Python: Fix RUN_FINISHED.interrupt to accumulate all interrupts when multiple tools need approval

[moonbox3]

Motivation and Context

When multiple tool calls required approval in the same turn, flow.interrupts was being reassigned with a single-element list for each approval request, causing only the last interrupt to appear in the RUN_FINISHED event. This meant clients could not see or approve earlier tool calls.

Fixes #4590

Description

The root cause was that both _emit_approval_request in _run_common.py and the confirmation handling in _agent_run.py used flow.interrupts = [...] (list reassignment) instead of appending to the existing list. The fix changes these assignments to flow.interrupts.append(...) so that each approval request is accumulated. A new test verifies that three consecutive approval requests result in three distinct interrupts in flow.interrupts.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• Is this a breaking change? If yes, add "[BREAKING]" prefix to the title of the PR.

---

Note: PR autogenerated by moonbox3's agent

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/ag-ui/agent_framework_ag_ui
_agent_run.py	475	49	89%	148–155, 194–195, 202, 299, 311, 315, 317–318, 334, 361–362, 380–384, 492–494, 506–508, 612, 620, 729, 731–732, 785, 800–801, 808, 873, 896, 904, 906, 914, 967, 970, 980–981, 988
_run_common.py	203	13	93%	263–265, 290–291, 297–302, 405–406
TOTAL	23998	2642	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
5223	20 💤	0 ❌	0 🔥	1m 24s ⏱️

[moonbox3]

Automated Code Review

Reviewers: 4 | Confidence: 83%

✗ Correctness

The diff changes flow.interrupts = [...] (list assignment) to flow.interrupts.append(...) in two locations, enabling accumulation of multiple interrupts per turn. The logic is sound and the test correctly validates the new behavior. However, the diff does not show a corresponding change to the FlowState dataclass to add an interrupts field with default_factory=list. The installed version of FlowState has no interrupts field — the old assignment (= [...]) worked because it dynamically created the attribute, but .append() will raise AttributeError if the attribute doesn't already exist. If this field was added in another commit in the same PR, the change is correct; otherwise this is a blocking bug.

✓ Security Reliability

The diff changes flow.interrupts from list replacement (= [...]) to list append (.append(...)) in two locations, enabling accumulation of multiple interrupt entries per turn. A corresponding test validates the new behavior. The change is straightforward and correct, with no security issues. The only reliability consideration is that flow.interrupts must be pre-initialized as a list (e.g., via field(default_factory=list)) on FlowState for .append() to succeed on the first call — the old assignment-based code would have worked even without prior initialization. Since the FlowState definition change is not included in this diff, this is worth confirming but the new test exercising FlowState(message_id="msg-1") implies the field is properly defined.

✓ Test Coverage

The diff changes flow.interrupts from assignment (= [...]) to append (.append(...)) in two locations: _run_common.py (_emit_approval_request) and _agent_run.py (run_agent_stream). A new test covers the accumulation behavior for _emit_approval_request, but no test covers the identical change in _agent_run.py's streaming path. The added test is well-structured with meaningful assertions (checks count and interrupt IDs), but only exercises one of the two modified code paths.

✓ Design Approach

The change fixes a real bug: every call to _emit_approval_request was overwriting flow.interrupts with a new single-element list instead of accumulating entries, meaning only the last approval request in a turn was ever surfaced. The fix (.append()) is the correct approach and aligns with the list-accumulation pattern implied by flow.interrupts being a list. The accompanying test directly validates the multi-interrupt accumulation scenario. No design issues are present — this is a straightforward and correct targeted fix with no leaky abstractions, fragile assumptions, or symptom-level masking.

Flagged Issues

• FlowState dataclass does not appear to define an interrupts field in this diff. The old flow.interrupts = [...] dynamically created the attribute via assignment, but .append() requires it to already exist. If FlowState was not updated elsewhere in this PR to include interrupts: list[dict[str, Any]] = field(default_factory=list), both call sites will raise AttributeError on first invocation.

Suggestions

• Ensure FlowState declares interrupts: list[dict[str, Any]] = field(default_factory=list) so that .append() works on a freshly constructed instance. If this was added in a separate commit not shown in this diff, confirm it is present.
• Consider whether flow.interrupts should be cleared between turns or agent runs to prevent stale interrupt data from carrying over. Since FlowState represents state for a single AG-UI run this is likely safe, but worth verifying the lifecycle.
• Add a test exercising the run_agent_stream code path in _agent_run.py (line 924) to verify that multiple approval requests in a streaming run also accumulate in flow.interrupts. Currently only the _emit_approval_request helper in _run_common.py is covered, so a regression in the streaming path would go undetected.

---

Automated review by moonbox3's agents

[Copilot]

Pull request overview

This PR fixes AG-UI’s interrupt metadata so that when multiple tool calls require approval in the same agent turn, all approval interrupts are preserved and surfaced in RUN_FINISHED.interrupt (instead of only the last one), enabling clients to see/approve every pending tool call.

Changes:

• Update _emit_approval_request to append to flow.interrupts rather than overwriting it.
• Update the predictive-tool confirmation path in run_agent_stream to append additional interrupts rather than overwriting.
• Add a unit test asserting that multiple approval requests accumulate into multiple flow.interrupts entries.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
python/packages/ag-ui/agent_framework_ag_ui/_run_common.py	Accumulates approval interrupts by appending to FlowState.interrupts.
python/packages/ag-ui/agent_framework_ag_ui/_agent_run.py	Accumulates multiple confirmation interrupts during agent streaming runs.
python/packages/ag-ui/tests/ag_ui/test_run.py	Adds regression test ensuring multiple approval requests produce multiple interrupts.

---

You can also share your feedback on Copilot code review. Take the survey.

[moonbox3]

Automated Code Review

Reviewers: 4 | Confidence: 97% | Result: All clear

Reviewed: Correctness, Security Reliability, Test Coverage, Design Approach

---

Automated review by moonbox3's agents