Fix ReAct agent parsing failures with reasoning models (<think> tags)#1667
Merged
rapids-bot[bot] merged 3 commits intoNVIDIA:release/1.5from Feb 26, 2026
Merged
Conversation
Signed-off-by: Yuchen Zhang <yuchenz@nvidia.com>
WalkthroughBase agent separates streaming LLM content and reasoning content and returns both (reasoning via additional_kwargs); ReAct agent normalizes outputs by stripping R1 think tags, falls back to extracting ... or reasoning_content, treats certain non-ReAct outputs as direct final answers, and adds fallback prompts on parse failures. Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes 🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
🧹 Nitpick comments (1)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py (1)
183-193: Extract duplicated content normalization logic into a helper method.This normalization logic is duplicated at lines 217-227. Consider extracting it into a private method to improve maintainability.
♻️ Proposed refactor
Add a helper method to the class:
def _normalize_reasoning_model_output(self, output_message: AIMessage) -> None: """Normalize output from reasoning models by stripping think tags and extracting content.""" if not isinstance(output_message.content, str): return raw_content = output_message.content output_message.content = remove_r1_think_tags(raw_content) if not output_message.content.strip(): think_match = re.search(r'<think>(.*?)</think>', raw_content, re.DOTALL) if think_match: output_message.content = think_match.group(1).strip() if not output_message.content.strip(): reasoning = output_message.additional_kwargs.get('reasoning_content', '') if reasoning: output_message.content = reasoningThen replace both occurrences with:
- if isinstance(output_message.content, str): - raw_content = output_message.content - output_message.content = remove_r1_think_tags(raw_content) - if not output_message.content.strip(): - think_match = re.search(r'<think>(.*?)</think>', raw_content, re.DOTALL) - if think_match: - output_message.content = think_match.group(1).strip() - if not output_message.content.strip(): - reasoning = output_message.additional_kwargs.get('reasoning_content', '') - if reasoning: - output_message.content = reasoning + self._normalize_reasoning_model_output(output_message)🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py` around lines 183 - 193, Extract the duplicated content-normalization logic into a private helper on the class (e.g., def _normalize_reasoning_model_output(self, output_message: AIMessage) -> None) that returns early if output_message.content is not a str, calls remove_r1_think_tags(raw_content), falls back to re.search(r'<think>(.*?)</think>', raw_content, re.DOTALL) to extract inner text if the stripped content is empty, and finally uses output_message.additional_kwargs.get('reasoning_content', '') as the last fallback; then replace both inline blocks (the one around output_message handling at lines shown and the duplicate at 217-227) with calls to self._normalize_reasoning_model_output(output_message) and add a short docstring to the helper for clarity.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Nitpick comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py`:
- Around line 183-193: Extract the duplicated content-normalization logic into a
private helper on the class (e.g., def _normalize_reasoning_model_output(self,
output_message: AIMessage) -> None) that returns early if output_message.content
is not a str, calls remove_r1_think_tags(raw_content), falls back to
re.search(r'<think>(.*?)</think>', raw_content, re.DOTALL) to extract inner text
if the stripped content is empty, and finally uses
output_message.additional_kwargs.get('reasoning_content', '') as the last
fallback; then replace both inline blocks (the one around output_message
handling at lines shown and the duplicate at 217-227) with calls to
self._normalize_reasoning_model_output(output_message) and add a short docstring
to the helper for clarity.
ℹ️ Review info
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (2)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/base.pypackages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py
willkill07
approved these changes
Feb 26, 2026
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (1)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py (1)
182-192: Extract duplicated normalization/fallback logic into one helper.The same block appears twice; centralizing it avoids drift and keeps first/subsequent cycle behavior consistent.
Refactor sketch
+ `@staticmethod` + def _normalize_agent_output_content(output_message: AIMessage) -> None: + if not isinstance(output_message.content, str): + return + raw_content = output_message.content + output_message.content = remove_r1_think_tags(raw_content) + if not output_message.content.strip(): + think_match = re.search(r"<think>(.*?)</think>", raw_content, re.DOTALL) + if think_match: + output_message.content = think_match.group(1).strip() + if not output_message.content.strip(): + reasoning = output_message.additional_kwargs.get("reasoning_content", "") + if isinstance(reasoning, str) and reasoning.strip(): + output_message.content = reasoning.strip()- if isinstance(output_message.content, str): - ... + self._normalize_agent_output_content(output_message)Also applies to: 216-226
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py` around lines 182 - 192, Extract the duplicated normalization/fallback logic that processes output_message.content (calls remove_r1_think_tags, falls back to <think> tag content, then to additional_kwargs['reasoning_content']) into a single helper function (e.g., normalize_output_message_content(output_message) or similar) and replace both code blocks (the one around output_message handling at the current location and the other at the later occurrence) with calls to that helper; ensure the helper accepts an OutputMessage-like object, performs the remove_r1_think_tags -> check empty -> extract <think> -> check empty -> use reasoning_content flow, and preserves trimming/strip behavior so first and subsequent cycle behavior stays identical.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py`:
- Around line 182-192: The code assigns output_message.content = reasoning from
output_message.additional_kwargs['reasoning_content'] without validating its
type; guard that value in the blocks inside the remove_r1_think_tags flow and
the similar block later (the code around output_message, additional_kwargs and
reasoning_content) by checking isinstance(reasoning, str) before assigning, and
if it is not a str either skip the assignment or coerce safely (e.g., convert to
str(reasoning) only if that behavior is acceptable), ensuring downstream text
parsing always receives a string.
- Around line 281-290: The fallback that accepts parse-failed outputs as a final
answer (the block using ex.missing_action, content_str, AGENT_LOG_PREFIX,
state.messages, and state.final_answer) is too permissive because it only
rejects content that starts with "Thought:"; update the gating so you only
accept content as a final answer when it contains no ReAct markers anywhere.
Specifically, modify the condition around ex.missing_action to scan content_str
(case-insensitive) for any occurrences of ReAct tokens such as "thought:",
"action:", "observation:", "action input:" (anchored at line starts or
standalone tokens) and only proceed to log, append AIMessage, set
state.final_answer and return state when none of those markers are present;
otherwise continue allowing tools to run or raise the parse error path. Ensure
you use the existing variables (content_str, ex.missing_action, state,
AGENT_LOG_PREFIX) so the change is localized to that block.
---
Nitpick comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py`:
- Around line 182-192: Extract the duplicated normalization/fallback logic that
processes output_message.content (calls remove_r1_think_tags, falls back to
<think> tag content, then to additional_kwargs['reasoning_content']) into a
single helper function (e.g., normalize_output_message_content(output_message)
or similar) and replace both code blocks (the one around output_message handling
at the current location and the other at the later occurrence) with calls to
that helper; ensure the helper accepts an OutputMessage-like object, performs
the remove_r1_think_tags -> check empty -> extract <think> -> check empty -> use
reasoning_content flow, and preserves trimming/strip behavior so first and
subsequent cycle behavior stays identical.
ℹ️ Review info
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py
Show resolved
Hide resolved
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py
Show resolved
Hide resolved
Signed-off-by: Yuchen Zhang <yuchenz@nvidia.com>
There was a problem hiding this comment.
♻️ Duplicate comments (2)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py (2)
189-192:⚠️ Potential issue | 🟡 MinorGuard
reasoning_contenttype before assignment.The
additional_kwargsdictionary is provider-defined and may contain non-string values. Add a type check to prevent downstream parsing failures.Suggested fix
if not output_message.content.strip(): reasoning = output_message.additional_kwargs.get('reasoning_content', '') - if reasoning: + if isinstance(reasoning, str) and reasoning.strip(): output_message.content = reasoning🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py` around lines 189 - 192, The code assigns output_message.content = reasoning from output_message.additional_kwargs['reasoning_content'] without validating type; update the branch in the block handling output_message so you first fetch reasoning = output_message.additional_kwargs.get('reasoning_content', None) and only assign it if it is a non-empty string (e.g., isinstance(reasoning, str) and reasoning.strip()), otherwise leave output_message.content unchanged (or coerce safely to string only after explicit validation) to avoid downstream parsing errors.
278-291:⚠️ Potential issue | 🟡 MinorConsider stricter validation for direct-answer fallback.
The current check only rejects content starting with
Thought:,Question:, orPrevious conversation. Content withAction:orObservation:markers on later lines could be incorrectly accepted as a final answer.Since
ex.missing_actionis already true (meaning no valid action was parsed), this may be acceptable. However, if you want stricter validation to avoid accepting malformed ReAct outputs:Suggested stricter check
content_str = str(output_message.content).strip() - if (ex.missing_action and content_str and not re.match( - r'\s*(thought\s*:|question\s*:|previous\s+conversation)', content_str, re.IGNORECASE)): + react_marker_pattern = re.compile( + r'^\s*(thought|action|action\s*input|observation)\s*:', + re.IGNORECASE | re.MULTILINE, + ) + if ex.missing_action and content_str and not react_marker_pattern.search(content_str):🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py` around lines 278 - 291, The direct-answer fallback currently accepts any non-empty content when ex.missing_action is true as long as it doesn't start with Thought:/Question:/Previous..., which can still let ReAct markers like "Action:" or "Observation:" appear later and be accepted; update the validation in the block that builds content_str (used with ex.missing_action, AGENT_LOG_PREFIX, state.messages, AIMessage, state.final_answer) to perform a stricter check—either reject when any line (not just the first) matches common ReAct markers (e.g., Action:, Observation:, Input:, Answer:) using a regex search (re.search) or require the first non-empty line to be free of ReAct markers—so malformed ReAct outputs are not treated as final answers.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Duplicate comments:
In
`@packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py`:
- Around line 189-192: The code assigns output_message.content = reasoning from
output_message.additional_kwargs['reasoning_content'] without validating type;
update the branch in the block handling output_message so you first fetch
reasoning = output_message.additional_kwargs.get('reasoning_content', None) and
only assign it if it is a non-empty string (e.g., isinstance(reasoning, str) and
reasoning.strip()), otherwise leave output_message.content unchanged (or coerce
safely to string only after explicit validation) to avoid downstream parsing
errors.
- Around line 278-291: The direct-answer fallback currently accepts any
non-empty content when ex.missing_action is true as long as it doesn't start
with Thought:/Question:/Previous..., which can still let ReAct markers like
"Action:" or "Observation:" appear later and be accepted; update the validation
in the block that builds content_str (used with ex.missing_action,
AGENT_LOG_PREFIX, state.messages, AIMessage, state.final_answer) to perform a
stricter check—either reject when any line (not just the first) matches common
ReAct markers (e.g., Action:, Observation:, Input:, Answer:) using a regex
search (re.search) or require the first non-empty line to be free of ReAct
markers—so malformed ReAct outputs are not treated as final answers.
ℹ️ Review info
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (2)
packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/base.pypackages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/react_agent/agent.py
🚧 Files skipped from review as they are similar to previous changes (1)
- packages/nvidia_nat_langchain/src/nat/plugins/langchain/agent/base.py
Contributor
Author
|
/merge |
Charlie-Yi-2002
pushed a commit
to Charlie-Yi-2002/NeMo-Agent-Toolkit
that referenced
this pull request
Mar 5, 2026
…s) (NVIDIA#1667) - Reasoning models (e.g., nemotron-3-nano-30b-a3b) wrap their chain-of-thought in <think>...</think> tags, which breaks the ReAct output parser. This PR adds multi-layered handling to recover usable content from reasoning model outputs. - Preserve reasoning_content from streamed LLM chunks in _stream_llm, which previously discarded additional_kwargs when constructing the final AIMessage - Strip <think> tags from LLM output before parsing, with fallbacks: extract content from inside tags, then fall back to reasoning_content from additional_kwargs - Accept direct (non-ReAct-formatted) answers from reasoning models instead of failing with ReActAgentParsingFailedError - Improve retry hints when content is empty, guiding the model toward producing a Final Answer: ## By Submitting this PR I confirm: - I am familiar with the [Contributing Guidelines](https://github.com/NVIDIA/NeMo-Agent-Toolkit/blob/develop/docs/source/resources/contributing/index.md). - We require that all contributors "sign-off" on their commits. This certifies that the contribution is your original work, or you have rights to submit it under the same license, or a compatible license. - Any contribution which contains commits that are not Signed-Off will not be accepted. - When the PR is ready for review, new or existing tests cover these changes. - When the PR is ready for review, the documentation is up to date with these changes. ## Summary by CodeRabbit * **Bug Fixes** * Streaming responses from reasoning models now capture and combine main content and separate reasoning streams reliably. * Agents more reliably recognize direct final answers even when not in strict ReAct format, reducing missed responses. * Improved fallback behavior that adds clear guidance prompts when parsing fails. * **Improvements** * More consistent normalization of agent outputs across initial and subsequent interaction flows for steadier results. Authors: - Yuchen Zhang (https://github.com/yczhang-nv) Approvers: - Will Killian (https://github.com/willkill07) URL: NVIDIA#1667
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
By Submitting this PR I confirm:
Summary by CodeRabbit
Bug Fixes
Improvements