Tool approval is not honored on resume when `Runner.run` is given a context by ojassharma7 · Pull Request #4245 · openai/openai-agents-python

ojassharma7 · 2026-08-06T04:28:01Z

What changed

src/agents/run_internal/agent_runner_helpers.py
tests/test_run_state.py

Verification

The project's own test suite was run before and after this change; it introduces no new test failures or lint violations.

…ontext Closes openai#4244

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bfa6fb8b80

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

seratch

Thanks for the contribution. The underlying issue is valid, but please revise the context-resolution change so the restored RunContextWrapper remains authoritative and only its application context value is replaced. Copying selected private fields into a new wrapper omits persisted metadata such as tool_input and will drift as additional run-owned state is added.

Please also add a regression covering the issue's serialized to_json() / from_json() resume path, plus the equivalent Runner.run_streamed() path. Both should prove that the approved call executes, the supplied application context is visible, and SDK-managed wrapper state survives the override.

Also, getting 👍 from Codex's PR review is required before merging in this repo, so please continue resolving all the review comments by Codex here (you don't need to mention codex; Codex automatically reviews all commits for this repo).

Replace only the application context value on the restored wrapper so approvals, usage, turn_input, and tool_input survive context= on resume. Add JSON and streamed regression coverage for issue openai#4244. Co-authored-by: Cursor <cursoragent@cursor.com>

ojassharma7 · 2026-08-06T15:15:04Z

Thanks @seratch — revised in 6a0c0ae:

resolve_resumed_context now keeps the restored RunContextWrapper and only replaces its application .context value (no field-by-field copy onto a new wrapper).
Added regressions for the to_json() / from_json() resume path and Runner.run_streamed(), asserting the approved tool runs, the override context is visible, and wrapper state (tool_input) survives.

Add a unit test that proves resolve_resumed_context keeps the restored wrapper identity and run-owned state, and share HITL setup across the JSON and streamed regressions with approvals-identity assertions. Co-authored-by: Cursor <cursoragent@cursor.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d05ddd9826

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

When Runner.run resumes with context=, also update cached Agent.as_tool() RunState application contexts for the resume scope so nested tools see the override while keeping nested approvals and tool_input intact. Co-authored-by: Cursor <cursoragent@cursor.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9c042f22d7

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

Walk only the resumed RunState's pending function tool calls instead of every unscoped agent-tool cache entry, so concurrent live interruptions cannot leak a context= override into each other. Co-authored-by: Cursor <cursoragent@cursor.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d0ce12a92a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

Recurse through live Agent.as_tool() wrappers so A→B→C overrides reach C, only consider _is_agent_tool function runs, and avoid unscoped signature fallback so ordinary functions cannot overwrite unrelated nested context. Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

seratch

Thanks for continuing to work through the feedback. The underlying issue is valid, and keeping the restored RunContextWrapper authoritative is the correct direction. However, the current head has expanded into recursive traversal and mutation of the shared agent-tool cache.

Please simplify this to one context-resolution path: keep the restored wrapper and replace only its application .context; when a pending Agent.as_tool() run is actually resumed, pass the current parent application context into its nested Runner.run() or Runner.run_streamed() call so the same resolver applies there. Please remove apply_application_context_to_agent_tool_states and the synthetic cache-propagation tests.

Keep the JSON round-trip and streaming regressions, and add an end-to-end nested Agent.as_tool() regression proving that the new application context is visible while nested approvals and tool_input remain owned by the restored wrapper.

Keep a single resolve_resumed_context path that only replaces the restored wrapper's application context, and pass the parent application context into nested Runner.run/run_streamed on Agent.as_tool() resume instead of mutating the shared agent-tool cache. Co-authored-by: Cursor <cursoragent@cursor.com>

ojassharma7 · 2026-08-06T23:58:39Z

Thanks @seratch — simplified as requested in the latest commits.

resolve_resumed_context only keeps the restored wrapper and replaces its application .context
Removed apply_application_context_to_agent_tool_states and the synthetic cache-propagation tests
On pending Agent.as_tool() resume, nested Runner.run / Runner.run_streamed now receive the current parent application context so the same resolver applies
Kept the JSON + streamed regressions, and added an end-to-end nested Agent.as_tool() regression for context override + wrapper-owned approvals/tool_input

seratch

Thanks for the update. The latest head now has the right ownership model: it keeps the restored RunContextWrapper authoritative, replaces only the application context, and routes nested Agent.as_tool() resumes through the same resolver. The JSON round-trip, streaming, and nested regressions cover the required behavior, and CI is green.

One small repository-style fix remains before re-review: in the resolve_resumed_context docstring, replace the Unicode ellipsis after tool input, with three ASCII dots (tool input, ...) as required by AGENTS.md. After that change, this looks ready for final review.

Co-authored-by: Cursor <cursoragent@cursor.com>

ojassharma7 · 2026-08-07T00:32:03Z

Thanks, updated the resolve_resumed_context docstring to use ASCII ... instead of the Unicode ellipsis.

seratch · 2026-08-07T00:42:47Z

@ojassharma7 oops, can you resolve the conflicts?

Co-authored-by: Cursor <cursoragent@cursor.com>

ojassharma7 · 2026-08-07T00:46:17Z

Resolved, merged latest main and fixed the import conflict in tests/test_run_state.py (kept both the new error-handler imports and resolve_resumed_context).

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b95b07708a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

seratch

hmm, codex pointed a new issue out. Can you resolve it as well?

When resuming an approved nested agent-tool interruption, keep the restored wrapper's approvals/tool_input but point its Usage back at the parent ToolContext accumulator so post-resume nested turns are not undercounted. Co-authored-by: Cursor <cursoragent@cursor.com>

ojassharma7 · 2026-08-07T00:59:05Z

@seratch addressed the Codex usage note as well — nested Agent.as_tool() resume now reattaches the restored nested wrapper's usage to the parent ToolContext accumulator so post-resume nested turns are counted on the parent.

Tool approval is not honored on resume when Runner.run is given a c…

bfa6fb8

…ontext Closes openai#4244

chatgpt-codex-connector Bot reviewed Aug 6, 2026

View reviewed changes

Comment thread src/agents/run_internal/agent_runner_helpers.py Outdated

seratch added the feature:core label Aug 6, 2026

seratch added this to the 0.20.x milestone Aug 6, 2026

seratch requested changes Aug 6, 2026

View reviewed changes

ojassharma7 and others added 2 commits August 6, 2026 11:13

Merge remote-tracking branch 'origin/main' into autocontrib/issue-4244

2633f42

chatgpt-codex-connector Bot reviewed Aug 6, 2026

View reviewed changes

Comment thread src/agents/run_internal/agent_runner_helpers.py

chatgpt-codex-connector Bot reviewed Aug 6, 2026

View reviewed changes

Comment thread src/agents/agent_tool_state.py Outdated

chatgpt-codex-connector Bot reviewed Aug 6, 2026

View reviewed changes

Comment thread src/agents/agent_tool_state.py Outdated

Comment thread src/agents/agent_tool_state.py Outdated

ojassharma7 and others added 2 commits August 6, 2026 12:09

Fix lint formatting and typecheck errors on resume context path

ec38d00

Co-authored-by: Cursor <cursoragent@cursor.com>

seratch requested changes Aug 6, 2026

View reviewed changes

ojassharma7 and others added 2 commits August 6, 2026 19:57

Merge upstream/main into autocontrib/issue-4244

9f6f234

seratch requested changes Aug 7, 2026

View reviewed changes

Use ASCII ellipsis in resolve_resumed_context docstring

5bd3073

Co-authored-by: Cursor <cursoragent@cursor.com>

seratch approved these changes Aug 7, 2026

View reviewed changes

Merge upstream/main and resolve test_run_state import conflict

b95b077

Co-authored-by: Cursor <cursoragent@cursor.com>

chatgpt-codex-connector Bot reviewed Aug 7, 2026

View reviewed changes

Comment thread src/agents/agent.py

seratch requested changes Aug 7, 2026

View reviewed changes

seratch approved these changes Aug 7, 2026

View reviewed changes

seratch merged commit 00c9d26 into openai:main Aug 7, 2026
10 checks passed

Uh oh!

Conversation

ojassharma7 commented Aug 6, 2026

What changed

Verification

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

seratch left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ojassharma7 commented Aug 6, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

seratch left a comment

Choose a reason for hiding this comment

Uh oh!

ojassharma7 commented Aug 6, 2026

Uh oh!

seratch left a comment

Choose a reason for hiding this comment

Uh oh!

ojassharma7 commented Aug 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

seratch commented Aug 7, 2026

Uh oh!

ojassharma7 commented Aug 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

seratch left a comment

Choose a reason for hiding this comment

Uh oh!

ojassharma7 commented Aug 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

seratch left a comment •

edited

Loading

ojassharma7 commented Aug 7, 2026 •

edited

Loading

ojassharma7 commented Aug 7, 2026 •

edited

Loading