openclaw - 💡(How to fix) Fix OpenAI Responses incremental context can replay orphan toolResult without matching toolCall [2 comments, 1 participants]

Summary

E2E test issue from IYENsystem/OpenClaw integration: OpenAI Responses incremental turns can fail when a stale/orphan toolResult is replayed without the matching assistant toolCall in the same incremental input window.

Observed behavior

In long-running OpenClaw sessions with tool use + compaction/runtime context reuse, the next OpenAI Responses turn can receive a toolResult whose call_id belongs to an older assistant toolCall that is no longer present in the incremental slice. When that happens, the provider rejects the turn / context continuation, which looks like the assistant “context exploding” or failing after tool use.

Expected behavior

When using previous_response_id incremental sending, OpenClaw should only send tool results whose matching assistant tool call is present in the incremental input. If the incremental slice contains orphan/stale/mixed tool results, the runtime should fall back to a full-context restart instead of sending an invalid incremental turn.

Local hotfix already tested

A local hotfix was applied to planTurnInput() in the installed OpenClaw dist build:

• Track assistant toolCall.call_id values after lastContextLength.
• Only include incremental toolResult entries when their call_id matches a tool call in the same incremental window.
• If orphan/stale/mixed tool results are detected, switch to full_context_restart.

Local verification:

• node --check passed.
• A small reproduction script confirmed:
  • valid paired toolCall/toolResult keeps incremental mode
  • orphan toolResult forces full restart
  • mixed valid+orphan toolResults forces full restart

Why this issue exists

This issue is also being used as an E2E routing test for issue-created/context-gathering automation. It is intentionally filed as a real bug report with the known reproduction shape so the downstream automation has concrete context to inspect.