Human-in-the-Loop Implementation#2021

Open

mjschock wants to merge 37 commits intoopenai:mainfrom

Commits on Dec 12, 2025

feat: implement human-in-the-loop (HITL) approval infrastructure

mjschock
and
claude
committed
feat: integrate HITL approval checking into run execution loop

mjschock
and
claude
committed
feat: add RunState parameter support to Runner.run() methods

mjschock
and
claude
committed
feat: add to_state() method to RunResult for resuming runs

mjschock
and
claude
committed
feat: add streaming HITL support and complete human-in-the-loop implementation

mjschock
and
claude
committed
fix: prime server conversation tracker in streaming path to prevent message duplication

mjschock
and
claude
committed
ci: fix issues that surfaced in CI
mjschock
committed
fix: Bring up coverage to minimum and add session hitl examples
mjschock
committed
fix: ensure RunState serialization compatibility with openai-agents-js
mjschock
committed
fix: standardize call_id extraction in ServerConversationTracker
mjschock
committed
test: add tests for RunState resumption and serialization to bring up coverage
mjschock
committed
fix: Updates following rebase, include test coverage
mjschock
committed
fix: address issues around resuming run state with conversation history
mjschock
committed
fix: address duplicating session history issue mentioned by @chatgpt-codex-connector
mjschock
committed
fix: update RunState with current turn persisted item tracking
mjschock
committed
fix: addressing edge cases when resuming
mjschock
committed
fix: addressing edge cases when resuming (continued)
mjschock
committed
fix: addressing rebase issues
mjschock
committed
fix: improving parity with openai-agent-js hitl functionality
mjschock
committed
fix: bring coverage back up, addressing edge cases
mjschock
committed
fix: cleanup
mjschock
committed
fix: rename summary_message back to assistant_message
mjschock
committed
fix: enhance agent state management during resume, ensuring correct agent usage and saving tool outputs to session
mjschock
committed
fix: finish up human-in-the-loop port
mjschock
committed
fix: add auto_previous_response_id parameter to test_start_streaming_fresh_run_exercises_persistence
mjschock
committed
fix: typing updates to pass `make old_version_tests`
mjschock
committed
fix: remove dead code and add failing hitl error scenarios
mjschock
committed
fix: address failing hitl error scenarios
mjschock
committed
fix: change logging level for _ServerConversationTracker creation to debug
mjschock
committed
fix: pass context_wrapper to _coerce_apply_patch_operation in ApplyPatchAction
mjschock
committed
fix: add test to ensure current turn is preserved when converting RunResult to RunState
mjschock
committed
fix: update RunResult to track current turn number and ensure it is preserved during state conversion
mjschock
committed
fix: add tests to ensure ToolApprovalItem hashability and preserve persisted item count during streamed run resumption
mjschock
committed
refactor: simplify condition for current turn persisted item count in AgentRunner
mjschock
committed
test: add tests to preserve tool output types during run state serialization
mjschock
committed
fix: enhance output item conversion logic to preserve non-function call types in RunState
mjschock
committed
fix: update test to use ResponseCustomToolCall for apply_patch call in HITL coverage
mjschock
committed