feat: #636 Add human-in-the-loop (HITL) support to the SDK#2230

Open

seratch wants to merge 30 commits intomainfrom

issue-636-hitl-2

Commits on Jan 12, 2026

feat: #636 Add human-in-the-loop (HITL) support to the SDK

seratch
and
mjschock
committed
split _run_impl.py into run_internal/
seratch
committed
Simplify the run state data
seratch
committed
Fix reported issues
seratch
committed
Fix HITL resume for computer actions and avoid duplicate rejections
seratch
committed
Fix automatic Responses compaction trigger during session persistence
seratch
committed
Preserve guardrail history when resuming from RunState
seratch
committed
fix python 3.9 errors
seratch
committed
Preserve conversation tracking when resuming runs
seratch
committed
Include CompactionItem in RunItem union for type safety
seratch
committed
fix: skip re-executing function tools on HITL resume when outputs exist
seratch
committed
fix: deserialize compaction_item in RunState
seratch
committed
fix: scope tool rejection to call ids
seratch
committed
fix: keep approved tool outputs on HITL resume with pending approvals
seratch
committed
feat: add RunState context serialization hooks and metadata
seratch
committed
fix: rebuild HITL function runs from object approvals
seratch
committed
fix: normalize tool-call dedupe keys for unhashable arguments
seratch
committed
fix: harden HITL resume dedupe and nested tool approvals
seratch
committed
fix: make RunState serialization tolerant of non-JSON outputs
seratch
committed
feat: add HITL session scenario example and tests
seratch
committed
fix: surface needs_approval errors on HITL resume
seratch
committed
fix: dedupe tool calls by call_id or id and centralize MCP approval partitioning
seratch
committed
Centralize tool call deduplication logic
seratch
committed
fix: honor filtered inputs in conversation tracking and preserve duplicate messages
seratch
committed
refactor; add comments
seratch
committed
fix: ignore fake response ids in dedupe and strip provider_data for OpenAI conversations
seratch
committed
Fix session persistence counter handling and add regression test
seratch
committed
fix: ignore fake response ids in conversation tracking
seratch
committed
fix: align OpenAI conversation persistence counts with sanitized items
seratch
committed

Commits on Jan 13, 2026

Refine run loop cleanup and streaming retry removal
seratch
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: #636 Add human-in-the-loop (HITL) support to the SDK#2230

feat: #636 Add human-in-the-loop (HITL) support to the SDK#2230
seratch wants to merge 30 commits intomainfrom
issue-636-hitl-2

Commits on Jan 12, 2026

feat: #636 Add human-in-the-loop (HITL) support to the SDK

split _run_impl.py into run_internal/

Simplify the run state data

Fix reported issues

Fix HITL resume for computer actions and avoid duplicate rejections

Fix automatic Responses compaction trigger during session persistence

Preserve guardrail history when resuming from RunState

fix python 3.9 errors

Preserve conversation tracking when resuming runs

Include CompactionItem in RunItem union for type safety

fix: skip re-executing function tools on HITL resume when outputs exist

fix: deserialize compaction_item in RunState

fix: scope tool rejection to call ids

fix: keep approved tool outputs on HITL resume with pending approvals

feat: add RunState context serialization hooks and metadata

fix: rebuild HITL function runs from object approvals

fix: normalize tool-call dedupe keys for unhashable arguments

fix: harden HITL resume dedupe and nested tool approvals

fix: make RunState serialization tolerant of non-JSON outputs

feat: add HITL session scenario example and tests

fix: surface needs_approval errors on HITL resume

fix: dedupe tool calls by call_id or id and centralize MCP approval partitioning

Centralize tool call deduplication logic

fix: honor filtered inputs in conversation tracking and preserve duplicate messages

refactor; add comments

fix: ignore fake response ids in dedupe and strip provider_data for OpenAI conversations

Fix session persistence counter handling and add regression test

fix: ignore fake response ids in conversation tracking

fix: align OpenAI conversation persistence counts with sanitized items

Commits on Jan 13, 2026

Refine run loop cleanup and streaming retry removal