fix: Fix openai responses api issues

[ncrispino]

Fix GPT-5 Reasoning Token Errors and Double Voting

Summary

This PR fixes two critical issues affecting GPT-5 reasoning models and multi-agent voting:

1. GPT-5-mini (and other GPT-5-X models) reasoning token errors - Fixed malformed Response API requests causing 400 errors
2. Double voting handling - Gracefully handle when reasoning models make multiple vote calls

Closes MAS-181

Issues Fixed

Issue 1: GPT-5 Reasoning Token Errors

Error:

Error code: 400 - {'error': {'message': "Item 'rs_...' of type 'reasoning' was provided without its required following item."}}

Root Causes:

1. The id field was being stripped from function_call items, breaking reasoning-to-function-call pairing required by OpenAI's Response API
2. When using previous_response_id, response items were being added both manually AND automatically, causing duplicate reasoning items

Fixes:

• massgen/formatter/_response_formatter.py: Preserve id field in function_call items (required for reasoning pairing per LangChain PR #9082)
• massgen/backend/response.py: Only add response items manually when NOT using previous_response_id to avoid duplicates

Issue 2: Double Voting in Multi-Agent Workflows

Issue: Certain GPT-5 models (particularly gpt-5.1 and gpt-5.2) make multiple vote calls despite enforcement messages, causing workflow failures.

Model Behavior:

• ❌ gpt-5.1, gpt-5.2: Ignore tool error enforcement, repeat violations
• ✅ gpt-5(-X), gpt-5.1-codex: Respond to enforcement correctly

Fix: Instead of rejecting multiple votes, gracefully handle by taking the last vote as the agent's final decision.

Rationale:

• The last vote represents the agent's most refined thinking
• Reasoning models iterate through their logic - later votes are more informed
• Simpler than vote counting and more aligned with how reasoning models work

Changes

Files Modified

1. massgen/formatter/_response_formatter.py

   • Preserve id field for reasoning item pairing
   • Added reference to LangChain's similar fix

2. massgen/backend/response.py

   • Conditional response item addition based on response_id presence
   • Deduplicate items by ID to prevent duplicate reasoning tokens
   • Improved logging (debug level instead of info)

3. massgen/orchestrator.py

   • Simplified multiple vote handling (38 lines → 14 lines)
   • Take last vote instead of complex counting/enforcement
   • Removed unnecessary debug logging
   • Cleaner variable names

Lines Changed

massgen/backend/response.py              | +16 -3
massgen/formatter/_response_formatter.py | +5 -2
massgen/orchestrator.py                  | +14 -38
Total: +35 -43 (net -8 lines, cleaner code)

Testing

Test Case 1: Reasoning Models

uv run massgen --automation --model gpt-5-mini "Create a simple website"

Before: 400 error about reasoning tokens
After: ✅ Completes successfully

Test Case 2: Multi-Agent Voting with gpt-5.1/5.2

uv run massgen --config config.yaml "Create a simple website"

Before: Agent execution failed: 'agent1' (KeyError)
After:

⚠️ Agent made 2 votes - using last (final decision): agent1
🏆 Turn 1 winner: agent_a

Verified Behavior

• ✅ No reasoning token errors with GPT-5 models
• ✅ Warning shown when multiple votes detected
• ✅ Execution completes successfully using last vote
• ✅ Anonymous ID mapping works correctly (agent1 → agent_a)

Related Documentation

• LangChain PR #9082: fix(openai): pair reasoning with function_call id; add tests langchain-ai/langchainjs#9082 (same reasoning token issue)
• OpenAI Response API: Requires strict item ordering for reasoning models
• MassGen Issue: Model-specific behavior where gpt-5.1/5.2 don't respect tool error enforcement

Known Limitations

GPT-5 Model Variants: Some models (gpt-5.1, gpt-5.2) don't properly respond to tool error enforcement. This fix works around the limitation by handling multiple votes gracefully rather than relying on enforcement.

Recommendation: For critical voting scenarios, prefer models that respond to enforcement (gpt-5(-X), gpt-5.1-codex) or accept that multiple votes will be deduplicated to the last vote.

Migration Notes

No breaking changes. This is purely a bug fix that:

• Makes reasoning models work correctly
• Handles edge cases more gracefully
• Improves user experience with better messaging

Users will see informational warnings when multiple votes are detected, but workflows will continue successfully.