fix: match pending nested agent-as-tool runs by call_id on resume

[tao-hpu]

Bug

_restore_pending_nested_agent_tool_runs pairs the unfiltered serialized entries (processed_response_data["functions"]) against the filtered deserialized runs with a positional zip(..., strict=False). _deserialize_actions drops an entry when its tool no longer resolves (if not tool: continue) or its tool call fails to parse, so a single dropped entry shifts every later pairing.

Trigger: resume a paused run where a function whose tool the resuming agent no longer exposes precedes an agent-as-tool call that carries agent_run_state with a pending nested interruption. The pending state then binds to the wrong tool call, or (at the tail) is silently lost, so the nested approval never re-surfaces.

Fix

Build a {tool_call.call_id: run} lookup from the deserialized runs and match each serialized entry by its tool_call.call_id instead of by position. Behavior for the aligned case is unchanged.

Tests

Added test_restore_pending_nested_run_when_earlier_entry_dropped: a serialized functions list whose first entry references a removed tool, followed by an agent-as-tool entry with a real agent_run_state (built via the existing HITL helpers) carrying a pending approval. It fails on the old positional zip (the pending nested run is lost) and passes with this change. This restore path previously had no test coverage.

pytest tests/test_run_state.py tests/test_agent_as_tool.py tests/test_agent_tool_state.py: 247 passed. ruff check, ruff format --check, and mypy on the touched files: clean.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4aa4718f78

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".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 221a01e519

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".

[tao-hpu]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Delightful!

Reviewed commit: 221a01e519

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]

Thank you for identifying this issue and for providing a concrete implementation and tests. Your key observation was exactly right: _deserialize_actions() may filter entries, while _restore_pending_nested_agent_tool_runs() previously relied on positionally zipping the original entries with the filtered actions. That insight directly informed the fix we are taking.

We decided to proceed with a maintainer-side implementation #3753 that preserves the association at the point of deserialization. Each successfully deserialized function action carries its normalized nested run-state data, rather than reconstructing the association after the original and filtered lists have diverged.

We prefer this approach because it:

• removes the positional assumption at its source;
• does not duplicate the ephemeral cache's private signature rules;
• avoids signature reconstruction and call_id/FIFO fallback heuristics;
• keeps the relationship explicit and strongly typed;
• attaches restoration metadata only to function actions that need it; and
• naturally handles missing, disabled, and malformed earlier entries, including mixtures of multiple skipped and pending calls.

We also expanded the regression coverage around independent cache scopes, namespaced and deferred tools, multiple pending states, JSON round-trips, and both approval and rejection after restoration.

The proposed signature-based implementation is thoughtful and works for the cases it covers, but it repairs the association after deserialization. The maintainer implementation instead preserves the original association through the filtering boundary, which we believe is simpler and less likely to drift as tool-call identity or cache behavior evolves.

Given the overlap, we are going to proceed with the maintainer implementation rather than merge this PR. We genuinely appreciate the report, analysis, and implementation idea—they were instrumental in getting this fixed.