fix(computer): isolate provider instances across concurrent runs

[russeell]

Summary

This pull request fixes ComputerTool provider isolation when the same agent is reused by concurrent runs.

Per-run computer providers were introduced in #2191 so applications can reuse an agent while creating and disposing independent computer resources for each request. The existing initialization cache is run-scoped, but model serialization still observes the mutable computer field on the shared tool.

initialize_computer_tools() now returns run-local ComputerTool copies for provider/factory-backed tools, bound to the computer resolved for the current RunContextWrapper. The original tool continues to own the existing per-run cache and disposal lifecycle, while model serialization and action processing receive stable run-local state instead of reading a shared mutable field. Tools configured with an existing concrete Computer retain their original identity for hooks and custom-data callbacks.

This prevents preview-compatible Responses requests from serializing another run's computer dimensions or failing when a sibling run restores the shared tool to its provider during cleanup. The change does not alter the public API, the wire payload, or the GA computer tool behavior.

Test plan

• Added a controlled concurrent Runner.run() regression test that interleaves two runs sharing one agent and provider-backed ComputerTool.
• Verified that both runs serialize their own display width, both complete successfully, and each created computer is disposed exactly once.
• Extended the existing non-streamed and streamed lifecycle tests to verify that the model receives a run-local tool bound to the correct computer.
• Added a full runner tool-call regression test confirming that concrete ComputerTool instances preserve their identity in model input and run hooks.
• Ran the focused lifecycle and affected runner tests.
• Ran codex review --uncommitted with no actionable findings.
• Ran the repository verification stack with the local provider override removed and localhost excluded from the system proxy:

env -u OPENAI_BASE_URL NO_PROXY=localhost,127.0.0.1 no_proxy=localhost,127.0.0.1 \
  bash .agents/skills/code-change-verification/scripts/run.sh

All formatting, lint, type-checking, and test steps passed.

Issue number

Closes #3842

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8113663664

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".

[chatgpt-codex-connector]

Preserve concrete ComputerTool identity for hooks

When a ComputerTool already wraps a concrete Computer, this still replaces it with a fresh ComputerTool even though there is no provider state to isolate. That copy is later stored in ToolRunComputerAction, so ComputerAction.execute passes the copy to on_tool_start/on_tool_end and custom-data callbacks; hooks that key off the configured agent.tools[0] object stop recognizing their tool on ordinary concrete-computer runs. Please avoid cloning already-resolved computer tools, or otherwise preserve the original tool identity for hook-facing surfaces.

Useful? React with 👍 / 👎.

[russeell]

Thanks, good catch. You're right that cloning a concrete ComputerTool unnecessarily changes its hook-facing identity. I'll update the implementation so only provider/factory-backed tools receive run-local copies, and add a regression test confirming concrete tools preserve their original identity.