gemma 4 tool-call-history cases, which now match Jinja2. Only 7 of the 45 templates use {% macro %} at all.
• Jinja2 parity: across the macro templates, no case went from matching Jinja2 to not matching. The existing GLM-4.5/Kimi-K2/LFM2 gaps are the same before and after (unrelated Minja limitations).
• Latency: no measurable change. Gemma 4 renders took about 250 µs each with or without the fix, over 3 rounds of 300 renders.

Reviewer questions

1. Is aligning macro scoping with Jinja2/upstream Minja acceptable even though a template relying on the old leakage would change output?
2. Gemma 4 needs both this PR and Preserve Gemma 4 tool definitions and add Minja upper filter #1129 for Jinja2-identical output with tools. They are independent; any merge order works.
3. Separately, MacroNode captures [&] while upstream captures [this, macro_context] by value. Fold that hardening in here or track it separately?

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Copilot review overview

🟢 Approval recommended

The focused implementation matches the intended semantics and is covered by relevant regressions.

Review effort: Balanced
Findings: None

What changed in this PR

Fixes Gemma 4’s extra generation prompt by aligning Minja macro scoping with Jinja2 and upstream Minja.

Changes:

• Creates an isolated child context for each macro invocation.
• Adds regression coverage for macro scope, recursion, namespace mutation, defaults, and Gemma 4 tool responses.

File	Description
shared/​api/​minja.hpp	Isolates macro-call variables from their defining context.
test/​pp_api_test/​test_tokenizer_chat.cc	Tests macro semantics and Gemma 4 output.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Sayan Shaw (sayanshaw24)]

/azp run onnxruntime-extensions.ci

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Copilot]

🟡 Changes recommended

Parameter-dependent defaults still read the caller’s scope rather than the newly bound macro arguments.

1 open finding

• Evaluate macro defaults in call context to resolve earlier parameters · New

🧠 Review effort: Balanced

[Copilot]

🟢 Approval recommended

The focused changes address the prior default-argument issue and have targeted regression coverage with no unresolved blocking findings.

No open issues 🎉

1 resolved since last review

• Evaluate macro defaults in call context to resolve earlier parameters

_{Review effort: Balanced}