fix(core): avoid redundant final computer screenshots

[kyletser]

Summary

• Treat screenshot actions as no-ops during computer-action dispatch and use the existing post-batch capture for the observation.
• Remove the branch-local screenshot cache and invalidation logic, as requested in review.
• Cover a screenshot request before a click, a trailing screenshot request, direct execution, HITL resume, and legacy replay. The screenshot-before-click regression checks that the returned image reflects the completed click.

Test plan

• Direct execution, HITL and replay: 337 passed.
• Runner tests: 692 passed.
• Full Linux suite: 6453 passed, 2 skipped (209 test files passed).
• Frozen dependency installation, build, recursive build-check, package dist:check, lint, Prettier for all changed files, and git diff --check passed.

Issue number

Fixes #1735

Checks

• I've added new tests.
• I've added a patch changeset for @openai/agents-core.
• No public API or documentation changes are needed.

[changeset-bot]

🦋 Changeset detected

Latest commit: 1b75d67

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 6 packages

Name	Type
@openai/agents-core	Patch
@openai/agents-extensions	Patch
@openai/agents-openai	Patch
@openai/agents-realtime	Patch
@openai/agents	Patch
realtime-react-native	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-08T13:15:20.811101Z	1b75d67	Manual request
🔒 Security Review	✅ Completed 2026-09-08T13:17:33.052522Z	1b75d67	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to 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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Copilot]

🟢 Approval recommended

The change is narrowly scoped to removing a redundant screenshot call, and the updated/added tests cover the reported regression scenarios and verify the single-capture behavior.

Pull request overview

Fixes duplicate computer screenshot captures in the agents-core runner by deferring explicit screenshot actions to the existing single “final post-batch” capture, so each computer tool call produces exactly one screenshot observation.

Changes:

• Treat computer_call screenshot actions as no-ops during action dispatch and rely on the single final capture.
• Update/extend test coverage for direct execution, HITL resume, and legacy replay to assert a single screenshot capture.
• Add a patch changeset for @openai/agents-core.

File summaries

File	Description
packages/agents-core/src/runner/toolExecution.ts	Stops invoking computer.screenshot() for action.type === 'screenshot', leaving only the final capture after all actions.
packages/agents-core/test/runner/toolExecution.test.ts	Adds/updates assertions to ensure exactly one screenshot per computer call and validates ordering for non-screenshot actions.
packages/agents-core/test/runner/turnResolution.test.ts	Adjusts resume-related expectations to match single-capture behavior.
packages/agents-core/test/toolInvocationReplay.test.ts	Updates replay expectations to reflect removing the redundant screenshot call.
.changeset/single-final-computer-screenshot.md	Declares a patch bump for the bug fix in @openai/agents-core.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[sylvesterkaczmarek]

Deferring every explicit screenshot changes the semantics of the action sequence, rather than only removing a duplicate capture. For [screenshot, click], the caller asked the backend to capture before the click, but this patch deliberately executes click first and captures only afterward. A screenshot can also be an observable or synchronizing operation for remote computer implementations. A safer shape would be to execute explicit screenshots in order, cache the most recent capture, and reuse it only when no later non-screenshot action invalidates it. That still gives one capture for [screenshot] and [click, screenshot], while preserving the requested ordering for [screenshot, click].

[kyletser]

Addressed in 1ee899c. Explicit screenshot actions now execute at their original sequence position. The most recent capture is reused only when no later non-screenshot action invalidates it; otherwise the runner captures the final state after the remaining actions. Added regressions for both [screenshot, click]\ and [click, screenshot]. The focused execution, HITL resume, and replay suite passes (337 tests), along with build, type, dist, lint, and formatting checks. The Windows full-suite run reached 6,363 passes; its 92 failures are confined to existing Unix sandbox, POSIX path/permission, and Windows child-process fixtures, while all changed-path tests pass.

[sylvesterkaczmarek]

Re-reviewed the current action loop. Explicit screenshots now run in sequence, a later non-screenshot action invalidates the cached capture, and only a final explicit screenshot is reused. The [screenshot, click] and [click, screenshot] regressions cover the ordering boundary I raised. My concern is resolved.

[seratch]

Thanks for the update. The current revision restores the caching and invalidation approach that we previously decided against in #1736. The official computer-use handlers treat screenshot actions as no-ops within the batch and capture the final screen afterward.

Please remove the screenshot cache and invalidation logic, make the screenshot switch case a no-op, and use the existing final capture. Update the [screenshot, click] regression to expect the click followed by one screenshot, with that image returned as the observation. This addresses #1735 without adding an intermediate-capture contract.

[kyletser]

Thanks for clarifying. Addressed in 1b75d67: the screenshot case is now a no-op, and the cache/invalidation logic is removed. The existing final capture supplies the observation. The [screenshot, click] regression now expects click followed by exactly one screenshot and checks that the returned image reflects the completed click. Focused execution/HITL/replay: 337 passed; runner: 692 passed; full Linux suite: 6453 passed, 2 skipped. Build, type checks, lint, and formatting also passed.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1b75d67cc7

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seratch]

Please fix all the codex feedback

[kyletser]

@seratch Could you confirm the intended screenshot semantics before I change the implementation again?

Your earlier review requested removing the cache/invalidation logic and treating screenshot actions as no-ops, with one final capture. That is implemented in 1b75d67. The new automated finding requests the opposite: capture at each screenshot action's position, cache the image, and invalidate it after subsequent actions, citing the guide's statement that batched actions execute in order.

For [screenshot, click], should the expected calls remain click -> screenshot, or should they be screenshot -> click -> screenshot? If final-only capture is still intended, I can clarify the screenshot exception in the appropriate documentation, coordinated with release timing, rather than restore the approach rejected in #1736. I have left the finding open pending clarification.

[seratch]

Thanks for flagging the conflicting feedback. The intended behavior remains one final capture per completed batch: [screenshot, click] should execute click followed by screenshot.

My later request to address all Codex feedback was too broad. The intermediate-capture finding does not change the specific direction previously requested. The current implementation matches that direction; no cache or invalidation logic is needed.