parser: bypass tool parser for no-tools streams

[voipmonitor]

Summary

Fixes a GLM-5.2 streaming corruption where plain post-reasoning content was routed through the tool parser even when the request did not provide any tools.

With --enable-auto-tool-choice, vLLM constructs a tool parser globally. After </think>, DelegatingParser._in_tool_call_phase() became true whenever that parser existed, regardless of whether the request had a tools list. For no-tools requests, the engine-based tool parser can re-emit buffered plain content and duplicate fragments in streamed deltas.

Observed on GLM-5.2 NVFP4/B12X MTP3 long-context streaming:

output_processor text: " 2:\n       "
served delta content: "< 2:\n        2:\n       "

This produced client-visible code corruption such as:

if limit < limit < 2:
    2:

Changes

• Add a small request-level tools check in DelegatingParser.
• Bypass tool-call extraction for post-reasoning content when request.tools is empty/missing.
• Skip engine tool-parser final flush if tool-call streaming never started.
• Update parser-engine tests so tool-parser flush tests explicitly use tools, and add a no-tools regression test.

Validation

• Syntax check: python3 -m py_compile vllm/parser/abstract_parser.py tests/parser/engine/test_parser_engine.py
• Whitespace check: git diff --check
• Runtime soak on the live GLM-5.2 NVFP4/B12X MTP3 instance with this file overlaid: 30/30 /mnt/test.py long-context streaming iterations passed without the if limit < ... limit < corruption.

Note: upstream vLLM PR vllm-project#46149 fixes a related but separate forced/named tool-choice + reasoning grammar issue. This PR intentionally does not include that change.

Summary by CodeRabbit

• Bug Fixes
  • Improved handling of generated content when no tools are configured. Content generated after reasoning is no longer incorrectly interpreted as tool call markup, preventing spurious parsing errors.
  • Added validation to bypass tool parsing logic when the request has no tools enabled.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 3f151d12-8e87-4479-a21e-5b75c49b8dd3

📥 Commits

Reviewing files that changed from the base of the PR and between 774f42d and 6987a34.

📒 Files selected for processing (2)

• tests/parser/engine/test_parser_engine.py
• vllm/parser/abstract_parser.py

---

📝 Walkthrough

Walkthrough

Parser gains a _request_has_tools() helper. parse_delta() now routes post-</think> content directly to delta_message.content when the request has no tools, and _flush_engine_parsers() skips flushing the tool parser when tool_call_text_started is false. Tests are updated to carry a with_tools flag and a new regression test verifies the no-tools bypass path.

Changes

No-tools tool-parser bypass

Layer / File(s)	Summary
Tool-parser bypass logic in Parser vllm/parser/abstract_parser.py	Adds Parser._request_has_tools() static helper; guards parse_delta() tool-call phase to write current_text directly into delta_message.content when the request has no tools; conditionally skips _tool_parser flush in _flush_engine_parsers() when tool_call_text_started is false.
Test helper and regression test tests/parser/engine/test_parser_engine.py	Extends _make_delegating_request with with_tools: bool = False; updates test_args_buffer_flushed_on_finished to use with_tools=True; adds test_no_tools_content_bypasses_tool_parser_after_reasoning asserting content accumulates normally and tool_call_text_started stays false for no-tools requests.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title 'parser: bypass tool parser for no-tools streams' directly and concisely describes the main change: bypassing the tool parser when requests have no tools configured, which is the core fix addressing the streaming corruption issue.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/glm52-no-tools-parser-bypass-20260619

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}