feat(env): integrate CodeDebug environment and vLLM inference stabilization#3448
Open
RUFFY-369 wants to merge 19 commits intoNousResearch:mainfrom
Open
feat(env): integrate CodeDebug environment and vLLM inference stabilization#3448RUFFY-369 wants to merge 19 commits intoNousResearch:mainfrom
RUFFY-369 wants to merge 19 commits intoNousResearch:mainfrom
Conversation
Extends HermesAgentBaseEnv with: - HumanEvalPack dataset (164 buggy Python functions) - Workspace scaffolding (buggy.py + tests.py uploaded to sandbox) - Multi-signal reward: test_signal (0.5), diagnosis (0.3), efficiency (0.2) - Terminal + file toolsets for iterative debugging
…-ASCII characters
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Note
Research Context: This PR integrates the physical environment provided in atropos (PR #421) and establishes the foundation for the MT-GRPO reward infrastructure in PR #3451.
What does this PR do?
This PR integrates the CodeDebug environment for agentic reasoning and introduces a Universal Tool Strategy to stabilize inference on vLLM backends.
It solves two primary problems:
Lack of Agentic Debugging Benchmarks
Provides a high-fidelity sandbox where agents iteratively debug code using real terminal and file tools.
vLLM Tool Parsing Bugs
Implements a robust client-side tool parsing fallback in
model_tools.pyandenvironments/agent_loop.pyto bypass known bugs in vLLM 0.6.5 that cause 400/500 errors during complex tool-use sessions.Related Issue
Fixes # (Initial integration of debugging environment)
Type of Change
Changes Made
environments/agent_loop.py: Added client-side tool parsing loop andatropos_inhibit_toolssupportmodel_tools.py: Addedparse_tool_calls_from_text(regex-based fallback parser)environments/code_debug_env/code_debug_env.py: Environment logic based on HumanEvalFixenvironments/code_debug_env/default.yaml: Configuration for agent prompts and tool inhibitionenvironments/code_debug_env/README.md: Usage and stabilization documentationHow to Test
Rollout Verification
vLLM Stability
<tool_code>tagsMock Trajectories
Checklist
Code
fix(scope):,feat(scope):, etc.)pytest tests/ -qand all tests pass (Note: system-level version conflict in env, but rollout verified)Documentation & Housekeeping
docs/, docstrings)cli-config.yaml.example— or N/ACONTRIBUTING.mdorAGENTS.md— or N/AFor New Skills
N/A — This is an RL training environment integration.
Screenshots / Logs
cc @teknium1