fix: isolate usage between RunState checkpoints

[chiruu12]

Summary

RunContextWrapper._copy_for_run_state detaches _approvals, _tool_invocations and _restored_unbound_approval_call_ids, but the copy.copy(self) it starts from leaves usage shared. Usage.add accrues in place and extends request_usage_entries, so every checkpoint taken from a result, and the result itself, accumulate into one instance.

Two consequences, both reproducible on main:

A RunResult the caller already holds keeps changing. After result.to_state() and a resume, result.context_wrapper.usage.requests goes from 1 to 2.

Two checkpoints from one result bill each other. Each resumed once from the same one-request run, so both should report 2 requests. The second reports 3.

The copy is deep because add extends request_usage_entries rather than replacing it.

turn_input is aliased by the same copy.copy, but it is reassigned rather than mutated in turn_resolution.py and run_loop.py, so I left it as is rather than widen the change.

Test plan

Two tests in TestRunStateResumption, next to the existing resume-isolation coverage: one asserts a resume does not move the source result's usage, the other that two checkpoints from one result report the same total and do not share request_usage_entries. Reverting only the run_context.py line fails them with assert 2 == 1 and assert 3 == 2.

make format, make lint, make typecheck and make tests all pass (8806 passed, plus the serial pass).

Issue number

None, filing the fix directly since it is a narrow follow-up to #4413.

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

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 41cb6a518a

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]

Thanks for the contribution. Please keep the new copied.usage = copy.deepcopy(self.usage) assignment in src/agents/run_context.py; that correctly isolates top-level checkpoints.

The additional runtime change belongs in Agent.as_tool()'s _run_agent_impl() in src/agents/agent.py. After the pending_resume_state if / elif block selects resume_state, move the existing usage assignment out of the status in ("approved", "rejected") branch and apply it to both paths:

if resume_state is not None and resume_state._context is not None:
    resume_state._context.usage = context.usage

Place this immediately after the pending_resume_state if / elif block and before if run_result is None:. This preserves top-level checkpoint isolation while ensuring cached nested Agent.as_tool() resumes continue accumulating usage on the current outer run.

[chiruu12]

Done. Moved the usage rebind out of the approved/rejected branch so it applies to the cached resume path too, and kept the _copy_for_run_state deepcopy. Added a test that fails without the runtime change. Full stack is green.