fix(delegation): preserve completed child results on parent interrupt

[andyvandaric]

What does this PR do?

Fixes a race condition in the batch delegation interrupt handler where a child that completes between the last _cf_wait() return and the _interrupt_requested check has its result discarded — replaced with a fabricated "interrupted" entry even though the work was already done.

In the batch polling loop (delegate_tool.py, ~line 2370), when _interrupt_requested fires, the old code iterates over pending futures calling f.done() point-in-time. This is inherently racy:

• A future can transition from not-done → done between the check and the else branch
• Futures completing between the last _cf_wait(timeout=0.5) return and the top-of-loop interrupt check are never collected through the normal done path

The fix replaces the racy iteration with a 2-second grace window via wait(pending, timeout=2.0, return_when=ALL_COMPLETED), cleanly separating futures into completed (collect real results) vs still-running (fabricate as interrupted).

Related Issue

• Already merged: fix(delegation): add hard timeout and stale detection for subagent execution #13770 (hard timeout + heartbeat for stuck children)
• Related open issues: [Bug]: delegate_task blocks parent session silently with no progress output when children hang on tool calls #45496 (delegation results lost), Non-blocking subagent delegation: a stuck/slow child freezes the whole parent turn #47302 (non-blocking delegation)
• Related open PR: feat: non-blocking background agent delegation (async_delegation toolset) #8482 (stale async delegation redesign)

This PR is complementary — it fixes the specific race where already-computed results are lost on interrupt, without requiring the larger async delegation redesign from #8482.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• Updated tools/delegate_tool.py batch interrupt handler: replaced racy f.done() iteration with concurrent.futures.wait(pending, timeout=2.0, return_when=ALL_COMPLETED) grace window that cleanly separates done vs still-pending futures.
• Added tests/tools/test_delegate_batch_interrupt_race.py with 3 test cases covering the race condition scenarios.

How to Test

1. python -m pytest tests/tools/test_delegate_batch_interrupt_race.py -v
   • Result: 3 passed.
2. python -m pytest tests/tools/test_delegate.py -v
   • Result: 131 passed (9 pre-existing failures due to missing concurrent_log_handler in dev env, unrelated to this change).
3. python -m pytest tests/tools/test_delegate_subagent_timeout_diagnostic.py tests/tools/test_delegate_toolset_scope.py tests/tools/test_delegate_composite_toolsets.py -v
   • Result: 17 passed.

Full-suite note: pytest tests/ -q does not fully complete on this local environment due to 9 pre-existing failures from missing concurrent_log_handler module (required by tests that instantiate real AIAgent without mocking). All tests touching the delegation interrupt path pass. The focused touched-path suites above cover the change surface.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass — all touched-path tests pass; 9 pre-existing failures from missing concurrent_log_handler are unrelated (see Full-suite note above)
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Windows 10

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — N/A
• I've updated cli-config.yaml.example if I added/changed config keys — N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide
• I've updated tool descriptions/schemas if I changed tool behavior — N/A

Screenshots / Logs

N/A. This is a concurrency correctness fix with unit test verification.

[teknium1]

Thanks for isolating a real race in the synchronous batch aggregation path; current main still has the point-in-time f.done() branch at tools/delegate_tool.py:2596-2623.

Problems

• The added drain at tools/delegate_tool.py:2378-2380 waits up to two seconds for work that was not complete when interruption was observed. The new test explicitly treats a child completing at 1.5s after an interrupt at 0.3s as completed (tests/tools/test_delegate_batch_interrupt_race.py:247-249, 270-271). That is a cancellation-policy change, not just preservation of already-computed output.
• The first two timing tests do not reproduce the original TOCTOU: current main's FIRST_COMPLETED wait collects children that finish at 0.1s before their 0.3s/0.5s interrupt points (tools/delegate_tool.py:2629-2649; tests/tools/test_delegate_batch_interrupt_race.py:107-109, 137-140).

Suggested changes

• Decide and document whether interruption may wait for post-interrupt completions; otherwise keep the interrupt path bounded while collecting only results known complete at its cutoff.
• Add a deterministic race regression test that controls the transition between the old completion observation and fabrication.

This is an automated hermes-sweeper review.

[teknium1]

This drain waits for work that was still running when the parent interrupt was observed. That is broader than preserving already-computed results and changes the existing prompt-interrupt behavior into a two-second grace policy. Please either make that policy explicit and intentional, or collect only work known complete at the interrupt cutoff.

[teknium1]

This child completes 1.2 seconds after the test sets the interrupt, so the assertion validates the new grace-window policy rather than the original f.done() TOCTOU. Please add a deterministic test that controls the completion transition between the old observation and fabricated-result path.