OpenAI Agents Python SDK releases

3 items in the render window (newest 100 + all corrections; not this source's all-time total) · status healthy · last successful fetch 2026-11-21 19:16 UTC
Feed for this sourceRSS
ACTION NEEDED APIDEPRECATION OpenAI Agents Python SDK releases · 2026-11-21 07:11 UTC

OpenAI Agents Python SDK v1.0.0-alpha.1 breaks Agent.run() and Session.submit() APIs

Do:Before upgrading, complete all pending native runs and approvals in your v0.24.0-alpha.8 environment.
WHAT CHANGEDOpenAI released v1.0.0-alpha.1 of the Agents Python SDK with breaking API changes. Agent.run() and Session.submit() from v0.24.0-alpha.8 are replaced by Runner.run() and Runner.run_streamed(). SessionState checkpoints from alpha.8 are incompatible with the new Runner's RunState format. The release adds Codex as an opt-in model provider via CodexProvider and is distributed exclusively through GitHub Releases, not PyPI. Requires Python 3.10–3.14.
WHY IT MATTERSExisting code using Agent.run() or Session.submit() will break on upgrade. Saved SessionState checkpoints from v0.24.0-alpha.8 cannot be resumed in v1.0.0-alpha.1, requiring fresh SDK sessions for conversation history. Any automation or CI that expects PyPI distribution will fail to find the package.
WHAT TO DOBefore upgrading, complete all pending native runs and approvals in your v0.24.0-alpha.8 environment. Migrate execution calls from Agent.run()/Session.submit() to Runner.run()/Runner.run_streamed() and import Agent, Runner, and RunConfig from agents instead of agents.classic. Update installation scripts to fetch wheels from GitHub Releases using the platform-specific URLs provided in the release notes.
Model-reported confidence 95% (not measured accuracy) Primary source ↗ written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
+## Agents SDK v1.0.0-alpha.1
+https://github.com/openai/openai-agents-python/releases/tag/v1.0.0-alpha.1
+2026-11-20T20:51:28Z
+
+Agents SDK v1.0.0-alpha.1
+This first preview from feat/simple-codex adds Codex as an opt-in model provider for the Python Agents SDK. The prebuilt openai-agents wheels (version 1.0.0a1) are available exclusively through GitHub Releases.
+Highlights
+
+Choose Codex per run. Import CodexProvider from agents.models.codex and create provider = CodexProvider(). Pass RunConfig(model_provider=provider) to the Runner. Codex uses your saved login. For API key authentication, use CodexProvider(api_key=...). The provider selects the model for agents that use a model name or the default model. Run the same agent with either provider
+Reuse Codex Core inside Python. The wheel embeds unmodified Codex Core rust-v0.166.0 through PyO3. Direct calls to Core eliminate codex app-server startup and JSON-RPC. Reuse the provider to share one ThreadManager across calls. Each call gets a fresh thread. Use an SDK session for history and close the provider with await provider.aclose() when you finish. Thread lifetime and history · Codex source
+Use SDK tools and approvals. Stream text, hand off to a specialist, pause for approval before a Python tool runs, and validate typed output. Native commands follow Codex's sandbox and approval settings. A native approval prompt ends the model call. Execution and approvals · Supported tools and model settings
+
+Install and try it
+Activate a virtual environment with standard CPython 3.10–3.14. Copy the link for your platform and replace WHEEL_URL in the command below. Each link includes the SHA-256 digest that pip checks when it downloads the wheel. Checksums for all four wheels
+Linux x86-64 (glibc 2.28+) · Linux ARM64 (glibc 2.28+) · macOS 11+ on Apple silicon · Windows x64
+python -m pip install --only-binary=:all: "WHEEL_URL"
+To try the wildlife example, check out v1.0.0-alpha.1 and run the command below from the repository root with the wheel installed. If your only credential is a saved Codex login, set OPENAI_AGENTS_DISABLE_TRACING=1 in your shell first. Wildlife example · Authentication
+python -I examples/model_providers/codex_provider.py --codex
+Fixed
+
+Automatic session compaction uses local conversation history for Codex, including after an approval pause. Compaction needs its own OpenAI client or credentials. Session compaction
+Saved sandbox manifests preserve inline binary files when you resume a run. If you share these manifests, upgrade all readers before saving non-UTF-8 content. Use LocalFile when older SDKs need to read the manifest. Manifest compatibility
+
+Moving from the harness preview
+If you use Agent.run() or Session.submit() in v0.24.0-alpha.8, move execution to Runner.run() or Runner.run_streamed() and select CodexProvider in your run config. If you use agents.classic, import Agent, Runner, and RunConfig from agents. Previous harness release · Runner example
+Finish pending native runs and approvals in your alpha.8 environment. Its SessionState checkpoints use a different format from the Runner's RunState. Start a fresh SDK session for conversation history in alpha.1. SDK sessions
+Release verification
+The SDK suite passed on the merged feat/simple-codex commit. CI built all four wheels and tested fresh installs on CPython 3.10 and 3.14 on each platform. We verified all four downloads against SHA256SUMS. Source 87e8f533 · SDK tests · Wheel builds and installation checks
+We also ran the wildlife example with a live model and the Linux x86-64 wheel on CPython 3.12. The specialist identified an American bison from the photo, called the Python catalog tool, and returned a typed report with record ID synthetic-bison-001. Example and assertions
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
triage verdictMATERIAL at 0.92, needed 0.5
triage reasoningMajor version with breaking API changes, new provider, deployment model shift, and session incompatibility.
signalsnew major version release (v1.0.0-alpha.1) · new model provider added (Codex) · breaking API changes (Agent.run() → Runner.run(), Session.submit() → Runner.run_streamed()) · authentication method changes (CodexProvider with api_key parameter) · deployment change (wheels available exclusively via GitHub Releases, not PyPI) · session state incompatibility (SessionState vs RunState format difference) · minimum Python version specified (3.10–3.14) · new execution model (Runner-based instead of harness-based)
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)
ACTION NEEDED APIDEPRECATION OpenAI Agents Python SDK releases · 2026-11-18 20:33 UTC

openai-agents-python 0.24.0 changes RunConfig.tool_error_formatter signature and RunState schema

Do:Update custom tool_error_formatter implementations to handle kind="tool_exception" or return None.
WHAT CHANGEDopenai-agents-python 0.24.0 changes RunConfig.tool_error_formatter to receive kind="tool_exception" and an optional error parameter. New RunState snapshots use schema 1.21. The Temporal extra now requires temporalio==1.34.0; the Vercel extra requires vercel-sandbox>=0.7.0,<0.9. New Decisions recipes require OpenAI SDK >=3.26.0,<4.
WHY IT MATTERSCustom error formatters that do not handle the new kind parameter will break. Saved agent graphs from earlier versions cannot resume on 0.24.0 without finishing on their original SDK first. Temporal and Vercel sandbox integrations will fail if dependency versions are outside the new bounds.
WHAT TO DOUpdate custom tool_error_formatter implementations to handle kind="tool_exception" or return None. Finish in-progress saved runs on their original SDK before upgrading, or use unique agent names for new durable runs. Update temporalio to 1.34.0 if using the Temporal extra, and vercel-sandbox to >=0.7.0,<0.9 if using the Vercel extra.
Model-reported confidence 90% (not measured accuracy) Primary source ↗ written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
 ## v0.24.0
 https://github.com/openai/openai-agents-python/releases/tag/v0.24.0
-2026-10-15T02:37:43Z
+2026-11-18T19:20:04Z
 
 0.24.0 (2026-10-15)
+Key Changes
+0.24.0 adds run-wide function error formatting and MCP discovery annotations, and improves recovery, sandbox safety, and tracing. This minor release includes compatibility changes for custom error callbacks and some saved agent graphs. Review the first two highlights before upgrading.
+Highlights:
+
+Update custom error formatters. RunConfig.tool_error_formatter can receive kind="tool_exception" and an optional error. Handle the new kind or return None to retain the default message. Explicit per-tool handlers keep precedence; unconfigured defaults stay unchanged.
+Plan upgrades around saved runs. New RunState snapshots use schema 1.21. Finish affected older same-name agent graphs on their original SDK. Keep the agent code and original hook ordering unchanged. Use unique agent names for new durable runs; renaming a saved graph is not a migration.
+Recover completed work more reliably. Deferred approval history, completed tool outputs, aliases, and compaction state survive supported recovery paths. Automatic compaction waits for approvals. Redis replaces history atomically on supported standalone clients. Conversations appends respect the 20-item request limit. Reconcile partial writes before retrying.
+Harden local execution and display. macOS sandbox commands deny Launch Services requests. Sandbox tools recover from invalid model arguments. Protected mount errors preserve caller exception data while detaching the caught traceback. The realtime demo displays tool content as text.
+Honor tracing choices through resume and cleanup. Per-run tracing opt-out applies inside caller traces, and approved resumed tools attach to the resumed agent span.
+Use richer integration examples and metadata. MCP tools expose optional mcp_annotations as discovery hints. New Decisions recipes demonstrate classification, routing, and tool review with OpenAI >=3.26.0,<4. They use a separate client, not an Agents model adapter.
+Refresh optional dependency locks. The Temporal extra uses temporalio==1.34.0; the Vercel extra supports vercel-sandbox>=0.7.0,<0.9. Core Python and OpenAI SDK minimums are unchanged.
+
+Thanks to @masasron for reporting the macOS local sandbox issue fixed in #5335.
 Features
 
 expose MCP annotations on converted tools (#5356) (75813f8)
@@
 docs/sandbox/clients.md, local and hosted clients
 Optionally note macOS Launch Services denial and document the Vercel >=0.7.0,<0.9 bound. Preserve the existing local-isolation limitations.
 
-Publication timing: publish documentation for new behavior after 0.24.0 is available. These documentation gaps do not block the release.
-
-Key Changes draft
-Key Changes
-0.24.0 adds run-wide function error formatting and MCP discovery annotations, and improves recovery, sandbox safety, and tracing. This minor release includes compatibility changes for custom error callbacks and some saved agent graphs. Review the first two highlights before upgrading.
-Highlights:
-
-Update custom error formatters. RunConfig.tool_error_formatter can receive kind="tool_exception" and an optional error. Handle the new kind or return None to retain the default message. Explicit per-tool handlers keep precedence; unconfigured defaults stay unchanged.
-Plan upgrades around saved runs. New RunState snapshots use schema 1.21. Finish affected older same-name agent graphs on their original SDK. Keep the agent code and original hook ordering unchanged. Use unique agent names for new durable runs; renaming a saved graph is not a migration.
-Recover completed work more reliably. Deferred approval history, completed tool outputs, aliases, and compaction state survive supported recovery paths. Automatic compaction waits for approvals. Redis replaces history atomically on supported standalone clients. Conversations appends respect the 20-item request limit. Reconcile partial writes before retrying.
-Harden local execution and display. macOS sandbox commands deny Launch Services requests. Sandbox tools recover from invalid model arguments. Protected mount errors preserve caller exception data while detaching the caught traceback. The realtime demo displays tool content as text.
-Honor tracing choices through resume and cleanup. Per-run tracing opt-out applies inside caller traces, and approved resumed tools attach to the resumed agent span.
-Use richer integration examples and metadata. MCP tools expose optional mcp_annotations as discovery hints. New Decisions recipes demonstrate classification, routing, and tool review with OpenAI >=3.26.0,<4. They use a separate client, not an Agents model adapter.
-Refresh optional dependency locks. The Temporal extra uses temporalio==1.34.0; the Vercel extra supports vercel-sandbox>=0.7.0,<0.9. Core Python and OpenAI SDK minimums are unchanged.
-
-Thanks to @masasron for reporting the macOS local sandbox issue fixed in #5335.
-Notes
-
-This review covers the exact candidate above, including its release metadata. Changes to main, the candidate, or its API baseline require a fresh release check and approval.
+Publication timing: publish documentation for new behavior after 0.24.0 is available. These documentation gaps do not block the release.
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
triage verdictMATERIAL at 0.85, needed 0.5
triage reasoningCustom error callback signature changed; saved runs incompatible with new schema; dependency and SDK version constraints introduced.
signalsbreaking API change: RunConfig.tool_error_formatter signature changed · schema migration: RunState snapshots now use schema 1.21 · saved state incompatibility: older agent graphs require finishing on original SDK · dependency version bump: temporalio==1.34.0, vercel-sandbox>=0.7.0,<0.9 · behavior change: automatic compaction now waits for approvals · behavior change: Redis history replacement atomicity, conversation append limits · new required OpenAI SDK floor: >=3.26.0,<4 for Decisions recipes
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)
ACTION NEEDED DEPRECATIONAPI OpenAI Agents Python SDK releases · 2026-11-18 07:29 UTC

OpenAI Agents Python SDK 0.25.0 removes Runloop sandbox provider

Do:If using Runloop: preserve workspace data, install openai-agents[modal], replace SandboxRunConfig client and options, adapt mounts, and restore data into new sandboxes b…
WHAT CHANGEDVersion 0.25.0 removes the Runloop sandbox provider integration, including the runloop extra, agents.extensions.sandbox.runloop module, and all Runloop exports. The checkpoint schema version increments from 1.21 to 1.22 to track approvals owned by nested agent tools. Pydantic context snapshots now include default values and in-place edits using JSON serialization, which can require an explicit serializer for process-local fields. OpenAIResponsesCompactionSession default model changes from gpt-4.1 to gpt-6-luna. MCP dependency requirement changes to >=1.30.0,<3,!=2.0.*,!=2.1.*. Daytona extra requires daytona>=0.226.1,<1.
WHY IT MATTERSApplications using Runloop must switch to another sandbox provider (Modal is documented) or remain on 0.24.0; existing Runloop session IDs and snapshots cannot be resumed directly by other providers. A v0.24.0 worker rejects a 1.22 checkpoint, so deploying writers before readers breaks checkpoint resume. Pydantic contexts with process-local defaults (client objects) that previously serialized by omitting unset fields now fail without an explicit serializer. Unset compaction model configurations silently switch to a different model.
WHAT TO DOIf using Runloop: preserve workspace data, install openai-agents[modal], replace SandboxRunConfig client and options, adapt mounts, and restore data into new sandboxes before upgrading. If sharing checkpoints across workers: upgrade resume workers to 0.25.0 before enabling writers that produce 1.22 checkpoints, or route by SDK version. If using Pydantic context with process-local fields: add context_serializer to exclude or transform them and context_deserializer or context_override to restore application context. If relying on default compaction model: set model="gpt-4.1" on OpenAIResponsesCompactionSession to retain previous behavior. Refresh dependency locks for MCP and Daytona version changes.
Model-reported confidence 95% (not measured accuracy) Primary source ↗ written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
+## v0.25.0
+https://github.com/openai/openai-agents-python/releases/tag/v0.25.0
+2026-11-18T03:37:05Z
+
+0.25.0 (2026-11-17)
+Features
+
+expose utilities for third-party session adapters (#5428) (16fdb20)
+sandbox: remove Runloop provider (#5369) (12ce34d)
+
+Bug Fixes
+
+avoid loading workspace startup files during internal sandbox operations (#5406) (841dfdd)
+preserve nested ToolContext approval decisions across resume (#5633) (c200ae8)
+avoid repeatedly fingerprinting completed tool calls (#5421) (dda19f9)
+add opt-in handoff error recovery (#5418) (a6ed84f)
+align implicit settings with provider-selected models (#5425) (a88d291)
+align model defaults with TS SDK (#6064) (d5287b7)
+allow concurrent Streamable HTTP requests on MCP v2 (#5641) (2b90566)
+default GPT-6 Chat Completions tool requests to non-reasoning (#5420) (3396dca)
+deps-dev: bump aiohttp from 3.14.3 to 3.14.5 (#5403) (442d0ea)
+deps-dev: bump dapr from 1.18.3 to 1.19.0 (#5711) (25e7b2e)
+deps-dev: bump fastapi from 0.141.1 to 0.162.2 (#6053) (494c0c0)
+deps-dev: bump playwright from 1.50.0 to 1.63.0 (#5402) (0dc4cf1)
+deps-dev: bump pytest from 9.0.3 to 9.2.0 (#5710) (b0051a7)
+deps-dev: bump rich from 14.3.3 to 15.0.0 (#6055) (18d979b)
+deps-dev: bump ruff from 0.16.8 to 0.17.1 (#5439) (5bb6fb7)
+deps-dev: bump twilio from 9.11.1 to 9.12.0 (#5486) (465583d)
+deps-dev: bump websockets from 15.0.1 to 16.1.1 (#5484) (86bd298)
+deps: allow websockets 17 (#6044) (9d0f5e8)
+deps: bump actions/create-github-app-token from 3.2.0 to 3.3.0 (#5464) (b8df4ff)
+deps: bump actions/download-artifact from 8.0.1 to 8.0.2 (#5466) (eccd8c8)
+deps: bump actions/setup-python from 7.0.0 to 7.1.0 (#5465) (18d233b)
+deps: bump astral-sh/setup-uv from 10.2.0 to 10.3.2 (#5463) (1ed8268)
+deps: bump certifi from 2026.10.19 to 2026.10.23 (#5709) (976f2f1)
+deps: bump daytona from 0.155.0 to 0.226.1 (#5487) (d4f07df)
+deps: bump httpx2 from 2.12.0 to 2.14.0 (#5438) (0c55458)
+deps: bump mcp from 2.0.0 to 2.2.0 (#5393) (97c8288)
+deps: bump mcp from 2.0.0 to 2.3.0 (#5404) (e2de6a8)
+deps: bump openai from 3.29.0 to 3.48.0 (#6054) (a4a4fa4)
+deps: bump opentelemetry-api from 1.40.0 to 1.45.1 (#5410) (2964f77)
+deps: bump opentelemetry-api from 1.45.1 to 1.45.2 (#5983) (cdc8d58)
+deps: bump pydantic from 2.14.2 to 2.14.3 (#6052) (4bb2637)
+deps: bump starlette from 1.6.0 to 1.7.0 (#5708) (82483dd)
+deps: bump the codeql group with 2 updates (#5462) (775707e)
+deps: require certifi 2026.10.19 or later (#5429) (6295010)
+deps: require pymongo 4.18.3 or later (#5445) (57027de)
+deps: update fastapi requirement from >=0.142.2 to >=0.143.0 in /examples/realtime/twilio_sip (#5399) (6bfa5ff)
+deps: update fastapi requirement from >=0.143.0 to >=0.150.0 in /examples/realtime/twilio_sip (#5435) (56df6b4)
+deps: update fastapi requirement from >=0.150.0 to >=0.162.0 in /examples/realtime/twilio_sip (#5482) (803fbf2)
+deps: update fastapi requirement from >=0.162.0 to >=0.162.2 in /examples/realtime/twilio_sip (#6051) (9a3bce3)
+deps: update httpx2 requirement from >=2.13.1 to >=2.15.1 in /examples/realtime/twilio (#5432) (4134643)
+deps: update httpx2 requirement from >=2.13.1 to >=2.15.1 in /examples/realtime/twilio_sip (#5436) (a91f0bf)
+deps: update httpx2 requirement from >=2.15.1 to >=2.16.0 in /examples/realtime/twilio (#5478) (16c3a3f)
+deps: update httpx2 requirement from >=2.15.1 to >=2.16.0 in /examples/realtime/twilio_sip (#5483) (39a123b)
+deps: update openai requirement from <4,>=3.26.0 to >=3.29.0,<4 in /examples/realtime/twilio_sip (#5398) (0e54e4b)
+deps: update openai requirement from <4,>=3.29.0 to >=3.35.1,<4 in /examples/realtime/twilio_sip (#5434) (2350082)
+deps: update openai requirement from <4,>=3.35.1 to >=3.39.0,<4 in /examples/realtime/twilio_sip (#5480) (13dbaa3)
+deps: update openai requirement from <4,>=3.39.0 to >=3.42.1,<4 in /examples/realtime/twilio_sip (#5707) (6e590a0)
+deps: update openai requirement from <4,>=3.42.1 to >=3.48.0,<4 in /examples/realtime/twilio_sip (#6050) (5e0f944)
+deps: update python-multipart requirement from >=0.0.32 to >=0.0.33 in /examples/realtime/twilio (#5706) (64fbde5)
+deps: update starlette requirement from >=1.7.0 to >=1.9.0 in /examples/realtime/twilio (#5433) (26c762a)
+deps: update starlette requirement from >=1.7.0 to >=1.9.0 in /examples/realtime/twilio_sip (#5437) (b6966fa)
+deps: update starlette requirement from >=1.9.0 to >=1.10.0 in /examples/realtime/twilio (#5479) (f137418)
+deps: update starlette requirement from >=1.9.0 to >=1.10.0 in /examples/realtime/twilio_sip (#5481) (814de8b)
+deps: update twilio requirement from <10,>=9.11.2 to >=9.12.0,<10 in /examples/realtime/twilio (#5397) (452c5bd)
+deps: update twilio requirement from <10,>=9.12.0 to >=9.13.0,<10 in /examples/realtime/twilio (#6049) (e27451c)
+flatten restored handoff history summaries (#6041) (7635b81)
+improve missing stream completion diagnostics (#5457) (e355d69)
+limit the Daytona extra to the supported 0.x SDK (#6065) (fd78fa0)
+preserve append order in SQLAlchemy sessions (#5645) (89a3ea4)
+preserve empty structured shell results (#5423) (963d032)
+Preserve nested handoff record boundaries (b040da5)
+preserve pending async calls in committed sessions (#6039) (f48003a)
+preserve pending input history across streamed retries (#6042) (9e05bd5)
+preserve Pydantic run context snapshots (#5444) (bc4c6e1)
+preserve structured tool input in agent lifecycle hooks (#6037) (e3538bd)
+preserve usage from failed and incomplete Responses calls (#6046) (6e6707c)
+preserve wire aliases when reading Conversations session history (#6047) (5789c7c)
+preserve wire aliases when replaying Responses items (#5638) (71bab8c)
+raise MCP minimum to 1.21.2 for Pydantic and OAuth compatibility (#5384) (970db99)
+realtime: honor explicit interruption with automatic VAD (#5427) (addc737)
+remove retry advice from the default tool error message (#5632) (fa7f312)
+remove unused EFS installer from sandbox mount example (#6038) (0777ee6)
+replay ID-less assistant output in Chat Completions sessions (#5419) (28c29bc)
+report completed input guardrails when a tripwire fires (#5634) (2cc15af)
+report completed output guardrails when a tripwire fires (#6040) (4bab6ea)
+report missing function approval owners during RunState restore (#5630) (788fb99)
+require patched MCP dependency versions (#5431) (b235388)
+Responses tool choice for functions named after hosted tools (#5639) (3ac005f)
+Restore checkpointed assistant text in Chat Completions (8ffdf21)
+resume nested approval checkpoints with output guardrails (#5629) (3599632)
+retain tool guardrail results after approval and resume (#5643) (3a0c89c)
+sessions: bounded session replay and isolate the Cloudflare mount test (#5637) (4920e5b)
+skip streamed model attempts after an input guardrail trips (#5644) (8ea4333)
+stop native tool queues during function failure cleanup (#5446) (cd26bbb)
+tool approval handling in the demo REPL (#6045) (d4d835e)
+trace offered tools on Chat Completions generation spans (#5635) (bb36fae)
+use gpt-6-luna for session compaction by default (#5646) (1fc77bb)
+
+Release readiness review (v0.24.0 -> TARGET 4cd46a6)
+This is a release readiness report done by $final-release-review skill.
+Diff
+v0.24.0...4cd46a6
+Release intent
+
+Review mode: final candidate
+Intended release: minor, 0.25.0
+Minimum required release type: minor. Existing Pydantic checkpoint callers can need a serializer where v0.24.0 omitted an unset field, and previously supported MCP dependency pins can require an upgrade.
+Versioning verdict: compatible
+
+Release call
+🟢 GREEN LIGHT TO SHIP. The candidate provides migration paths for the compatibility changes. The package metadata, editable lock entry, Release Please manifest, source version and frozen API baseline agree on 0.25.0.
+Scope summary
+
+261 files changed (+14,052/-17,143) across the SDK, tests, examples, documentation and release workflows. Relative to current main at fd78fa038f853598ed849e25eec78bab1e49aa00, #5394 changes only six release metadata files. The frozen API snapshot records its generating commit e03c7157f7629ff60aa3e434f2f083f72e4c7ba9, the candidate's parent.
+
+Risk assessment (ordered by impact)
+
+Runloop sandbox migration
+
+Risk: 🟢 LOW with migration. Applications using the beta Runloop integration must switch providers before upgrading or stay on 0.24.0 during migration.
+Evidence: The candidate removes the runloop extra, agents.extensions.sandbox.runloop, and the Runloop exports. The migration guide identifies the supported Modal client and explains how provider-specific state affects the move.
+Files: pyproject.toml, src/agents/extensions/sandbox/__init__.py, examples/sandbox/extensions/README.md
+Action: Preserve needed workspace data first. For Modal, install openai-agents[modal], replace the client and options in SandboxRunConfig, adapt mounts, and restore data into a new sandbox. Existing Runloop session IDs and snapshots cannot be resumed directly by Modal.
+
+Checkpoint readers and Pydantic context serialization
+
+Risk: 🟢 LOW with ordered deployment. A v0.24.0 worker rejects a new 1.22 checkpoint. A Pydantic context with a process-local default, such as a client object, can now require an explicit serializer.
+Evidence: v0.25.0 still reads 1.21 and writes 1.22 to track approvals owned by nested agent tools. Pydantic context snapshots now include default values and in-place edits using JSON serialization.
+Files: src/agents/run_state.py, tests/test_run_state_pydantic_context.py
+Action: Upgrade resume workers before enabling new writers, or route checkpoints by SDK version. Use context_serializer to exclude or transform process-local fields and context_deserializer or context_override to restore the application context.
+
+Model defaults and dependency pins
+
+Risk: 🟢 LOW with explicit configuration. Applications that leave the compaction model unset will use a different model. Dependency locks may also need updates.
+Evidence: OpenAIResponsesCompactionSession changes its default from gpt-4.1 to gpt-6-luna. MCP now requires >=1.30.0,<3,!=2.0.*,!=2.1.*. The Daytona extra requires daytona>=0.226.1,<1.
+Files: src/agents/memory/openai_responses_compaction_session.py, src/agents/models/default_models.py, pyproject.toml
+Action: Set model="gpt-4.1" on the compaction session to retain its previous model. Refresh incompatible dependency pins. With MCP v2 Streamable HTTP, enable tool retries only when replaying a call is safe.
+
+Documentation coverage (non-blocking)
+
+Coverage source: Reviewed the current open PRs. No separate open PR covers the 0.25.0 user-guide updates below.
+Status: partially covered.
+Covered obligations: The sandbox extension examples document the Runloop-to-Modal migration and data preservation. The human-in-the-loop guide already describes context serializers and version-aware resume routing.
+Post-release updates:
+
+In docs/human_in_the_loop.md, add the 1.22 reader-before-writer rollout order and a Pydantic example that excludes a process-local field.
+In docs/sessions/index.md and docs/models/index.md, distinguish the new compaction default from the agent model default, show the model= override, and update GPT-5-and-newer reasoning model guidance.
+In docs/mcp.md, correct the dependency range and explain when retrying concurrent Streamable HTTP tool requests is safe. In docs/handoffs.md, document the Runner-only failure_error_function and its default exception behavior.
+In docs/sandbox/clients.md, link the Runloop migration, record the Daytona range and refresh the localized provider tables. In examples/sandbox/external_provider/README.md, qualify the local co-install command: the standalone example declares >=0.24.0,<0.25.0 and cannot be installed alongside this candidate with that constraint.
+
+Publication timing: Publish the 0.25.0 guidance after the package is available.
+
+Key Changes draft
+Key Changes
+Version 0.25.0 removes the beta Runloop sandbox provider and changes how Pydantic run contexts are checkpointed. Runloop users should preserve their workspace data before switching providers. Applications that share checkpoints across workers should upgrade readers first.
+Highlights:
+
+Migrate Runloop workloads to another supported sandbox provider, such as Modal, or remain on 0.24.0 while migrating. Provider-specific sessions and snapshots require a separate data migration.
+Resume nested agent-tool approvals with output guardrails using RunState schema 1.22. Pydantic checkpoints now preserve default values and in-place edits. Use a custom serializer for process-local resources.
+OpenAIResponsesCompactionSession now defaults to gpt-6-luna. Set model="gpt-4.1" to keep the previous choice. Agent model settings also follow the selected model more closely. Native Chat Completions tool requests for gpt-6-sol and gpt-6-luna default to non-reasoning mode when no effort is supplied.
+Runner handoffs can opt into recovery with failure_error_function. Omitting it preserves exception propagation. The demo REPL can now approve or reject individual tool calls.
+MCP v2 Streamable HTTP supports concurrent requests. MCP v1 retains serialized requests. Failed and incomplete Responses calls now retain provider-reported token usage.
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
triage verdictMATERIAL at 0.92, needed 0.5
triage reasoningMultiple breaking changes affecting deployment, checkpointing, model selection, and sandbox provider support require migration planning.
signalsbreaking change: Runloop sandbox provider removed · breaking change: checkpoint schema version bump (1.21 → 1.22) · breaking change: Pydantic context serialization behavior changed · model default changed: gpt-4.1 → gpt-6-luna for compaction · dependency requirement changed: MCP >=1.30.0,<3,!=2.0.*,!=2.1.* · dependency requirement changed: Daytona >=0.226.1,<1 · migration required: Runloop users must switch providers or stay on 0.24.0 · deployment ordering required: checkpoint readers must upgrade before writers
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)